NVM · REFERENCE LIBRARY
Browse the Technology Atlas
Choose a topic to explore IP cell principles, technology lineage, foundry processes or storage physics. Each study connects diagrams and operating conditions to public sources.
NVM KNOWLEDGE HUB · REFERENCE LIBRARY
IP cell studies, technology lineage, process integration and device-physics references.
NVM · REFERENCE LIBRARY
Choose a topic to explore IP cell principles, technology lineage, foundry processes or storage physics. Each study connects diagrams and operating conditions to public sources.
IP AND PROCESSES · CELL STUDIES
Choose an IP and follow the same cell through programming, reverse update and read. The storage region and operating paths explain the differences among OTP, MTP, eFlash, MRAM and ReRAM.
One-time programming can store charge or alter a dielectric; these are different cell mechanisms.
Follow PGM, ERS and subsequent PGM on reusable storage cells. Identify each implementation’s carrier mechanism separately.
Embedded Flash macros still store charge, but extra masks, array organization and read paths are checked separately from MTP.
Read embedded macros through their junctions, access transistors and electrical paths.
Follow oxygen-related defects or metallic paths through SET, RESET and sensing.
IP AND PROCESSES · TECHNOLOGY LINEAGE
Follow three IP families into the Synopsys portfolio through original cells, product announcements and succession events. Each lineage links to its cell lesson, followed by subsequent public products and the physical scope supported by sources.
XPM stores a one-time state in a gate-oxide antifuse. Its historical 2T cell separates the storage and selection MOS roles. XPM later joined Synopsys with other Kilopass products.
Figure 1 of WO2007090089A2 identifies the existing XPM with M0 for storage and M1 for selection. This is the structure reconstructed in the lesson.
A Kilopass announcement names XPM, Gusto and 2T CMOS antifuse, linking the named historical products to the cell approach.
The official announcement includes XPM, Gusto and SecretCode in the acquired portfolio and describes an expanded 1T/2T OTP offering.
As checked on 2026-09-10, Synopsys publicly lists a 1T/2T antifuse OTP portfolio. Its advanced-node article discusses cell sizing, analog supply and sensing, repair, ECC and controller design. These span cell and macro design; the article does not trace every current product to an original vendor cell.
The cell lesson uses the historical XPM 2T functional topology in patent Figure 1. It excludes the self-sensing node of Figure 2 and undisclosed current FinFET sections. Business succession and internal cell implementation have separate evidence scopes.
1T-Fuse uses one continuous gate over thick and thin oxide, integrating channel selection and antifuse storage in a split-channel cell. This is the starting point for understanding the named Sidense 1T architecture.
Figure 2 in the Sidense author article shows thick/thin oxide, continuous poly and one BL diffusion. The lesson distinguishes its selection region from the persistent conduction region.
Thin-oxide conduction creates an irreversible 1T-Fuse cell state. Emulated MTP updates use multiple storage locations and management; they do not erase or repair the original antifuse.
The official announcement names the single-transistor, split-channel 1T-Fuse technology, linking the original cell approach to the acquired portfolio.
Current Synopsys OTP material lists 1T and 2T offerings. Later design descriptions cover oxide-breakdown control, leakage and sensing, plus macro repair and ECC. Public material does not map every current node to the historical Sidense section.
The lesson retains the n-type structure and single BL diffusion shown by the original author in 2007; read arrows are inferred from that structure. It does not import p-type patent biases, extra terminals or undisclosed current macro wiring.
AEON is a logic-process floating-gate MTP family originating at Impinj. Virage Logic acquired the business, which subsequently entered Synopsys with Virage Logic and continued in named AEON MTP ULP products.
The company announcement names AEON/MTP and floating-gate transistors. Its specifications apply to that product announcement.
The SEC filing records completion of an asset purchase on this date; Impinj as a whole was not acquired.
NVM manager Craig Zajac identifies FN for both operations in a company-authored article. This defines the cell lesson scope.
The completion announcement includes NVM in the added portfolio, bringing this succession chain into Synopsys.
The announcement explicitly uses AEON branding, supporting product-family continuity without asserting identical internal wiring.
The 2013 official announcement explicitly uses DesignWare AEON MTP ULP branding. As checked on 2026-09-10, the current Synopsys MTP ULP page describes single-poly, floating-gate and zero-mask-adder positioning. That is a check date, not a launch date or evidence of unchanged wiring across generations.
The named FN/FN basis is a 2009 article by a Virage Logic NVM manager. The diagrams use coupling, tunneling roles, FG and read MOS to explain electron entry and removal. Undisclosed device count, p/n polarity and pin arrangement remain unspecified.
TECHNOLOGY FOUNDATIONS
Use the reference table to compare stored states, operating paths and implementation conditions. Standalone devices remain background references.
Charge, conductive structure, magnetization, ion distribution, crystal phase and polarization determine the physical mechanism. OTP is not one physics: floating-gate OTP stores charge; AntiFuse OTP forms a dielectric path. Study conventional standalone EEPROM separately from embedded MTP IP; within MTP IP, distinguish foundry double-poly EEPROM and third-party single-poly routes. NOR and NAND still describe array organization.
Volume production, completed qualification, research demonstrations, and development plans must be tied to a named company, process node, version, and application. Finding one MRAM product in volume production does not establish the same maturity for every SOT-MRAM implementation.
SCM describes an application role that addresses the gap between memory and storage; it is not another bitcell type. Application, access semantics, persistence through power loss, and current supply status must be explained separately.
Maturity labels refer to verified, named implementations. Open a study for its full limitations. Search includes operations, applications, and source keywords.
Sixteen Technology and IP Studies
| Technology and Physical Family | Stored State | Status and Principal Tradeoffs |
|---|---|---|
| eFuse: Permanent Conductance Programming Irreversible Structures | A bit is stored as a difference in the resistance of a conductive path. An unprogrammed fuse typically has low resistance; a controlled current causes material migration or a break in a designated region, producing higher resistance. Logic 0/1 is defined by sensing and encoding. High resistance does not intrinsically correspond to a particular bit value, and the programmed state must not be assumed to be an ideal open circuit. | Identified Implementation IBM's 2007 technology review describes eFUSE evolution from 180 nm to 45 nm and applications in memory redundancy, chip identification, and analog trimming, establishing implementations in identified processes. This material also uses IBM and TSMC patents to explain polysilicon/silicide and metal-via structures. The same physical location cannot be erased in normal operation; field updates require reserved locations and data-validity encoding. |
| Antifuse: Permanent Conduction Through Dielectric Breakdown Irreversible Structures | Before programming, the storage dielectric separates two electrodes and permits only very small leakage. Programming creates a permanent, detectable conduction path through a high electric field. Information resides in the conduction difference before and after breakdown, opposite to the typical eFuse transition from low to high resistance. Both can provide OTP, but their storage materials and programming conditions differ. | In Volume Production Synopsys' 2018 acquisition statement identifies Kilopass antifuse 1T/2T IP, including XPM, Gusto, and SecretCode, and reports aggregate cumulative shipments exceeding 10 billion units. The current OTP page separately lists advanced-node silicon validation and specific automotive qualifications, indicating continued commercial availability. No normal erase is available; misprogramming recovery and appended updates require reserved capacity and encoding from the outset. |
| Conventional Standalone EEPROM: Local Windows and Fine Updates Charge Storage | A floating gate is a conductive island surrounded by insulating layers, with no direct DC metal connection to it. Retained charge changes how the control gate acts on the channel, shifting the MOS threshold voltage. In a typical n-channel example, adding electrons makes conduction more difficult. Reading measures the channel; normal read operation does not require draining the stored electrons. | Production Device Microchip publicly offers the 24LC256 serial EEPROM and its full datasheet, covering packages, I2C, byte writes, a page buffer, and an internal high-voltage generator. This establishes a standalone product; the local-window teaching diagram is supported separately by a public patent. The external device consumes package, board, and interface resources; serial transfer adds end-to-end latency. |
| Embedded MTP IP: Foundry Double-Poly and Third-Party Single-Poly Charge Storage | This topic focuses on floating-gate embedded MTP/EEPROM IP. Charge remains on an insulated conductive floating node and alters channel conduction through capacitive coupling. n-type and p-type storage transistors have different read-state behavior: adding electrons makes the n-channel US5844271A example harder to turn on, while eMemory describes its p-type NeoMTP device as turning on after electron injection. | Commercial IP Current Synopsys and eMemory product pages directly identify single-poly MTP/EEPROM IP. The 2003 X-FAB XC06 brief provides a historical foundry double-poly NVM example. YMC has a logic-process MTP offering and a separate single-poly patent; Floadia ZT has a public floating-gate/FN program-and-erase example. Single-poly does not eliminate area or qualification cost; coupling capacitors, isolation, and high-voltage peripherals still consume die resources. |
| NOR: Stacked-Gate and Split-Gate Code Storage Charge Storage | NOR describes array connectivity and access organization, not a unique storage material. This topic uses floating-gate NOR: charge changes cell threshold voltage, and the selected cell is sensed through the bitline and source path. Both stacked-gate and split-gate cells can serve NOR arrays, but their selection channels, programming efficiency, and erase control differ. | In Volume Production Microchip's SST39SF020A was listed as in production when reviewed, with a public summary specifying 2 Mb and a 4.5–5.5 V parallel flash interface. SST's SuperFlash technology brochure separately provides an identified technical lineage for split gates, source-side injection, and inter-gate FN erase, allowing commercial implementation evidence to be compared with the mechanism lesson. More cell contacts and wiring reduce density relative to NAND; complete cost must be compared at the required capacity. |
| SONOS and NROM: Charge Trapping in Insulating Layers Charge Storage | Electrons remain in trapping centers within insulating materials such as silicon nitride, changing the potential seen by the channel and its threshold voltage. The trapping layer is not a conductive floating gate, and charge can have a spatial distribution. Channel-wide SONOS program/erase examples and localized NROM charge-trapping examples therefore require different operating and sensing explanations; simply recoloring a floating gate is not sufficient. | In Volume Production Infineon publicly lists SONOS eFlash production at 130, 65, 55, 40, and 28 nm and explicitly describes a 2T cell with FN program/erase. Its MCU shipments and licensable macro information support classification as an established platform. Localized-charge NROM is treated separately here through Saifun's original patent, with its own evidence boundary. Trap, interface, and blocking-layer quality jointly determine retention and endurance; charge-loss risk remains. |
| NAND: Planar Strings, Vertical Stacks, and Multilevel Storage Charge Storage | NAND data can still be retained by charge in a floating gate or dielectric trapping layer that changes threshold voltage. NAND itself describes an array organization with multiple cells connected in series. Storing N bits per cell requires 2 to the Nth power distinguishable states, such as eight for TLC and sixteen for QLC. More bits do not provide additional windows of unchanged width for free. | Commercial Generations Established Kioxia's fundamentals explanation records commercial planar 15 nm technology and BiCS FLASH generations: 48 layers in 2015, 96 layers in 2018, 112 layers in 2020, and 162 layers in 2022. This establishes 3D NAND as a mature commercial family and supports a historical comparison of planar scaling and vertical stacking. Page programming and block erase make small updates involve data movement, write amplification, and garbage collection. |
| Toggle MRAM: Magnetic-Field Sequencing Magnetism | The bit is stored in the magnetization direction of the free magnetic layer. Parallel and antiparallel alignment relative to the reference layer produce different resistance levels in the magnetic tunnel junction. A magnetic energy barrier maintains the direction after power is removed. Toggle specifically denotes a write method that reverses data through the rotation of coupled magnetic moments; it is not a general name for all field-written MRAM. | In Volume Production Everspin's 2025 annual filing explicitly states that Toggle products entered volume production in 2008 and that devices with capacities of 128kb–32Mb continue to ship. This is commercial evidence for a named product family, rather than maturity inferred from a patent or experimental device. The current, spacing, and half-select conditions of magnetic-field lines limit density scaling. |
| STT-MRAM: Spin Current Through the Junction Magnetism | The stored quantity remains the orientation of the free magnetic layer relative to the reference layer, and reading relies on the resistance difference of the magnetic tunnel junction. The principal difference from Toggle is writing: current passing through the magnetic stack carries spin angular momentum and exerts torque on the free layer, changing its magnetic state. Nonvolatility comes from the magnetic energy barrier, not from keeping current inside the device. | In Volume Production Everspin has shipped STT products with DDR-derived interfaces and SPI-class products. In 2026, its 64Mb high-reliability xSPI product also has evidence of completed production qualification and ordering availability. Embedded implementations must be linked individually to a specific MCU or process document. Reading and writing share the tunnel barrier, requiring joint design of write stress and read disturbance. |
| SOT-MRAM: Separate Read and Write Paths Magnetism | SOT-MRAM also retains data in the magnetization direction of an MTJ free layer and senses it through magnetoresistance. Its distinguishing feature is that write angular momentum is generated primarily by a spin-orbit material beside or beneath the free layer and injected into it, rather than by sending the main write current through the tunnel barrier. The stored physical quantity is therefore similar to STT, while the write structure and array cost differ. | Research Demonstration imec demonstrated extremely scaled devices and functional arrays in 2023–2024. TSMC's 2025 annual report also records field-free Type-C SOT-MRAM presented at IEDM 2025. These are concrete device and array research results, but they are insufficient to label a last-level cache as being in volume production. The third terminal, write line, and selection circuits add area and may offset the benefit of cell scaling. |
| VCM ReRAM: Oxygen Redistribution and Conductive Paths Resistive Switching | VCM stores data in the ionic distribution, local redox state, or interfacial barrier of an oxide, producing distinguishable resistance states. A typical filament model explains conduction and rupture through redistribution of oxygen ions/oxygen vacancies, but not every device has a single clearly defined filament. Materials, electrodes, and measurement evidence determine the mechanism; a hysteretic I–V curve alone is insufficient to identify VCM. | Qualification Completed Weebit/DB HiTek 130nm BCD RRAM has public evidence of completed technology qualification, and SkyWater S130 has a named 1T1R reliability test vehicle. These support the maturity of resistive-memory integration; a product name alone cannot reveal its complete VCM material cross section. Forming and path growth are stochastic, so the verification algorithm may dominate write latency. |
| ECM/CBRAM: Growing and Dissolving a Metal Bridge Resistive Switching | ECM changes a conductive path through the motion and redox reactions of active-metal ions. An existing conductive bridge commonly produces a low-resistance state; dissolving a critical part of the bridge produces a high-resistance state. Both ECM and oxygen-vacancy VCM exhibit resistive switching, but their ion sources and path materials differ. CBRAM is a common commercial name for this conductive-bridge memory. Sharing the ReRAM label does not justify combining their physical models. | Historical Commercial Shipments The CBRAM section of Adesto's 2019 annual filing explicitly records commercial product shipments. ECM/CBRAM therefore cannot uniformly be labeled as never commercialized. This evidence supports historical product maturity, but does not establish every subsequent node or availability of the original part numbers in 2026. Rapid formation and long-term stability of thin metal bridges constrain each other. |
| PCM: Controlling Phase with Thermal History Phase Change | PCM stores data in the fraction and geometry of crystalline and amorphous phase-change material. In a typical electronic device, the crystalline state has lower resistance and the amorphous state higher resistance; material kinetics retain the state after power removal. The actual state is more than an abstract resistance value. It includes the location and size of the phase-change region, degree of crystallization, and evolution over time, which jointly determine reading and lifetime. | In Volume Production On 2026-09-10, ST's SR6P6C8 product page explicitly lists production status, includes PCM in the product description, and provides specific ordering codes. This is a concrete commercial example of embedded PCM, avoiding judgments about the entire technology family based only on one discontinued storage-class product. Peak RESET current may require a larger access transistor. |
| Capacitor FeRAM: Sensing Polarization-Switching Charge Ferroelectricity | The bit is represented by the polarization direction retained in a ferroelectric material after the external electric field is removed. Unlike DRAM, which relies on temporarily stored free charge, FeRAM is based on switchable remanent polarization. Reading uses the different charge responses when polarization switches and when it does not. Distinguishing the material's retention mechanism from the circuit's sensing method explains why nonvolatile memory may still require restoration after a read. | In Volume Production Infineon EXCELON F-RAM is a named commercial family with a datasheet for the 16Mb CY15B116QI/CY15V116QI. The document specifies the interface, operating temperature, and retention conditions at different temperatures. It supports discussion of conditional product performance without extrapolating from a single material paper. Reading may require restoration, so internal timing and power-failure scenarios must be included in reliability design. |
| FeFET: Translating Polarization into Threshold Voltage Ferroelectricity | FeFET uses ferroelectric polarization in the gate stack to change channel electrostatics, giving the transistor distinguishable high and low threshold voltages. Reading selects a gate bias between those thresholds and senses channel current. Charge trapping and detrapping also affect the actual memory window, so not every threshold-voltage change can be attributed solely to polarization. | Research Demonstration The KIOXIA study associated with IEDM 2023 and reviewed here explicitly demonstrates control of trapped charge and polarization stability through interface engineering. An original FeFET PUF paper provides additional evidence. These support concrete device and circuit research, but are insufficient to identify a commercial production part using this stack. Voltage division between the ferroelectric and interfacial layers may require higher write voltage and increase dielectric stress. |
| FTJ: Modulating the Tunnel Barrier with Polarization Ferroelectricity | FTJ controls tunneling current through the polarization direction of a thin ferroelectric barrier. The two directions produce different effective barrier profiles and resistance levels, commonly described through tunnel electroresistance, or TER. FTJ and FeFET both use polarization, but FTJ does not rely on threshold-voltage amplification in a semiconductor channel. Read current, barrier thickness, electrode screening, and leakage therefore become central tradeoffs. | Research Demonstration The 2024 original FTJ paper and TSMC FTJ structure patent publication reviewed here support concrete thin-film and reliability research. A datasheet and supply evidence identifying a commercial production part using this structure have not been obtained, so the research-demonstration label is retained. Tunneling current and polarization retention impose competing requirements on film thickness. |
No studies match. Shorten the search term or clear the family and maturity filters.
Read NeoBit and NeoFuse side by side: stored floating-gate charge versus altered gate-dielectric transport.
MTP describes the ability to program more than once. The floating-gate IP lessons compare named PGM/ERS mechanisms; MRAM and ReRAM remain in emerging-NVM categories, while Flash/SONOS retain their own structures. Read each reverse-update operation and its subsequent write. YMC product evidence is separate from the independent CHI/BBHH teaching example.
Trace STT switching and sensing, then compare oxygen-related and metallic-path ReRAM SET/RESET.
Use the background topics for device, array, process and system limits. Standalone products provide context rather than the main learning path.
Switching a device twice establishes only that it has usable storage states. A practical memory must also select its target from a large population of cells, avoid disturbing its neighbors, and read data correctly across temperature, aging, and process variation. Selectors, wires, and peripheral circuits therefore determine how much of the cell-level advantage survives.
Storage-class memory (SCM) addresses the gap in requirements between DRAM and NAND storage. It is not another bitcell type, and adopting CXL does not automatically establish an SCM implementation. Understanding SCM requires distinguishing storage physics, attachment, access granularity, and which data can actually be recovered after failure.
eMemory · OTP IP
Follow the series select transistor and p-type floating-gate storage transistor as electron injection changes read current. Then distinguish normal OTP operation from the physical possibility of ultraviolet erasure.
Series pMOS selector and pMOS floating-gate cell. Historical section: n-well in p substrate, p+ source/shared region/drain, separate select gate and p+ FG; no FG wire or stacked control gate. Dielectric isolates FG. The historical p+ model adds electron–hole recombination leaving localized negative ionic charge; stored Q− is not entirely free electrons.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Hot-hole-induced electron injection
Carriers: accelerated holes create electron–hole pairs; some hot electrons enter FG through oxide.
Do not draw oxide rupture, a permanent filament or nMOS CHE; do not invent voltages or universal current p+ doping. SL, SG/WL, BL and NW; FG floats. SG is separate from FG; capacitive BL coupling is not a DC connection.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Normal OTP operating boundary
The normal OTP interface lacks electrical erase. This does not imply physical irreversibility: the 2021 brief lists UV erase, subject to implementation and package.
Do not draw oxide rupture, a permanent filament or nMOS CHE; do not invent voltages or universal current p+ doping. SL, SG/WL, BL and NW; FG floats. SG is separate from FG; capacitive BL coupling is not a DC connection.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Channel-current sensing
Sense output current while preserving the stored state; the macro defines logic coding.
Do not draw oxide rupture, a permanent filament or nMOS CHE; do not invent voltages or universal current p+ doping. SL, SG/WL, BL and NW; FG floats. SG is separate from FG; capacitive BL coupling is not a DC connection.
A floating-gate OTP stores its state in charge. The selector controls access, while programming moves the storage transistor to another sensed state. An interface without electrical erase is a different OTP design path from an irreversible dielectric change. Public NeoBit principles describe two series p-MOSFETs and CHEI. The cell uses I/O devices: about 6.5 V PGM for a 3.3 V cell and 7.5 V for a 5 V cell; an n-type floating-gate cell at the same node needs a higher Vpgm. That I/O window is why a p-type cell can embed in standard logic CMOS — not core-GOX-breakdown AntiFuse.
eMemory · OTP IP
Start at the n-type cell's gate dielectric and follow high-field defect creation, changes in effective tunneling distance and the gate current used for sensing.
Storage uses an nFET gate dielectric and gate-current sensing. Published 3T adds regulation; show selection, regulation and antifuse functions with conceptual connectivity. The programmed dielectric-defect state persists; data is not an FG electron count.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
High-field defect generation and enhanced tunneling
Carriers: defects shorten the effective barrier; electrons tunnel across it.
No thick metallic short or FG storage; the damaged high-k/interfacial sublayer of current advanced processes is unverified. Antifuse gate AF, underlying Si, selection/regulation controls and BL; storage dielectric lies between AF and Si, not in the selector oxide.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Normal OTP operating boundary
Normal operation has no defect-repair erase step; reducing bias does not restore the initial dielectric.
No thick metallic short or FG storage; the damaged high-k/interfacial sublayer of current advanced processes is unverified. Antifuse gate AF, underlying Si, selection/regulation controls and BL; storage dielectric lies between AF and Si, not in the selector oxide.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Gate-current sensing
Sense output current while preserving the stored state; the macro defines logic coding.
No thick metallic short or FG storage; the damaged high-k/interfacial sublayer of current advanced processes is unverified. Antifuse gate AF, underlying Si, selection/regulation controls and BL; storage dielectric lies between AF and Si, not in the selector oxide.
The storage event changes dielectric conduction. An ideal short does not explain the ultrathin-dielectric physics. Selection and regulation transistors support array operation; defects and tunneling in the storage region create the programmed read-current difference.
Kilopass; acquired by Synopsys in 2018 · OTP IP
The original patent explicitly names XPM and distinguishes the storage MOS from the select MOS.
Original patent Figure 1 describes XPM as storage MOS M0 plus select MOS M1. WLP drives the storage gate, WLR selects M1, and BL senses current. The enlarged M0 region shows function without inventing a second diffusion or an intermediate sense terminal.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Gate-Oxide Breakdown
Local M0 oxide breakdown creates a conductive path; electrons are illustrated from BL through M1 and silicon toward positive WLP.
Uses Figure 1 explicitly named XPM in the 2007 patent and the 2012 product announcement. Later self-sensing, latch and 3T drawings are excluded. Positive WLP and low BL define this nMOS teaching branch; read uses lower stress without implementation voltages. Acquisition announcements establish portfolio continuity; current 1T/2T or advanced-process articles do not prove this section is retained at every node.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Normal OTP Operating Boundary
Normal OTP provides no electrical erase; an ordinary reverse operation cannot restore the unprogrammed cell.
Uses Figure 1 explicitly named XPM in the 2007 patent and the 2012 product announcement. Later self-sensing, latch and 3T drawings are excluded. Positive WLP and low BL define this nMOS teaching branch; read uses lower stress without implementation voltages. Acquisition announcements establish portfolio continuity; current 1T/2T or advanced-process articles do not prove this section is retained at every node.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Oxide-Conduction Current Sensing
The sense amplifier compares low/high BL current. The macro defines 0/1 coding; read does not repair the oxide.
Uses Figure 1 explicitly named XPM in the 2007 patent and the 2012 product announcement. Later self-sensing, latch and 3T drawings are excluded. Positive WLP and low BL define this nMOS teaching branch; read uses lower stress without implementation voltages. Acquisition announcements establish portfolio continuity; current 1T/2T or advanced-process articles do not prove this section is retained at every node.
Learn the two transistor roles and the original XPM evidence; acquisition does not prove one cell is retained at every node.
Sidense; acquired by Synopsys in 2017 · OTP IP
One gate spans thick and thin oxide; persistent conduction through the thin region creates the OTP state.
Following original-author Figure 2 from 2007, one continuous poly gate connects to WL and the sole N+ diffusion connects to BL. Thick I/O oxide lies near BL; thin core oxide lies farther away. Breakdown links the gate to the channel through the thin region. Thicknesses and paths are enlarged for readability.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Gate-Oxide Breakdown
The thin core oxide locally breaks down. Electrons travel from BL diffusion through the channel and the thin region toward WL; the thick oxide remains intact.
The section follows the product article’s n-type structure. Positive WL and low BL define the teaching bias from which directions are inferred. Detailed voltages from the historical p-type patent example are excluded. Split channel means neither two separate gates nor a floating gate. Acquisition announcements establish portfolio continuity; current 1T/2T or advanced-process articles do not prove this section is retained at every node.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Normal OTP Operating Boundary
Normal OTP provides no electrical erase; an ordinary reverse operation cannot restore the unprogrammed cell.
The section follows the product article’s n-type structure. Positive WL and low BL define the teaching bias from which directions are inferred. Detailed voltages from the historical p-type patent example are excluded. Split channel means neither two separate gates nor a floating gate. Acquisition announcements establish portfolio continuity; current 1T/2T or advanced-process articles do not prove this section is retained at every node.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Oxide-Conduction Current Sensing
Compare low/high current along the WL/BL path. The macro defines logic coding and normal read preserves the state.
The section follows the product article’s n-type structure. Positive WL and low BL define the teaching bias from which directions are inferred. Detailed voltages from the historical p-type patent example are excluded. Split channel means neither two separate gates nor a floating gate. Acquisition announcements establish portfolio continuity; current 1T/2T or advanced-process articles do not prove this section is retained at every node.
The thick region controls access and the thin region stores conductance; split channel means neither two gates nor a floating gate.
Chuangfeixin CFX · OTP IP
CFX publicly lists Anti-fuse, eFuse and floating-gate OTP. This unit teaches only the Semi IP Hub named HV-macro gate-to-substrate oxide breakdown and marks the irreversible OTP limit.
CFX publicly lists Anti-fuse, eFuse, and floating-gate OTP. This drawing teaches only the Semi IP Hub named HV-macro gate-oxide breakdown and does not represent every SKU.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
A high-voltage pulse permanently conducts the gate-to-substrate oxide.
Establish the programmed state by the published mechanism without splicing unpublished bias tables.
This teaching case matches only the Semi IP Hub gate-oxide-breakdown narrative.
CFX also has eFuse and floating-gate OTP, which are not drawn in this frame.
OTP has no electrical erase back to insulation.
If the target macro is eFuse or floating-gate OTP, open a separate unit; do not reuse this drawing.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
A high-voltage pulse permanently conducts the gate-to-substrate oxide.
OTP has no electrical erase back to the initial state; this section only marks the restore limit.
An OTP cell has no electrical cycle back to an intact oxide.
If rewrite is required, use an MTP or eFlash unit instead.
The three OTP routes still must be checked separately during selection.
If the target macro is eFuse or floating-gate OTP, open a separate unit; do not reuse this drawing.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
A high-voltage pulse permanently conducts the gate-to-substrate oxide.
Sense the retained state under product read conditions, then latch and isolate.
Read does not use program-level high voltage.
A conducting cell draws larger current.
Actual read time and window are defined by the macro.
If the target macro is eFuse or floating-gate OTP, open a separate unit; do not reuse this drawing.
One vendor can license more than one OTP physics. Check whether the target macro is oxide breakdown, a fuse or a floating gate. This drawing covers only the published gate-oxide-breakdown account.
Attopsemi · OTP IP
I-fuse is a poly / metal-gate / metal fuse. Programming uses heat-assisted electromigration to raise resistance while staying below thermal runaway and explosive rupture.
A poly / metal-gate / metal fuse. Heat-assisted electromigration changes resistance while staying below thermal runaway and explosive rupture.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Local heating accelerates metal-atom migration and raises fuse resistance.
Establish the programmed state by the published mechanism without splicing unpublished bias tables.
I-fuse is a fuse, not a MOS oxide.
The vendor explicitly excludes explosive rupture and AntiFuse.
Poly, metal-gate, and metal fuses belong to this family; this drawing does not pick one cross-section.
I-fuse is neither AntiFuse nor conventional explosive eFuse.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Local heating accelerates metal-atom migration and raises fuse resistance.
OTP has no electrical erase back to the initial state; this section only marks the restore limit.
An OTP fuse has no electrical cycle back to low R.
This is not a rewritable MTP.
The physics boundary versus conventional explosive eFuse must be kept.
I-fuse is neither AntiFuse nor conventional explosive eFuse.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Local heating accelerates metal-atom migration and raises fuse resistance.
Sense the retained state under product read conditions, then latch and isolate.
Sense current is far below program current.
Sense circuits distinguish high R from low R.
Read must not drive the fuse near thermal runaway.
I-fuse is neither AntiFuse nor conventional explosive eFuse.
Keep I-fuse separate from AntiFuse and from conventional explosive eFuse. The stored quantity is fuse resistance, not whether a MOS gate oxide has broken down.
Floadia · OTP IP
Zero extra-mask Anti-fuse OTP. The teaching drawing shows a dielectric changing from insulation to permanent conduction and does not locate the breakdown site in a gate or capacitor.
Anti-fuse OTP with zero extra mask. The teaching drawing shows dielectric isolation then conduction and does not locate the breakdown site in a gate or capacitor.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
A single high-voltage event permanently conducts the dielectric.
Establish the programmed state by the published mechanism without splicing unpublished bias tables.
The product name is LEE Fuse ZA.
Zero extra mask is the integration claim; the breakdown site is unpublished.
A DRAM 1xnm production track cannot be extrapolated as the same cross-section on every logic node.
The product name is LEE Fuse ZA; a page typo of LEE Flash ZA does not change the mechanism class.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
A single high-voltage event permanently conducts the dielectric.
OTP has no electrical erase back to the initial state; this section only marks the restore limit.
Anti-fuse OTP is kept separate from LEE Flash ZT/G1/G2.
A page typo of LEE Flash ZA does not turn this unit into eFlash.
180 nm to sub-10 nm is a vendor node narrative.
The product name is LEE Fuse ZA; a page typo of LEE Flash ZA does not change the mechanism class.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
A single high-voltage event permanently conducts the dielectric.
Sense the retained state under product read conditions, then latch and isolate.
Read does not use program-level high voltage.
A conducting cell draws larger current.
Actual specifications follow the licensed target version.
The product name is LEE Fuse ZA; a page typo of LEE Flash ZA does not change the mechanism class.
The product name is LEE Fuse ZA. If a page writes LEE Flash ZA, the cell remains Anti-fuse OTP and must not be read as the rewritable charge cells in LEE Flash ZT, G1 or G2.
eMemory · MTP IP
Follow the control-coupling region, floating node and tunneling region as FN transport stores and removes electrons. A read transistor then senses the stored state.
The same storage cell supports programming, electrical erase and subsequent programming. ERS restores a window suitable for another program operation; it does not require every carrier to disappear. Whether the host issues a separate erase command depends on the macro or component interface.
PGM → ERS → PGM
Single-poly FG with capacitive-coupling MOS structures and selectors. Coupling and tunneling are functional roles; public evidence does not fix the device count or p/n arrangement. After high-field removal, isolated FG charge shifts the read-channel threshold.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
Carriers: electrons enter FG from a MOS tunneling region by FN tunneling.
Exclude the historical CHE/FN branch; do not assert exactly two physical capacitors or that more electrons always mean ON. C, T, R and S denote coupling, tunneling, read-channel and selection functions, not official pins. Shared FG has no external supply connection.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
An FN charge-removal path enables electrical rewriting; dielectric wear and macro conditions limit cycling.
Exclude the historical CHE/FN branch; do not assert exactly two physical capacitors or that more electrons always mean ON. C, T, R and S denote coupling, tunneling, read-channel and selection functions, not official pins. Shared FG has no external supply connection.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Channel-current sensing
Sense output current while preserving the stored state; the macro defines logic coding.
Exclude the historical CHE/FN branch; do not assert exactly two physical capacitors or that more electrons always mean ON. C, T, R and S denote coupling, tunneling, read-channel and selection functions, not official pins. Shared FG has no external supply connection.
Both update directions use tunneling, but field direction, selection and biased regions must still be distinguished. Single poly describes layer count; the division of work among coupling, tunneling and read devices explains this IP's update path.
eMemory · MTP IP
Compare hot-carrier programming of the p-type floating-gate cell with FN electron transfer toward a dedicated erase gate. Both operations act on the same storage node.
The same storage cell supports programming, electrical erase and subsequent programming. ERS restores a window suitable for another program operation; it does not require every carrier to disappear. Whether the host issues a separate erase command depends on the macro or component interface.
PGM → ERS → PGM
Single-poly p-type FG-MOSFET related to NeoBit, with an additional erase gate EG. Dielectric separates EG and FG; they are not shorted. Dielectric isolates FG charge; EG provides an FN exit under erase conditions.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Hot-hole-induced electron injection
Carriers: hot holes induce hot electrons, which cross oxide into FG.
Do not draw injected holes in FG or substitute conventional nMOS CHE; historical edge n+ EG requires separate attribution. SL, SG, BL and well contact follow the pMOS concept; EG is a published erase terminal. Current EG doping, geometry and voltages are not fully disclosed. These are teaching terminals, not an official pin table.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
EG removes electrons for later reinjection; the OTP electrical-erase boundary does not apply.
Do not draw injected holes in FG or substitute conventional nMOS CHE; historical edge n+ EG requires separate attribution. SL, SG, BL and well contact follow the pMOS concept; EG is a published erase terminal. Current EG doping, geometry and voltages are not fully disclosed. These are teaching terminals, not an official pin table.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Channel-current sensing
Sense output current while preserving the stored state; the macro defines logic coding.
Do not draw injected holes in FG or substitute conventional nMOS CHE; historical edge n+ EG requires separate attribution. SL, SG, BL and well contact follow the pMOS concept; EG is a published erase terminal. Current EG doping, geometry and voltages are not fully disclosed. These are teaching terminals, not an official pin table.
CHI/FN is the operation-pair shorthand used here. eMemory describes programming as channel-hot-hole-induced hot-electron injection and labels it CHEI. The erase gate provides a separate electron exit, distinguishing this cell from NeoBit's normal OTP interface and NeoEE's FN/FN mechanism.
Yield Microelectronics (YMC) · MTP IP
YMC publicly identifies a logic-process MTP family. The CHI/BBHH sequence below is an independent mechanism illustration, not evidence that a current ymtp product uses BBHH. Separate product capability from an illustrative 1T1C model.
The same storage cell supports programming, electrical erase and subsequent programming. ERS restores a window suitable for another program operation; it does not require every carrier to disappear. Whether the host issues a separate erase command depends on the macro or component interface.
PGM → ERS → PGM
One nMOS and one functional coupling capacitor share FG. Official information supports ymtp and the 1T1C family; this original equivalent drawing does not assert current product junctions, wells or dimensions. FG and CG have no DC short.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Channel Hot-Carrier Injection; Electrons Enter FG in This nMOS Model
CG controls the channel through the coupling capacitor. Electrons accelerate near the model high-field end; a fraction crosses the dielectric, producing higher Vth and lower read current.
CG and FG remain separated by the capacitor dielectric; an electron channel forms between the model source and drain.
Blue arrows follow electrons; green arrows show conventional current in the opposite direction. Only a fraction acquires sufficient injection energy.
The local vertical field is an explicit injection condition, not implied by D* relative to S alone. Electrons cross the dielectric into FG and increase negative charge; this is neither oxide rupture nor a CG-to-FG DC connection.
At the same read bias, the effective nMOS has higher Vth and smaller I_R: the sensed consequence of its charge state.
An independent CHI / BBHH teaching model; cited sources do not establish BBHH in current YMC ymtp products. It uses an equivalent 1T1C and qualitative directions rather than a current version-specific cross-section or bias table. Independent original research supports BBHH physics; YMC FN/DAHHI patent variants retain their distinct mechanisms.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Band-to-Band Carrier Generation Followed by Hot-Hole Injection into FG
A high-field region in silicon first produces electron/hole pairs. Some holes then cross the dielectric into FG. Silicon BBT and hot-carrier injection into FG are distinct physical steps.
Terminal relationships describe the model junction condition but do not alone guarantee the local vertical field. E⊥ toward FG is a separate condition; D* is not a current ymtp terminal specification.
Electrons are collected by the high-field junction while holes move toward the channel/body side. This tunneling occurs in silicon, not across the FG oxide.
Under this local field condition, a few red holes cross the dielectric into FG and reduce negative charge. Silicon BBT generation and subsequent dielectric injection each require suitable conditions; neither FN electron removal nor DAHHI avalanche generation is substituted.
Vth is lower and I_R is larger at the same read bias. This qualitative direction does not guarantee neutrality, a fixed endpoint or self-convergent erase.
An independent CHI / BBHH teaching model; cited sources do not establish BBHH in current YMC ymtp products. It uses an equivalent 1T1C and qualitative directions rather than a current version-specific cross-section or bias table. Independent original research supports BBHH physics; YMC FN/DAHHI patent variants retain their distinct mechanisms.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
nMOS Threshold Modulation and Reference-Current Sensing
A and B are fixed alternative initial states, not sequential updates. At the same read bias, A has more negative charge, higher Vth and smaller current; B has the opposite. Reading preserves either state.
All four frames retain the same two-state comparison. There is no A-to-B program or erase and no carrier transfer across the dielectric.
This is an independent read example for A. Higher Vth gives smaller I_R; reading does not add FG electrons.
B is an alternative example, not a conversion from A in the previous frame. Lower Vth gives larger I_R; the green arrow denotes conventional current.
Iref must separate the states with adequate margin; the comparison is unchanged from earlier frames. Process, temperature and usage history affect the window; product specifications define margins and 0/1 encoding.
An independent CHI / BBHH teaching model; cited sources do not establish BBHH in current YMC ymtp products. It uses an equivalent 1T1C and qualitative directions rather than a current version-specific cross-section or bias table. Independent original research supports BBHH physics; YMC FN/DAHHI patent variants retain their distinct mechanisms.
This is the CHI/BBHH mechanism model selected for this course. Public YMC material supports its logic-process MTP IP positioning; independent primary research supports the BBHH physics. The figure is not identified as a complete cross-section of a current ymtp version. FN, drain-avalanche hot-hole injection and band-to-band hot-hole injection are distinct paths, even when related patents share an assignee.
Impinj → Virage Logic → Synopsys · MTP IP
Follow the named 2009 AEON company account: electrons enter and leave FG by FN, then a read MOS senses the state. Business and brand succession have a separate timeline.
The same storage cell supports programming, electrical erase and subsequent programming. ERS restores a window suitable for another program operation; it does not require every carrier to disappear. Whether the host issues a separate erase command depends on the macro or component interface.
PGM → ERS → PGM
Public FN/FN functional model: coupling role C, isolated floating gate FG, tunnel counterparts and a read MOS. Floating-gate product positioning and named FN program/erase evidence are attributed separately. This original functional reconstruction leaves undisclosed physical wiring unspecified.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
FN program in the 2009 AEON company account: electrons cross dielectric into isolated FG.
The initial charge is a relative illustration, not a claim that FG must be neutral.
Dielectric field enables FN; channel hot-electron injection is not substituted. C couples to FG through dielectric only.
FN crosses dielectric. No metallic short or external DC injection wire is drawn into FG.
The read device senses charge-dependent behavior. No fixed charge-to-ON or charge-to-logic-1 mapping is assumed.
T_P/T_E are operating roles and must not be assumed to be one physical pin. MOS polarity, device count, wells, voltages and logic coding are unspecified. Do not substitute early Impinj hot-electron patents or the internal cell of every current Synopsys MTP product.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
The same named source identifies FN erase; this model represents reverse updating by removing FG electrons.
Reverse updating begins from the stored charge state.
T_E denotes the erase tunnel counterpart. Its physical relationship to T_P is not fully disclosed by these sources.
Erase is represented as electron removal, not neutralization by injected holes.
Electrical erase and reprogramming enable MTP. Particle counts do not imply endurance, speed or retention specifications.
T_P/T_E are operating roles and must not be assumed to be one physical pin. MOS polarity, device count, wells, voltages and logic coding are unspecified. Do not substitute early Impinj hot-electron patents or the internal cell of every current Synopsys MTP product.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
MOS current sensing
Sense the read MOS under low stimulus; all frames retain the same charge without assigning p/n polarity or logic coding.
Isolated charge affects MOS behavior. MOS polarity is unspecified, so no fixed ON/OFF mapping is imposed.
Read conditions do not establish an FN update field. A and B are teaching terminals, not a macro pin table.
Current flows through the read device, without transporting stored charge through FG or the tunnel dielectric.
Reference strategy, differential implementation and 0/1 coding belong to a specific macro. Automotive differential-cell options are not generalized to all AEON products.
T_P/T_E are operating roles and must not be assumed to be one physical pin. MOS polarity, device count, wells, voltages and logic coding are unspecified. Do not substitute early Impinj hot-electron patents or the internal cell of every current Synopsys MTP product.
AEON is a named logic-process MTP family originating at Impinj. A 2009 Virage Logic company article explicitly supports FN program and erase. The diagram retains that physical scope through C, T_P, T_E and read-MOS roles, without assuming undisclosed p/n polarity, device count or current wiring.
Actt (CMT lineage) · MTP IP
Actt acquired CMT in 2016. The current public MTP product is LogicFlash: logic-compatible, 0–1 extra mask, Flash-like byte program and sector/chip erase. The storage-node material is unpublished.
The same storage cell supports programming, electrical erase and subsequent programming. ERS restores a window suitable for another program operation. Whether the host issues a separate erase command depends on the macro interface.
PGM → ERS → PGM
A logic-process MTP macro. The public page guarantees Flash-like byte program and sector/chip erase, not the storage-node material. CMT is the 2016 acquisition lineage.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Public evidence stops at a Flash-class PGM/ERS interface and does not specify FN, HCI, or a trap layer.
Establish the programmed state by the published mechanism without splicing unpublished bias tables.
The vendor proves Flash-like byte program, not a cell cross-section.
The drawing marks the interface command and does not invent an FN or HCI path.
10k cycles is a product-page ceiling narrative, not a guarantee in this drawing.
SuperMTP is marked under development on the vendor page and is not used in this unit’s operation drawings.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Public evidence stops at a Flash-class PGM/ERS interface and does not specify FN, HCI, or a trap layer.
Electrically erase by the published mechanism so the cell returns to a reprogrammable window.
Public erase granularity is sector or chip, not a proven bit-level erase.
The storage-node material remains unpublished.
CMT lineage explains origin and adds no physics.
SuperMTP is marked under development on the vendor page and is not used in this unit’s operation drawings.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Public evidence stops at a Flash-class PGM/ERS interface and does not specify FN, HCI, or a trap layer.
Sense the retained state under product read conditions, then latch and isolate.
Read does not invent a carrier mechanism in this drawing.
Current difference is decoded by the product sense circuit.
Read-disturb limits must be checked on the target macro.
SuperMTP is marked under development on the vendor page and is not used in this unit’s operation drawings.
CMT is a lineage name, not a current SKU. SuperMTP is marked under development and is not used here. Public evidence stops at the interface and update granularity; do not invent FN or HCI.
NSCore · MTP IP
TwinBit Gen-2 uses a Pch Schottky storage device with zero extra mask. Program is hot hole; erase is hot electron. Sibling PermSRAM is OTP and is not drawn in this unit.
The same Schottky storage cell supports programming, electrical erase and subsequent programming. Erase uses hot electrons to compensate or remove the hot-hole effect and restore a reprogrammable window.
PGM → ERS → PGM
Gen-2 uses a Pch Schottky transistor as the storage device with zero extra mask. PermSRAM is a sibling OTP that traps hot carriers in a SiN spacer and is not drawn here.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Program by hot hole; erase by hot electron.
Establish the programmed state by the published mechanism without splicing unpublished bias tables.
Gen-2 is published as Pch Schottky with zero extra mask.
Arrows show hot-hole direction, not a bias table.
The PermSRAM SiN spacer is not drawn.
TwinBit is not drawn as PermSRAM spacer trapping, and bias numbers are unspecified.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Program by hot hole; erase by hot electron.
Electrically erase by the published mechanism so the cell returns to a reprogrammable window.
TwinBit is electrically erasable and is kept separate from OTP PermSRAM.
The published statement is erase by hot electron.
40–22 nm is a vendor node narrative, not a measurement in this drawing.
TwinBit is not drawn as PermSRAM spacer trapping, and bias numbers are unspecified.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Program by hot hole; erase by hot electron.
Sense the retained state under product read conditions, then latch and isolate.
Read does not repeat hot-carrier program.
Threshold shift changes current.
The read window is set by supplier conditions.
TwinBit is not drawn as PermSRAM spacer trapping, and bias numbers are unspecified.
TwinBit and PermSRAM share a vendor, not a cell physics. Do not draw OTP SiN-spacer hot-carrier trapping as the MTP erase path.
Floadia · MTP IP
Zero extra-mask floating-gate MTP. The vendor states both program and erase use FN; news names the storage node as a floating gate.
The same floating-gate cell supports programming, electrical erase and subsequent programming. Both directions use FN, not hot-carrier program.
PGM → ERS → PGM
Zero extra-mask floating-gate MTP; program and erase use FN. The teaching drawing shows only an equivalent FG and coupling terminal.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Electrons enter and leave the floating gate by FN tunneling.
Establish the programmed state by the published mechanism without splicing unpublished bias tables.
ZT news names the storage node as a floating gate.
The vendor states both program and erase use FN.
Cycle counts do not follow the page’s conflicting numbers.
Cycle-count copy on the page conflicts with itself; no count is treated as a guarantee.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Electrons enter and leave the floating gate by FN tunneling.
Electrically erase by the published mechanism so the cell returns to a reprogrammable window.
Erase remains FN, not hot carrier.
The teaching drawing does not specify well-potential numbers.
Zero extra mask is an integration claim, not a cross-section proof.
Cycle-count copy on the page conflicts with itself; no count is treated as a guarantee.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Electrons enter and leave the floating gate by FN tunneling.
Sense the retained state under product read conditions, then latch and isolate.
Read does not use program-level FN.
Channel current reflects FG charge.
Automotive narrative must be checked against target product conditions.
Cycle-count copy on the page conflicts with itself; no count is treated as a guarantee.
ZT is rewritable MTP, not LEE Fuse ZA Anti-fuse. Cycle-count copy on the page conflicts with itself; no count is treated as a common guarantee.
Floadia · eFlash IP
SONOS charge-trap eFlash with 2–3 extra masks and FN program/erase. O-N-O is a teaching stack, not a measured thickness.
The same SONOS cell supports programming, electrical erase and subsequent programming. Electrons enter and leave the nitride trap layer by FN and restore a reprogrammable window.
PGM → ERS → PGM
SONOS charge-trap eFlash; 2–3 extra masks; FN program/erase. O-N-O is a teaching stack, not a measured thickness.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Electrons enter and leave the nitride trap layer by FN tunneling.
Establish the programmed state by the published mechanism without splicing unpublished bias tables.
G1 is published as SONOS with 2–3 extra masks.
Program and erase both use FN.
O-N-O thickness is unpublished.
A BCD narrative cannot be extrapolated to every logic node.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Electrons enter and leave the nitride trap layer by FN tunneling.
Electrically erase by the published mechanism so the cell returns to a reprogrammable window.
Erase is not TwinBit hot-hole compensation.
A BCD narrative cannot be extrapolated to every logic platform.
Mask count is integration cost, not an endurance guarantee.
A BCD narrative cannot be extrapolated to every logic node.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Electrons enter and leave the nitride trap layer by FN tunneling.
Sense the retained state under product read conditions, then latch and isolate.
Read does not use program-level FN.
Trapped charge shifts threshold.
Read disturb must be checked on the target macro.
A BCD narrative cannot be extrapolated to every logic node.
G1 is eFlash/SONOS, not zero-mask MTP ZT and not Anti-fuse ZA. A BCD narrative cannot be extrapolated to every logic node.
Floadia · eFlash IP
A SONOS storage cell sandwiched by switch transistors, with four extra masks. The vendor emphasizes VDD read without high voltage on diffusion terminals, and marks ongoing development.
The same sandwiched SONOS cell supports programming, electrical erase and subsequent programming. Side switches select the cell; storage remains in the nitride trap layer.
PGM → ERS → PGM
A SONOS storage cell sandwiched by switch transistors; 4 extra masks; read at VDD without high voltage on diffusion terminals.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Program/erase still use charge trapping; read selects through side switches at VDD.
Establish the programmed state by the published mechanism without splicing unpublished bias tables.
G2 sandwiches SONOS between switch transistors.
Four extra masks is the published integration number.
The vendor marks ongoing development.
The vendor marks ongoing development; this drawing teaches only the published structural principle.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Program/erase still use charge trapping; read selects through side switches at VDD.
Electrically erase by the published mechanism so the cell returns to a reprogrammable window.
Erase remains charge-trap physics, not a fuse.
No high voltage on diffusion is a vendor read/logic claim; erase may still need internal HV generation.
An in-development note is not a production guarantee.
The vendor marks ongoing development; this drawing teaches only the published structural principle.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Program/erase still use charge trapping; read selects through side switches at VDD.
Sense the retained state under product read conditions, then latch and isolate.
The vendor emphasizes read at VDD.
Diffusion terminals do not take program-level high voltage.
Non-volatilized logic is product positioning, not proof of a generic standard-cell library.
The vendor marks ongoing development; this drawing teaches only the published structural principle.
G2’s read claim is VDD plus side switches; it does not cancel charge-trap physics. An in-development note is not a production guarantee and is not a generic standard-cell library.
SST / Microchip · eFlash IP
Split-gate Flash: a select gate beside a floating gate. Program uses source-side injection; erase uses interpoly FN.
The same split-gate cell supports programming, electrical erase and subsequent programming. Erase lets electrons leave FG through the interpoly oxide and restores a reprogrammable window.
PGM → ERS → PGM
Split-gate Flash: a select gate beside a floating gate. Program uses source-side injection; erase uses interpoly FN.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Hot electrons inject from the source side into FG; during erase, electrons leave FG through the interpoly oxide.
Establish the programmed state by the published mechanism without splicing unpublished bias tables.
SuperFlash is published as a split-gate architecture.
The program mechanism is source-side injection, not a generic CHE label.
The licensed process range is not proof of one node.
The process range is a licensing narrative, not proof of one production node.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Hot electrons inject from the source side into FG; during erase, electrons leave FG through the interpoly oxide.
Electrically erase by the published mechanism so the cell returns to a reprogrammable window.
Erase uses interpoly FN, not a reverse SSI current.
Electrons cross the oxide between select gate and floating gate.
The teaching drawing is not a foundry metrology cross-section.
The process range is a licensing narrative, not proof of one production node.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Hot electrons inject from the source side into FG; during erase, electrons leave FG through the interpoly oxide.
Sense the retained state under product read conditions, then latch and isolate.
Read does not repeat SSI program.
Channel current reflects FG charge.
Read-speed ratings must be checked on the licensed target version.
The process range is a licensing narrative, not proof of one production node.
SuperFlash program is SSI, not a generic CHE label; erase is inter-gate FN, not channel erase. A licensed process range is not proof of one production node.
Numem · MRAM IP
Connect foundry-standard STT-MRAM cells to embedded IP by identifying the magnetic junction, access transistor, bit line, source line and sensing path.
MRAM overwrites existing data by changing magnetic state, without a Flash-style erase-before-program step. Clearing to all zeros or ones is a series of target-state writes; the P/AP-to-data mapping is product-specific.
P ⇄ AP
FL/tunnel barrier/RL represent STT functions; A/B are teaching terminals. WL/BL/SL follow the 2019 Numem architecture without claiming a layer-to-line mapping.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
MTJ free-layer magnetization stores information
Write free-layer magnetization through spin-transfer torque.
WL is off and the cell retains AP; this sequence writes P.
WL turns on and teaching drive A/B crosses the MTJ; electron and conventional-current arrows oppose each other. The actual BL/SL layer mapping requires the PDK.
Magnetization reaches P; the intermediate angle is not a measured trajectory or deterministic switching time.
Turn WL off and remove bias to retain the moment; the other drive overwrites the opposite data without a floating-gate erase step.
This reconstructs the public Numem IP architecture for teaching. Current sources do not disclose materials, thicknesses, vertical order, write-terminal polarity, or logic encoding. Directions A/B mean two calibrated opposite drives. The 2019 forced-current read is not a specification for every product.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
MTJ free-layer magnetization stores information
Use the opposite MTJ drive to overwrite magnetization.
WL is off and the cell retains P; this sequence overwrites AP.
WL turns on and teaching drive A/B crosses the MTJ; electron and conventional-current arrows oppose each other. The actual BL/SL layer mapping requires the PDK.
Magnetization reaches AP; the intermediate angle is not a measured trajectory or deterministic switching time.
Turn WL off and remove bias to retain the moment; the other drive overwrites the opposite data without a floating-gate erase step.
This reconstructs the public Numem IP architecture for teaching. Current sources do not disclose materials, thicknesses, vertical order, write-terminal polarity, or logic encoding. Directions A/B mean two calibrated opposite drives. The 2019 forced-current read is not a specification for every product.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
MTJ free-layer magnetization stores information
Select the same cell, sense its retained resistance with a small stimulus, then latch and isolate.
The same cell starts in retained P with WL off; reading does not first reverse its moment.
Following the 2019 architecture, a small forced current produces a voltage including access-path resistance; P voltage is below AP at equal current.
After the sensor latches, WL turns off; free/reference layers remain P without a read-restore cycle.
This reconstructs the public Numem IP architecture for teaching. Current sources do not disclose materials, thicknesses, vertical order, write-terminal polarity, or logic encoding. Directions A/B mean two calibrated opposite drives. The 2019 forced-current read is not a specification for every product.
Numem can use foundry-standard STT cells while integrating its layout, circuits and memory architecture. Understand the two junction states and access/sense paths before considering macro behavior; a control-architecture improvement is not a new storage mechanism.
GLOBALFOUNDRIES · MRAM IP
Use a publicly reported 22FDX research cell to examine 1T1MTJ, free and reference layers, and bidirectional switching under the source's current convention.
MRAM overwrites existing data by changing magnetic state, without a Flash-style erase-before-program step. Clearing to all zeros or ones is a series of target-state writes; the P/AP-to-data mapping is product-specific.
P ⇄ AP
The 2024 research uses CoFeB free/reference layers, a tunnel barrier, SAF pinning, and an access transistor. Their vertical placement here defines a drawing coordinate.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
P/AP magnetization and resistance in 1T1MTJ
Write free-layer magnetization through spin-transfer torque.
WL is off and the cell retains AP; this sequence writes P.
WL turns on; conventional current RL-to-FL applies STT, with opposite electron flow.
Magnetization reaches P; the intermediate angle is not a measured trajectory or deterministic switching time.
Turn WL off and remove bias to retain the moment; the other drive overwrites the opposite data without a floating-gate erase step.
Polarity follows this paper: positive Ic flows RL-to-FL and writes P; reverse writes AP. This sign convention and recipe are not universal to MRAM. Obtain BL/SL layer mapping and values from the PDK; undisclosed barrier material and exact thicknesses are omitted.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
P/AP magnetization and resistance in 1T1MTJ
Use the opposite MTJ drive to overwrite magnetization.
WL is off and the cell retains P; this sequence overwrites AP.
WL turns on; conventional current FL-to-RL applies STT, with opposite electron flow.
Magnetization reaches AP; the intermediate angle is not a measured trajectory or deterministic switching time.
Turn WL off and remove bias to retain the moment; the other drive overwrites the opposite data without a floating-gate erase step.
Polarity follows this paper: positive Ic flows RL-to-FL and writes P; reverse writes AP. This sign convention and recipe are not universal to MRAM. Obtain BL/SL layer mapping and values from the PDK; undisclosed barrier material and exact thicknesses are omitted.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
P/AP magnetization and resistance in 1T1MTJ
Select the same cell, sense its retained resistance with a small stimulus, then latch and isolate.
The same cell starts in retained P with WL off; reading does not first reverse its moment.
WL enables a low-bias current through the MTJ/access device; P has greater current than AP at equal bias.
After the sensor latches, WL turns off; free/reference layers remain P without a read-restore cycle.
Polarity follows this paper: positive Ic flows RL-to-FL and writes P; reverse writes AP. This sign convention and recipe are not universal to MRAM. Obtain BL/SL layer mapping and values from the PDK; undisclosed barrier material and exact thicknesses are omitted.
A foundry embedded macro joins the magnetic junction to logic processing, the access transistor and reliability conditions. This figure retains the named study's materials and polarity convention for step-by-step reading; those details do not automatically describe every current 22FDX memory version.
Weebit Nano · ReRAM IP
Follow oxygen exchange, a defect-related conduction path and access-transistor current compliance in a public silicon-oxide research structure to understand embedded ReRAM SET, RESET and read.
RESET is the reverse data-state update from low to high resistance; a later SET restores low resistance. This is reversible resistance switching rather than Flash-style block erase. The circuit defines logical zero/one encoding.
SET → RESET → SET
Uses the coauthored CEA 130nm 1T1R: Ti top electrode, SiOx switching layer, and TiN bottom electrode. The access transistor selects and limits current.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Oxygen-ion exchange and an oxygen-vacancy conduction path
Positive TE bias restores the conduction path and produces low resistance.
Start in a formed HRS with a local BE-side gap; forming is not repeated on every cycle.
Positive TE drives O²− toward Ti for interfacial exchange; the access transistor limits current.
The oxygen-deficient conduction path reconnects; current rises under compliance.
After removing bias and WL selection, the path retains LRS.
This is the public Weebit/CEA-Leti/Silvaco research model, not a product recipe for every foundry node. SET: positive TE, O²− toward Ti. RESET: negative TE, oxygen returns into SiOx and recombines near the BE-side path. Forming is an initial condition, not every write.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Oxygen-ion exchange and an oxygen-vacancy conduction path
Reverse TE bias interrupts the conduction path and produces high resistance.
A vacancy path already exists; this operation changes LRS to HRS.
TE is negative relative to BE; oxygen returns from the Ti interface into SiOx. Blue arrows denote oxygen motion.
Oxygen recombines with vacancies and opens a critical BE-side gap; RESET does not restore the entire layer to its as-fabricated material.
HRS remains after bias removal; vacancies and interfacial oxygen may remain.
This is the public Weebit/CEA-Leti/Silvaco research model, not a product recipe for every foundry node. SET: positive TE, O²− toward Ti. RESET: negative TE, oxygen returns into SiOx and recombines near the BE-side path. Forming is an initial condition, not every write.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Oxygen-ion exchange and an oxygen-vacancy conduction path
Select the same cell, sense its retained resistance with a small stimulus, then latch and isolate.
The same cell starts in retained LRS with selection off. HRS can follow the same read sequence.
A small bias senses the vacancy path; ILRS > IHRS at equal bias, without using the read pulse to rearrange oxygen.
After latching, remove bias and retain the original path; actual read-disturb limits remain supplier-specific.
This is the public Weebit/CEA-Leti/Silvaco research model, not a product recipe for every foundry node. SET: positive TE, O²− toward Ti. RESET: negative TE, oxygen returns into SiOx and recombines near the BE-side path. Forming is an initial condition, not every write.
In this named research example, the storage medium, oxygen-exchange electrode and access transistor jointly shape switching. Current compliance and read stimulus are part of cell operation. Materials and recipes remain scoped to the cited implementation.
Crossbar · ReRAM IP
Read Crossbar's public patent and historical embedded-macro materials through metallic-path extension, retraction and low-stimulus sensing.
RESET is the reverse data-state update from low to high resistance; a later SET restores low resistance. This is reversible resistance switching rather than Flash-style block erase. The circuit defines logical zero/one encoding.
SET → RESET → SET
Selects the Ag/amorphous-Si/p+ poly-Si embodiment of US20120007035A1; an access transistor represents the separately published embedded 1T1R integration.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Extension/retraction from an upper metal region changes interparticle tunneling
Positive TE bias restores the conduction path and produces low resistance.
Forming has established an upper metal region; the HRS particle path does not yet extend effectively toward the lower contact.
Positive TE bias extends the path from the upper metal region toward BE, following this patent rather than assuming generic upward cathodic nucleation.
Closer neighboring metal particles strengthen tunneling conduction; the dots do not claim a fully solid silver bridge.
Turn selection off and remove bias to retain the low-R path.
This is a published patent embodiment associated with historical embedded IP, not proof of current macro recipes or newly licensable nodes in 2026. The patent describes metal particles and interparticle tunneling; the path is not equated to a solid silver bridge or generic cathode-nucleated ECM.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Extension/retraction from an upper metal region changes interparticle tunneling
Reverse TE bias interrupts the conduction path and produces high resistance.
Start with the existing low-R particle path; the upper metal region and narrow path are shown separately.
Negative TE bias retracts or disconnects the narrow particle path toward the upper metal region; each particle charge state is unspecified.
The effective lower-side spacing increases and tunneling current falls; the upper residual metal region remains.
HRS remains after bias removal. This is reverse RESET, without a preceding block-erase cycle.
This is a published patent embodiment associated with historical embedded IP, not proof of current macro recipes or newly licensable nodes in 2026. The patent describes metal particles and interparticle tunneling; the path is not equated to a solid silver bridge or generic cathode-nucleated ECM.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Extension/retraction from an upper metal region changes interparticle tunneling
Select the same cell, sense its retained resistance with a small stimulus, then latch and isolate.
The same cell starts in retained LRS with selection off. HRS can follow the same read sequence.
A small bias senses the metal-particle path; ILRS > IHRS at equal bias and transport can involve interparticle tunneling.
After latching, remove bias and retain the original path; actual read-disturb limits remain supplier-specific.
This is a published patent embodiment associated with historical embedded IP, not proof of current macro recipes or newly licensable nodes in 2026. The patent describes metal particles and interparticle tunneling; the path is not equated to a solid silver bridge or generic cathode-nucleated ECM.
This figure follows the named patent's metal/amorphous-silicon model. Do not substitute another ECM cell's nucleation direction or an ideal continuous silver bridge. Evidence for the embedded macro and for the historical cell embodiment is identified separately.
ENGINEERING COMPARISONS
Start with a named device, macro or system. Read each value alongside its measurement conditions; these examples do not form a ranking across implementation levels.
NAMED LOGIC-PROCESS IPs Historical Course Table and Corrections
NAMED LOGIC-PROCESS IPs
The Yu 2016 paper and the 2021 course table compare SRAM, DRAM, NOR, NAND, PCM, RRAM, STT-MRAM, SOT-MRAM, FeRAM and FeFET. The logic-process OTP / MTP / eFlash IPs below are not columns of that table, and they do not share a unified voltage or cycle number that could be pasted into it. This section only maps each named IP onto a selection-matrix family leaf and links to its cell study.
Open the 11-leaf selection matrixHistorical Course Table and Corrections
A dielectric is permanently made conductive; one-time program. The teaching leaf is not floating-gate OTP and not a blown fuse.
Blow-class programs by electromigration rupture of a link; I-fuse is listed separately and is not the same physics.
The family leaf uses NeoEE / LEE Flash ZT as the FN/FN reference. Actt and TwinBit do not inherit that cell physics.
The HD-MTP leaf is CHI/FN floating gate. Schottky TwinBit belongs to neither leaf.
The split-gate eFlash leaf is this family, not SONOS.
Charge trapping. G1 and G2 are both SONOS but differ in mask count, read path and array organization; they are not one specification cell.
Historical-table numbers remain the course baseline. Biases, cycle counts and production nodes for named IPs stay bound to each cell study and its VERIFY limits; this page does not invent datasheet figures.
These examples span products, research macros, dies, and system devices. Their values explain measurement boundaries; they do not establish a ranking across those levels.
FeRAM/F-RAM · Complete Product
Toggle MRAM · Complete Product
Oxide ReRAM · Complete Product
STT-MRAM · Array / Research Design
3D NAND / QLC · ISSCC 2026 Die Demonstration
3D XPoint / Optane · Complete SSD and Host Platform
HISTORICAL REFERENCE
This historical comparison table is reconstructed from the screenshot of slide 14 in the course dated 2021-11-01. That slide cites Yu and Chen's 2016 paper and adds SOT-MRAM, FeRAM, FeFET, and several numerical updates, so the complete table cannot be called the original 2016 table. The course values are retained as a baseline for learning how to compare technologies; they are not uniform product specifications for each family in 2026. F is the lithographic feature size used in the course, and energy is estimated at the cell level. The original slide also describes the values as representative, rather than best-case or worst-case figures.
| Metric | SRAM | DRAM | NOR Flash | NAND Flash | PCM | RRAM | STT-MRAM | SOT-MRAM | FeRAM | FeFET |
|---|---|---|---|---|---|---|---|---|---|---|
| Cell Area | >150 F² | 6 F² | 10 F² | <4 F² (3D) | 4–50 F² | 4–50 F² | 6–50 F² | 12–100 F² | 6–50 F² | 6–50 F² |
| Bits per Cell | 1 | 1 | 2 | 3–4 | 2–3 | 2–3 | 1 | 1 | 1 | 2–3 |
| Operating Voltage | <1 V | <1 V | >10 V | >10 V | <3 V | <3 V | <1 V | <1 V | <2 V | <3 V |
| Read Time | ~1 ns | ~10 ns | ~50 ns | ~10 µs | <10 ns | <10 ns | <10 ns | ~1 ns | <100 ns | <50 ns |
| Write Time | ~1 ns | ~10 ns | 10 µs–1 ms | 100 µs–1 ms | ~50 ns | <100 ns | <20 ns | <3 ns | <100 ns | <100 ns |
| Data Retention | N/A | ~64 ms | >10 years | >10 years | >10 years | >10 years | >1 years | >1 years | >10 years | >1 years |
| Endurance | >10¹⁶ cycles | >10¹⁶ cycles | ~10⁵ cycles | 10³–10⁴ cycles | 10⁶–10⁹ cycles | 10³–10⁹ cycles | 10⁶–10¹⁴ cycles | ~10¹² cycles | 10⁹–10¹² cycles | 10⁶–10⁹ cycles |
| Write Energy | ~fJ-scale/bit | ~10 fJ/bit | 100 pJ/bit | ~10 fJ/bit | ~10 pJ/bit | ~pJ-scale/bit | ~pJ-scale/bit | ~pJ-scale/bit | ~100 fJ/bit | ~fJ-scale/bit |
Cell Area: The course's area notation is retained. The original slide also states that PCM, RRAM, and FeFET could reach below 4 F² through 3D integration. This is not a strictly equivalent comparison of physical cell area, effective area per bit, and complete-die density within one process.
Bits per Cell: The course's representative values are retained. A multibit research demonstration does not establish the usable number of bits in a specified product under its temperature, endurance, and retention conditions.
Operating Voltage: The course does not distinguish supply, read, write, forming, and internally boosted voltages in every cell of the table. Updated product comparisons must separate these quantities; they cannot be compared directly with a part number's VDD.
Read Time: These are representative time scales from the course. They are not tied to common capacity, peripheral circuits, sensing methods, or interface completion points.
Write Time: An updated comparison must separately record erase, SET/RESET, pulse duration, verification and retries, and total time from entry into a data buffer to completion of the nonvolatile write.
Data Retention: The original table does not specify temperature, prior cycling, or extrapolation conditions for every entry. SRAM requires power. The time scale for dynamic storage or refresh in DRAM is not the same guarantee as power-off retention in NVM.
Endurance: Historical ranges are retained cell by cell. The sources do not use a common bit/page/block granularity, definition of a read/write cycle, error threshold, ECC policy, temperature, or retention test.
Write Energy: The original slide explicitly limits these estimates to the cell level, excluding the complete array periphery. fJ denotes 10⁻¹⁵ J and pJ denotes 10⁻¹² J. Entries without a numerical coefficient remain order-of-magnitude estimates; no precise value is invented.
The emerging-technology columns in the 2016 paper contain only STT-MRAM, PCRAM, and RRAM. The 2021 course adds SOT-MRAM, FeRAM, and FeFET.
Distinguish the original paper from the lecturer's later additions so that the provenance remains traceable.
Three differences have been verified: SRAM area changes from >100 F² to >150 F²; NAND bits per cell change from 3 to 3–4; and STT-MRAM endurance changes from >10¹⁵ to 10⁶–10¹⁴ cycles.
These differences show that the course table was updated, but a representative value from any particular year is not a guarantee for an entire technology family.
As of 2026-09-10, these families have verifiable commercial products or implementations in volume production: CY15B104QSN, MR25H40/Everspin STT-MRAM, MB85AS8MT, and ST SR6P6C8.
Maturity must be tied to the subtechnology, materials, process, and specific part number, not assigned solely by the family name.
Separate cell switching, array sensing, macro completion, and host-visible latency. The MB85AS8MT specifies product tWC of 5 ms typical and 10 ms maximum, and write-completion status must be checked.
Buffering, internal programming, and control sequences can all add time.
25 ns is the clock period. A complete transaction also includes the instruction, address, data transfer, and any internal write wait.
Frequency describes the transfer cadence; latency describes how long it takes to complete an operation.
The CY15B104QSN's 151-year rating applies at 65°C; the rating is 38 years at 75°C and 10 years at 85°C. Retention must also be paired with the specified cycling and test conditions.
Temperature and usage history change the probability of state failure. Comparing years alone is insufficient.
Record physical cell footprint, layer count, bits per cell, array efficiency, and complete-die bit density separately.
3D stacking and multibit storage change effective bit density. Peripheral circuits, redundancy, and ECC also occupy area.
Compare cell pulses, wire charging and discharging, selector leakage, charge pumps, verification, ECC, and interface energy at their respective levels.
Low cell energy does not guarantee the same advantage in a large array or a workload with low utilization.
Volume production is a maturity state; emerging describes a development stage or a convention in the literature; SCM is a system role within the memory/storage hierarchy.
The same storage physics can support different maturity levels and applications. The classification dimensions must remain separate.
CXL can carry volatile or persistent memory; Samsung CMM-D uses DRAM. Persistence must be established through the media, backup energy, and platform semantics together.
An interconnect protocol is not storage physics and does not automatically guarantee recovery after failure.
Micron discontinued 3D XPoint development on 2021-03-16; Intel discontinued further Optane development in July 2022. The inventory forecasts and support statements in the 2023 letter must be recorded separately.
Development, manufacturing, inventory shipments, warranties, and support are different lifecycle events.
Public Foundry Evidence · 2026-09-10
A node can support consumer, automotive, high-retention, high-endurance, or high-speed variants. Symposium announcements and targets are marked complete only when subsequent annual reports, technology pages, or product documents confirm them.
Newest event years first. The count shows entries, not shipments or qualification results.
The March announcement stated Grade 1 ready, PDK availability, and specified performance. Volume production in Dresden was targeted for the second half of 2026.
Evidence BoundaryAs of September 10, this review had not obtained an announcement confirming volume production of the new version. Entering the second half of the year does not establish volume production.
Retains the 2026 volume-production target from the 2025 announcement; the current FDX page continues to describe RRAM.
Evidence BoundaryNo explicit completion announcement was obtained in this review. A feature listed on a product page is not evidence of its completion date.
As of the verification date, official product text lists 22/16nm as automotive-qualified and in production, with 12nm automotive and 5nm high-write-speed versions in development.
Evidence BoundaryThis is a dynamic current-status snapshot. 2026 is not assigned as the first volume-production year of each platform; not all performance conditions are published.
As of the verification date, official text lists 40/28/22/12nm in volume production and 6nm in development.
Evidence BoundaryA 12nm automotive qualification date or 6nm volume-production date must not be added without public primary evidence of completion.
The automotive article lists MRAM on both platforms as technologies supporting MCUs.
Evidence BoundaryAn application claim cannot replace production and quality documentation for a 12LP+ macro.
The technology summit announced prototyping availability and preliminary macro design kits, with volume production targeted for 2026.
Evidence BoundaryIts mechanism and version differ from the 2020 Dialog CBRAM collaboration. It is not treated as evidence that the earlier plan was completed.
Passed automotive Grade 1. The annual report states a chip failure rate below 1 ppm after one million cycles.
Evidence BoundaryLimited to the second generation; does not rewrite the earlier record of 16FFC completion in 2023.
N12e completed consumer-grade qualification for production; 22RRAM completed qualification for one hundred thousand cycles; 22ULL RRAM met automotive Grade 1 requirements.
Evidence BoundaryThe three qualification scopes remain separate. One hundred thousand cycles and all automotive conditions must not be automatically combined into a commitment for the same macro.
Faraday announced completion and silicon validation of its RRAM SoC development platform.
Evidence BoundaryController and processor integration evidence does not establish named customer volume shipments.
The official article describes the direction of eMRAM deployment and its use in automotive MCUs.
Evidence BoundaryNo macro-specific qualification or production-completion date is available for verification. MRAM figures from 22FDX must not be applied.
Automotive qualification for the smaller-bit 16nm technology was targeted for 2025; development began at the 12nm and 5nm nodes.
Evidence BoundaryThe earlier 16FFC Grade 1 completion in 2023 remains valid. A target for a new version must not be misrepresented as a delay of the entire node.
The annual report's R&D chapter states that 12nm consumer-grade technology qualification was completed and 6nm entered development.
Evidence BoundaryThe 2025 annual report additionally records consumer-grade qualification for production. Both source statements are retained; no single first-qualification date is assigned independently.
Samsung Foundry Forum materials discuss advanced-node eMRAM direction alongside MBCFET logic roadmaps.
Evidence BoundarySF3/SF2 logic announcements do not by themselves prove shipped eMRAM macros; separate NVM milestones required.
Intel Foundry Direct Connect 2024 describes 18A production preparation, PowerVia BSPDN, and embedded IP ecosystem goals.
Evidence BoundaryRibbonFET/PowerVia logic progress does not automatically prove native AntiFuse, eMRAM, or FeFET NVM at named nodes.
The platform chart continues to show MRAM development. Joint development is documented for 12LP, but this does not confirm embedded MRAM volume production in that year.
Evidence BoundaryFoundry manufacturing of discrete STT-MRAM products and GF's embedded macro availability must be distinguished.
The annual report records completed AEC-Q100 Grade 1 reliability qualification, providing subsequent completion evidence for the target stated in 2022.
Evidence BoundaryThe same year's R&D chapter separately describes a consumer-grade version and a next generation with smaller bits. A precise mapping of macro versions is unavailable.
The R&D chapter explicitly records volume production of 22nm consumer-grade MRAM, completed consumer-grade technology qualification at 16nm, and continued development of a next generation with smaller bits.
Evidence BoundaryThis statement of 22nm production is not used as the sole evidence for the first volume-production year. Distinct 16nm versions are not conflated.
40/28/22nm was in volume production, while 12nm and the next generation remained in development. 22/28ULL was in its second year of volume production.
Evidence BoundaryVolume production of the underlying 12nm logic platform does not establish volume production of embedded RRAM.
8Mb plus 16Kb RRAM IP qualified on the 0.8V/2.5V platform.
Evidence BoundaryThe 16Mb automotive and 0.8V/1.8V versions remained under development in this announcement.
The existing 22FDX production baseline traces back to 2020. MRAM-G2 appears in the 2022 roadmap, but this review has not established its separate qualification or completed volume-production status.
Evidence BoundaryPlanned features shown in a chart must not all be marked as in production.
Completed reliability qualification, with one million cycles and solder-reflow capability. The technology was production-ready; Grade 1 was then targeted for 2023.
Evidence BoundaryDoes not establish completed automotive qualification or customer volume shipments in that year.
Several customers completed product qualification and began volume production.
Evidence BoundaryApplies to specific platforms and customer products; does not establish automotive qualification for every capacity.
Avalanche third-generation P-SRAM announced immediately available.
Evidence BoundaryStandalone-product evidence is not a specification for a general UMC embedded macro.
GF officially announced that eMRAM had entered production.
Evidence BoundaryProvides a production baseline preceding 2022; it does not imply that every subsequent version was completed in 2020.
Samsung announced commercial shipment of 28nm FD-SOI embedded MRAM.
Evidence BoundaryFD-SOI platform with BEOL adders; do not equate with logic-roadmap 0-mask claims or bulk FinFET NVM.
IEDM 2018/2019 publications describe 22FFL embedded STT-MRAM approaching production readiness.
Evidence BoundaryEvidence centers on 22FFL test vehicles; not an open PDK for every customer.
MRAM development with Avalanche starting from 28nm CMOS.
Evidence BoundaryAn agreement does not establish availability of every embedded macro.
Panasonic collaboration targeting samples in 2018.
Evidence BoundarySampling and production were forward plans; existing 180nm production does not establish 40nm production.
Public symposium releases do not always include complete tables for every eNVM macro. The entries below identify the available material for each year; annual reports and product documents cross-check completed milestones.
TSMC's public release establishes only the relevant platform scope; GF's investor platform chart was obtained through indexed text. Node-specific completion status is supported by annual reports and formal product announcements.
TSMC's main symposium release was reviewed, but no complete MRAM roadmap was found. The graphical associations in GF's chart were not fully verified, so the limitations are retained.
TSMC's main release is not a complete memory timeline. GF's official automotive technical article supports the direction of 12LP+ MRAM but provides no volume-production date.
GF's annual summit provides directly citable evidence of RRAM prototyping availability and a 2026 target. TSMC completion events are verified against the subsequent annual report.
Access to TSMC's full videos beyond the public highlights requires an invitation; no access to restricted slides is claimed. GF's March AutoPro150 announcement and its undated current-status page are recorded separately.
One hundred thousand cycles; ten-year retention over −40°C to 125°C; five solder-reflow cycles; silicon-validated macros from 4–48 Mbit.
The announcement states Grade 2 design support; Grade 1 was a development target at the time.
The selectable 4–48 Mbit macro range, endurance, and retention metrics must not be treated as guaranteed in every combination. A complete test matrix is unavailable.
Press release: up to five hundred thousand cycles, reads below 10 ns, and operation up to 150°C. Current page: more than five hundred thousand cycles and twenty-year retention at 150°C.
The sources use different wording, which is retained for each metric. Exact capacity, ECC, failure rate, and read/write conditions require the design kit.
Read latency is not write latency. Qualification for operation at 150°C and twenty-year data retention are different metrics.
No MRAM macro performance specification with complete traceability was obtained in this review.
12LP is a FinFET joint-development platform; official 12LP+ articles describe an automotive offering.
Neither 22FDX figures nor improvements in 12LP+ logic performance are transferred to MRAM performance.
Early 16FFC: one million cycles and solder-reflow capability. Second-generation 16MRAM: chip failure rate below 1 ppm after one million cycles.
Statements from different annual reports and technology generations are kept separate. The second generation completed automotive qualification in 2025.
The public summaries do not disclose complete capacity, temperature, ECC, or sample-distribution information. These results cannot be directly compared with the write error rate of an individual MTJ.
22RRAM completed qualification for one hundred thousand cycles in 2025; N12e completed consumer-grade qualification for production in 2025.
High-endurance options, automotive options, and consumer-grade platforms have distinct qualification scopes.
Without complete macro data, no single shared endurance or retention value is assigned to all RRAM.
16FFC has a documented Grade 1 completion in 2023. Separate automotive milestones for smaller bits and second-generation 16MRAM appear in 2024–2025. Preserve the version distinctions and the limitation that a complete macro mapping is unavailable.
Record the 2024 technology qualification, the 2025 consumer-grade qualification for production, and the product page's volume-production confirmation as of the verification date separately. No unique first volume-production date has been established.
The 2025 summit announced prototyping availability and preliminary design kits. The verifiable 2026 milestone remains a target and requires a subsequent completion announcement.
The conductive-bridge material mechanism and oxide resistive switching cannot be merged under the shared RRAM name. Record each collaboration, platform, version, and schedule separately.
12LP/12LP+ belong to the FinFET family; 22FDX is FD-SOI. Joint development, application marketing, and production-macro qualification require separate evidence.
Production evidence exists for GF 22FDX MRAM and multiple TSMC MRAM/RRAM nodes. New nodes, automotive variants, and high-endurance variants still require their own qualification.
NVM · INDUSTRY & RESEARCH
Trace named products, process platforms and research programs. Compare the implementation, its public maturity and the conditions behind each claim.
81 named routes across 13 families · Reviewed through 2026-09-10
MRAM · Discrete memory supplier
Toggle MRAM
Commercial productsThe current PERSYST catalog lists Toggle production parts. MR3A16ACYS35 is marked MP and specifies 8Mb, asynchronous x16, 35ns, 3.3V and −40 to 85°C.
ScopeThese values belong only to MR3A16ACYS35. Do not mix speed, temperature or automotive ratings across parts. The old MR4A16B datasheet is inaccessible, so its over-20-year retention is not cited.
Undated source; checked 2026-09-10
MRAM · Discrete persistent-memory supplier
STT-MRAM / DDR4-derived
ShippingThe 2025 Form 10-K confirms continuing 1Gb STT-MRAM shipments. The technology page identifies a DDR4-like persistent-DRAM product for enterprise storage.
ScopeDDR4-like does not imply drop-in compatibility with every DDR4 controller. Do not transfer Toggle or xSPI retention, endurance or automotive ratings. The old family URL is now 404.
Source date / event period: 2026-03-04 · Checked 2026-09-10
MRAM · Discrete memory supplier
STT-MRAM / xSPI
HR 64Mb qualified and orderable; HR 128/256Mb scheduled at releaseThe March 5, 2026 investor release confirms HR 64Mb xSPI STT-MRAM completed AEC-Q100 Grade 1 production qualification and is orderable with distributor inventory. HR 128Mb qualification was expected in May and 256Mb in July, with 256Mb volume availability expected in the second half of 2026.
ScopeThe 128Mb and 256Mb dates were forecasts as of March 5, 2026; subsequent completion was not verified. HR qualification does not transfer to other EMxxLX variants. This release does not specify bandwidth or retention.
Source date / event period: 2026-03-05 · Checked 2026-09-10
MRAM · Memory supplier / foundry
pMTJ STT-MRAM
Named product in productionThe September 13, 2022 release announces immediate availability of Gen 3 P-SRAM on UMC 22nm. The cited parallel x32 product specifies over 10^14 writes and 1,000-year retention at 85°C.
ScopeThis is a discrete product, not proof of identical specifications for general UMC embedded macros. Retention is a supplier reliability specification.
Source date / event period: 2022-09-13 · Checked 2026-09-10
MRAM · Memory and technology developer
STT-MRAM
Cell-scaling milestoneThe 2026 web announcement reports completion of a first-phase MTJ scaling milestone for future higher-density space-grade MRAM.
ScopeThe page is dated March 2, 2026, but the body says March 2, 2025. A 16x density increase is a future goal, not a shipped product.
Source date / event period: 2026-03-02 · Checked 2026-09-10
MRAM · Foundry platform
STT-MRAM / eMRAM
28FDS production; FinFET expansionThe current specialty-process page confirms 28nm FD-SOI eMRAM mass production since 2019 and expansion to 14LPU and 8LPU, with 5nm still planned.
ScopePlatform expansion does not establish named high-volume customers at every node. Earlier roadmap dates are not completion evidence.
Undated source; checked 2026-09-10
MRAM · Process and device R&D
STT-MRAM
Published research; current product unverifiedThe official IEDM 2018 program lists Intel-authored work on MRAM embedded in 22FFL FinFET, establishing primary evidence of process-integration research.
ScopeNo newer verified Intel commercial MRAM offering was found in this bounded review. Do not infer availability from the paper or the Intel 16 name.
Source date / event period: 2018-12-04 · Checked 2026-09-10
MRAM · Foundry platform
eMRAM / STT route
Automotive Grade 1 qualifiedThe 2025 annual report confirms qualification and customer availability of second-generation 16nm automotive Grade 1 MRAM. 12nm automotive and 5nm high-speed MRAM remain in development.
ScopeQualification is not proof of volume shipment of a named MCU. SOT research must remain separate from qualified platforms.
Source date / event period: 2025 · Checked 2026-09-10
MRAM · Advanced memory R&D
SOT-MRAM
Research demonstrationThe 2025 annual report describes an IEDM 2025 Type-C SOT-MRAM demonstration using a circular MTJ with built-in magnetic anisotropy for field-free operation.
ScopeA research demonstration is not a commercial process or a shipping SRAM replacement; 16nm automotive MRAM maturity does not transfer.
Source date / event period: 2025-12 · Checked 2026-09-10
MRAM · Foundry platform
STT-MRAM / 22FDX
Platform entered productionThe February 27, 2020 announcement confirms production entry of 22FDX eMRAM and 4–48Mb silicon-validated macros, with 100k endurance and 10-year retention across the stated temperature range.
ScopeGrade 1 was a future target in that release. The current FDX page still lists MRAM, but customer-specific qualification requires separate evidence.
Source date / event period: 2020-02-27 · Checked 2026-09-10
MRAM · MCU supplier
Embedded MRAM
Available; Japanese release confirms productionThe October 22, 2025 release announces available RA8M2 and RA8D2 MCUs with embedded MRAM, a 1GHz Cortex-M85 and a 250MHz Cortex-M33.
ScopeCPU clock is not native MRAM read speed. Do not transfer 2024 research-macro measurements directly to the RA8 products.
Source date / event period: 2025-10-22 · Checked 2026-09-10
MRAM · Automotive MCU supplier
Embedded MRAM
Announced; official brief says preproductionNXP announced the 16nm FinFET S32K5 with embedded MRAM on March 11, 2025. Its October 30, 2025 product brief still labels the family preproduction.
ScopeThe claimed 15x write advantage is a supplier comparison against embedded flash, not an absolute latency. Announcement is not volume-production evidence.
Source date / event period: 2025-10-30 · Checked 2026-09-10
MRAM · Discrete memory supplier
STT-MRAM
Named product datasheetThe March 2024 S3RxxxxR1M datasheet specifies 1–16Mbit STT-MRAM, asynchronous x8/x16 interfaces and an industrial −40 to 85°C range; the website also lists serial products.
ScopeDo not transfer larger densities or process nodes from other series or media reports into this datasheet. Shipment volume is not disclosed.
Source date / event period: 2024-03 · Checked 2026-09-10
MRAM · Magnetics, MTJ and technology R&D
STT-MRAM
Technology and R&D; commercial memory SKU unverifiedTDK's September 1, 2025 investor-day presentation includes STT-MRAM among its spintronics technologies; Headway authors also have public embedded-STT-MRAM research presentations.
ScopeHDD-head production and MTJ expertise do not establish commercial discrete MRAM. No orderable MRAM SKU, PDK or specific foundry commitment was verified.
Source date / event period: 2025-09-01 · Checked 2026-09-10
MRAM · Memory IP and chip/chiplet architecture provider
Foundry-based STT-MRAM
Supplier claims production readiness; shipments unverifiedNumem describes foundry-based STT-MRAM IP and chips/chiplets enhanced by AIME. Its June 10, 2025 announcement claims production readiness.
ScopePower and SRAM-class performance are supplier claims without uniform independent benchmarking. This is not evidence of a new magnetic material or named volume shipments.
Source date / event period: 2025-06-10 · Checked 2026-09-10
MRAM · Research institute / technology R&D
SOT-MRAM
300mm research-device demonstrationOn December 13, 2023 imec reported roughly 50nm critical-dimension SOT devices on 300mm wafers, below 100fJ/bit switching energy and endurance above 10^15 cycles.
ScopeDevice switching energy excludes full macro, bus and system overhead. The 50nm dimension is not a 50nm CMOS process-node claim.
Source date / event period: 2023-12-13 · Checked 2026-09-10
MRAM · Device and Integration Research
STT-MRAM
Device and CMOS Integration DemonstrationsIBM contributes traceable device physics and CMOS integration results. MTJ size, process node and write-error rate from different studies must not be combined into an imaginary best-specification product.
ScopeIBM papers are not a list of IBM production foundry services.
Checked 2026-09-10
MRAM · Collaborative Research, Prototyping and Transfer
SOT-MRAM
Research Arrays, Prototypes and Trial ProductionITRI evidence spans distinct collaborations and versions: SOT with TSMC, cryogenic STT with NYCU, a joint β-W array, an 8-inch prototyping service and RRAM technology transfer.
ScopeCollaborative research must not be represented as an ITRI standalone production product.
Checked 2026-09-10
ReRAM / CBRAM · ReRAM IP licensor
ReRAM
Licensing and customer prototypesThree customer designs had taped out by July 2026, with a prototype running software.
ScopeFirst customer product mass production remained a future milestone.
Source date / event period: 2026-07-31 · Checked 2026-09-10
ReRAM / CBRAM · IDM adopting ReRAM
ReRAM
Technology transferWeebit reported onsemi ReRAM technology transfer progressing to schedule.
ScopeLicensing or transfer does not establish product mass production.
Source date / event period: 2026-07-31 · Checked 2026-09-10
ReRAM / CBRAM · IDM adopting ReRAM
ReRAM
Technology transferWeebit reported TI ReRAM technology transfer progressing to schedule.
ScopeTI commercial FRAM and this ReRAM transfer are separate technology routes.
Source date / event period: 2026-07-31 · Checked 2026-09-10
ReRAM / CBRAM · Foundry and IP partnership
ReRAM
Qualified IP available for SoC integrationThe official IP page lists qualified S130 130 nm CMOS ReRAM, available for integration, with two added masks in BEOL.
ScopeIP qualification does not qualify every customer chip or prove its mass production.
Undated source; checked 2026-09-10
ReRAM / CBRAM · Foundry and IP partnership
ReRAM
Qualified and available for integration130 nm BCD ReRAM IP is silicon-proven and qualified, adding two masks; listed specifications include 10K writes and over ten years retention at 125°C.
Scope100K cycles is an extension option; base BCD volume does not establish ReRAM product shipments.
Undated source; checked 2026-09-10
ReRAM / CBRAM · Embedded-memory foundry
ReRAM
40/22 in production; 12 risk productionThe IoT NVM page lists 40RRAM and 22RRAM in production; 12RRAM entered consumer-grade risk production in 2024, with cells between BEOL metal layers.
Scope12RRAM risk production is not full production or automotive qualification.
Source date / event period: 2024 · Checked 2026-09-10
ReRAM / CBRAM · Automotive MCU and process partnership
ReRAM
Announced integration partnershipThe 2022 announcement describes preparing TSMC RRAM for next-generation AURIX TC4x, supporting bit-wise writes without prior erase.
ScopeThis announcement does not prove every TC4x variant uses RRAM or has reached production.
Source date / event period: 2022-11-25 · Checked 2026-09-10
ReRAM / CBRAM · CBRAM acquisition and foundry integration
CBRAM
2023 acquisition; 22FDX then in qualificationGF acquired production-proven CBRAM technology from Renesas in 2023, following a 2020 Dialog license; 22FDX qualification was underway.
ScopePrior CBRAM production does not establish 22FDX production; the acquirer was GF, not Infineon.
Source date / event period: 2023-02-09 · Checked 2026-09-10
ReRAM / CBRAM · Embedded ReRAM MCU supplier
ReRAM
Commercial product familyThe M2L31 family lists an Arm Cortex-M23, 64–512 KB ReRAM and 72 MHz operation; writes do not require a page erase.
Scope72 MHz is the MCU clock, not cell write latency; density and reliability are part-specific.
Undated source; checked 2026-09-10
ReRAM / CBRAM · Historical ReRAM process collaboration
ReRAM
2017 joint-development announcementThe 2017 agreement combined Panasonic ReRAM with UMC manufacturing to develop a 40 nm mass-production process.
ScopeThe development target is not evidence of 2026 availability or the process used by every current Nuvoton part.
Source date / event period: 2017-02-01 · Checked 2026-09-10
ReRAM / CBRAM · Standalone ReRAM supplier
ReRAM
Specific part in mass productionThe product list marks the MB85AS8MT 8 Mbit SPI ReRAM as mass-produced with one million cycles; the 12 Mbit part requires sales contact.
ScopeDo not transfer the 8 Mbit production status or endurance to the 12 Mbit part.
Undated source; checked 2026-09-10
ReRAM / CBRAM · ReRAM IP and secure-processor developer
ReRAM
Vendor architecture and chip disclosureA 2026 company article describes the 22 nm Daric secure processor integrating ReRAM, computing and cryptography on one die.
ScopeThe article does not establish production qualification or independent security certification; avoid blanket immunity claims.
Source date / event period: 2026-05-06 · Checked 2026-09-10
ReRAM / CBRAM · ReRAM and neuromorphic research partnership
ReRAM
Research demonstrationThe announcement combines CEA-Leti spiking neural networks with Weebit SiOx ReRAM in a neuromorphic object-recognition demonstration.
ScopeThe demonstration is not an orderable complete AI accelerator; transfer and qualification require separate evidence.
Source date / event period: 2019-07-18 · Checked 2026-09-10
ReRAM / CBRAM · Foundry and Ecosystem Integration
22nm RRAM
Qualified RRAM IP; Named MRAM Products AvailableUMC evidence spans process availability, qualified RRAM IP, a SoC development platform and standalone MRAM products. These are distinct delivery levels.
ScopeA standalone MRAM product does not establish an available embedded macro.
Checked 2026-09-10
ReRAM / CBRAM · Technology transfer
1S1R RRAM
Undated research / transfer listingCross-point RRAM and selector development with published electrical and geometry targets.
ScopeNot evidence of current high-volume manufacturing.
Checked 2026-09-10
PCM · MCU supplier; classification boundary
PCM, not established MRAM offering
Announced; selected-customer early accessThe November 18, 2025 STM32V8 announcement explicitly identifies 18nm FD-SOI and embedded PCM, with Samsung Foundry manufacturing cooperation.
ScopeThe 2025 announcement offered early access to selected customers and targeted major OEM availability for Q1 2026, with broad availability later. Calendar passage does not prove completion.
Source date / event period: 2025-11-18 · Checked 2026-09-10
PCM · Historical 3D XPoint developer and supplier
3D XPoint
Historical route; development ceasedMicron announced an immediate end to 3D XPoint development in 2021 and redirected resources toward CXL memory products.
ScopeEnding this route does not end all PCM research; CXL is an interconnect, not a memory-cell mechanism.
Source date / event period: 2021-03-16 · Checked 2026-09-10
PCM · Historical Optane supplier
3D XPoint
Historical route; business wind-downIntel's 2022 annual filing states that the Optane memory business wind-down began in 2022.
ScopeHistorical product pages or inventory sales do not establish continuing development or other vendors' PCM exits.
Source date / event period: 2022 · Checked 2026-09-10
PCM · Automotive embedded-PCM MCU supplier
PCM
Stellar family; P3E sampling and production planThe 2026 Stellar P3E page identifies xMemory PCM, with full automotive qualification and production readiness planned for H2 2026.
ScopeReaching the planned quarter does not prove completion; P3E timing does not apply to every Stellar part.
Source date / event period: 2026 · Checked 2026-09-10
PCM · Analog in-memory-computing research
PCM
Research chipA 14 nm CMOS research chip with backend PCM integrates 64 256×256 analog cores and digital processing and communication for neural-network inference.
ScopeAnalog weight-compute results do not establish a purchasable general-purpose PCM memory or complete-system performance.
Source date / event period: 2023-09-17 · Checked 2026-09-10
FeRAM · Embedded FRAM MCU supplier
FeRAM
Commercial MCUs and reference designAn MSP430 reference design emulates EEPROM using embedded FRAM and lists supported MCUs and I2C/SPI host interfaces.
ScopeFRAM is the storage technology; EEPROM emulation is interface behavior, not floating-gate or FeFET construction.
Source date / event period: 2016-12-20 · Checked 2026-09-10
FeRAM · Standalone FeRAM and embedded-application supplier
FeRAM
Mass-produced productsThe FAQ states FeRAM has been mass-produced since 1999 for frequent-write applications; retention must be interpreted at the specified temperature.
ScopeCycle ratings, interfaces and temperatures are part-specific; FeRAM is not synonymous with all FeFET or FTJ devices.
Undated source; checked 2026-09-10
FeRAM · Standalone F-RAM supplier
FeRAM
Commercial product familyInfineon lists serial, parallel and EXCELON F-RAM using PZT ferroelectric films, with family-dependent endurance up to 100 trillion cycles.
ScopeMaximum ratings do not apply to every part; this does not establish HfO2 FeFET or FTJ construction.
Undated source; checked 2026-09-10
FeRAM · Ferroelectric process and research partnership
FeRAM
22FDX industrial-process integration demonstrationThe 2026 collaboration reports HfO2 ferroelectric FRAM integrated in 22FDX, operating below 1 V with nanosecond switching.
ScopeThe announcement lacks orderable parts, a complete qualification report or shipment volumes.
Source date / event period: 2026-06-11 · Checked 2026-09-10
FeRAM · Ferroelectric-device and compute-in-memory research
FeRAM / FeCAP
Research demonstrationJoint work with Georgia Tech demonstrated nondestructive FeCAP reading, reporting over 10^11 read cycles at IEDM 2023.
ScopeRead endurance is not write endurance and does not establish conventional FeRAM behavior or mass production.
Source date / event period: 2023 · Checked 2026-09-10
FeFET / FTJ · Ferroelectric-device research institute
FeFET / FTJ
Research and publicationsThe official 2025 publication list includes HZO bilayer FTJ thickness scaling and charge-trapping challenges in CMOS-embedded FeFETs.
ScopePublication listings establish research participation, not foundry service, PDK availability or mass production.
Source date / event period: 2025 · Checked 2026-09-10
FeRAM · Ferroelectric-memory commercialization developer
HfO2 Ferroelectric Memory
Commercialization and vendor solution claimsThe current site proposes DRAM+ persistent modules and CACHE+ persistent chiplets based on ferroelectric technology.
ScopeThe current page describes ferroelectric capacitors; it does not establish that DRAM+ or CACHE+ uses a FeFET or FTJ. Product qualification and shipment volumes are not verified.
Undated source; checked 2026-09-10
NOR / eFlash / SONOS · Technology and IP licensor
SONOS eFlash
Named production platforms and licensingThe official page identifies 2T SONOS, FN program/erase, production nodes and process/design licensing.
ScopeSeparate Cypress history from current macros; do not combine family-wide maximum specifications.
Undated source; checked 2026-09-10
NOR / eFlash / SONOS · Technology and IP licensor
SuperFlash NOR / eFlash
Commercial technology licensingSST lists SuperFlash process integration and licensing, complementing standalone NOR coverage.
ScopeLicensing does not establish identical current qualification across nodes; verify each technology generation.
Undated source; checked 2026-09-10
NOR / eFlash / SONOS · Integrated device manufacturer
eSTM eFlash / Page EEPROM
Named commercial implementationST links 40nm floating-gate eSTM with vertical select transistors to STM32H5 and Page EEPROM implementations.
ScopeKeep eSTM separate from Stellar PCM; it is not the cell used by every ST MCU.
Undated source; checked 2026-09-10
NOR / eFlash / SONOS · Integrated device manufacturer
SG-MONOS eFlash
Historical production and smaller-node researchThe 2016 announcement identifies production 40nm SG-MONOS MCUs and research on 16/14nm fin-shaped cells.
ScopeThis announcement does not prove 16/14nm production; verify current part numbers separately.
Source date / event period: 2016-12-07 · Checked 2026-09-10
NOR / eFlash / SONOS · Foundry and platform provider
XT011 eFlash / EEPROM
Named platform releaseThe 2024 announcement identifies embedded Flash and EEPROM on the XT011 110nm BCD-on-SOI platform.
ScopeKeep this separate from the historical 2003 XC06 example; do not assume identical cells.
Source date / event period: 2024-12-03 · Checked 2026-09-10
NOR / eFlash / SONOS · NOR/NAND product supplier
Serial NOR / OctaBus
Official product portfolioMacronix's official Serial NOR page provides its product portfolio and OctaBus interface offerings.
ScopeThis source directly supports NOR; verify NAND parts separately and do not infer cell geometry from interfaces.
Undated source; checked 2026-09-10
NOR / eFlash / SONOS · Code-storage memory supplier
W25Q16JW Serial NOR
Named product catalogThe official catalog identifies W25Q16JW Serial NOR and associated ordering entries.
ScopeCatalog presence does not guarantee stock; verify NAND families and other capacities separately.
Undated source; checked 2026-09-10
NAND · NAND and storage supplier
Ninth-Generation TLC V-NAND
Named generation in mass productionSamsung announced mass production of 1Tb TLC ninth-generation V-NAND in April 2024.
ScopeEvidence concerns this TLC generation; later demonstrations and projected QLC timing are separate.
Source date / event period: 2024-04-23 · Checked 2026-09-10
NAND · NAND and storage supplier
321-Layer TLC 4D NAND
Named generation starts mass productionSK hynix announced the start of 321-layer 1Tb TLC NAND mass production in November 2024.
ScopeSeparate production from customer delivery; do not assign this die to every Solidigm SSD.
Source date / event period: 2024-11-21 · Checked 2026-09-10
NAND · Enterprise SSD supplier
D5-P5336 QLC SSD
Named commercial productThe official D5-P5336 page identifies a commercial enterprise QLC SSD family and capacity options.
ScopeSSD capacity and system metrics are not die specifications; corporate relationships do not establish common NAND.
Undated source; checked 2026-09-10
NAND · NAND and storage supplier
G9 TLC NAND
Volume production and named SSD shipmentsThe July 2024 release reports G9 TLC NAND and volume shipment of the Micron 2650 SSD using it.
ScopeDo not extend TLC evidence to every QLC or NOR product; interface speed is not cell programming speed.
Source date / event period: 2024-07-30 · Checked 2026-09-10
NAND · Joint development and manufacturing partners
Tenth-Generation BiCS 3D NAND
Production begins at a named fabThe July 2026 joint announcement states that tenth-generation 3D Flash production began at Kitakami K2.
ScopeProduction start does not establish universal customer availability or identical finished products.
Source date / event period: 2026-07-03 · Checked 2026-09-10
NAND · NAND technology and product supplier
Xtacking 3D NAND
Published architecture and named product familyThe official page explains separate peripheral/array wafers joined by bonding and names Xtacking 4.0 X4 products in 2025.
ScopeAwards and architecture descriptions alone do not prove each product's volume, layer count or shipment status.
Undated source; checked 2026-09-10
NOR / eFlash / SONOS · NOR/NAND product supplier
GD25 / GD55 NOR
Official product portfolioThe official portfolio lists GD25/GD55 NOR products.
ScopeFamily coverage does not replace part-specific production status, temperature, endurance or retention specifications.
Undated source; checked 2026-09-10
OTP / MTP · Technology and IP licensor
LEE Flash ZT MTP
Named commercial IPThe ZT page describes FN program/erase, zero added masks and named platform production records.
ScopeGeneral and platform-specific endurance figures differ; do not infer universal endurance or polysilicon count.
Undated source; checked 2026-09-10
OTP / MTP · Memory IP / technology lineage
OTP
Named IP study; qualification is process-specificFollow the series select transistor and p-type floating-gate storage transistor as electron injection changes read current. Then distinguish normal OTP operation from the physical possibility of ultraviolet erasure.
ScopeHistorical cells and patents do not establish the structure of every current implementation.
Checked 2026-09-10
OTP / MTP · Memory IP / technology lineage
OTP
Named IP study; qualification is process-specificStart at the n-type cell's gate dielectric and follow high-field defect creation, changes in effective tunneling distance and the gate current used for sensing.
ScopeHistorical cells and patents do not establish the structure of every current implementation.
Checked 2026-09-10
OTP / MTP · Memory IP / technology lineage
OTP
Named IP study; qualification is process-specificThe original patent explicitly names XPM and distinguishes the storage MOS from the select MOS.
ScopeHistorical cells and patents do not establish the structure of every current implementation.
Checked 2026-09-10
OTP / MTP · Memory IP / technology lineage
OTP
Named IP study; qualification is process-specificOne gate spans thick and thin oxide; persistent conduction through the thin region creates the OTP state.
ScopeHistorical cells and patents do not establish the structure of every current implementation.
Checked 2026-09-10
OTP / MTP · Memory IP / technology lineage
MTP
Named IP study; qualification is process-specificFollow the control-coupling region, floating node and tunneling region as FN transport stores and removes electrons. A read transistor then senses the stored state.
ScopeHistorical cells and patents do not establish the structure of every current implementation.
Checked 2026-09-10
OTP / MTP · Memory IP / technology lineage
MTP
Named IP study; qualification is process-specificCompare hot-carrier programming of the p-type floating-gate cell with FN electron transfer toward a dedicated erase gate. Both operations act on the same storage node.
ScopeHistorical cells and patents do not establish the structure of every current implementation.
Checked 2026-09-10
OTP / MTP · Memory IP / technology lineage
MTP
Named IP study; qualification is process-specificYMC publicly identifies a logic-process MTP family. The CHI/BBHH sequence below is an independent mechanism illustration, not evidence that a current ymtp product uses BBHH. Separate product capability from an illustrative 1T1C model.
ScopeHistorical cells and patents do not establish the structure of every current implementation.
Checked 2026-09-10
OTP / MTP · Memory IP / technology lineage
MTP
Named IP study; qualification is process-specificFollow the named 2009 AEON company account: electrons enter and leave FG by FN, then a read MOS senses the state. Business and brand succession have a separate timeline.
ScopeHistorical cells and patents do not establish the structure of every current implementation.
Checked 2026-09-10
ReRAM / CBRAM · Joint device and memory R&D
Copper ReRAM
Historical research demonstrationMicron’s 2014 VLSI announcement identifies Sony collaboration on copper ReRAM for a 16Gb storage-class-memory demonstration.
ScopeThis does not establish current products or adoption in a Sony image sensor.
Source: 2014-06-05; checked 2026-09-10
NAND · NOR/NAND product supplier
GD5F NAND
Official product portfolioThe official portfolio separately lists GD5F NAND products.
ScopeFamily coverage does not replace part-specific production status, temperature, endurance or retention specifications.
Undated source; checked 2026-09-10
MRAM · Joint high-density MRAM research
64 Gbit 1Selector–1MTJ Cross-Point MRAM
IEDM 2024 research prototype; described publicly in 2025The jointly developed 64 Gbit cross-point MRAM replaces select transistors with two-terminal selectors; MTJ diameter is 20 nm, half-pitch 20.5 nm, and cell area 0.001681 µm².
ScopeThis is a research array, not evidence of product availability. The 20 nm figure is MTJ diameter, not a CMOS process node; typical-bit tests do not establish full-array yield.
Source date 2025-04-15
ReRAM / CBRAM · Multi-level RRAM analog in-memory computing developer
MLX200 Multi-Level RRAM Analog IMC
Tape-out and initial silicon validation completedIn May 2026, TetraMem reported MLX200 tape-out and initial silicon validation on TSMC 22nm, integrating multi-level RRAM with mixed-signal computing in a SoC.
ScopeEvaluation kits were scheduled for H2 2026 at announcement. Initial validation does not establish mass production or delivery, and computing results are not general-purpose storage specifications.
Source date 2026-05-19
ReRAM / CBRAM · Silicon-oxide RRAM and embedded-memory architecture collaboration
SiOx RRAM
Technology development and commercialization collaborationIntrinsic identifies silicon-oxide RRAM as its core technology and lists a sureCore collaboration combining CMOS-compatible cells, memory architectures and compiler-design expertise.
ScopeThe collaboration does not establish qualification or mass production of a named process macro; the site does not provide a complete orderable part and datasheet proving current supply.
Undated source; reviewed 2026-09-10
NRAM · Historical carbon-nanotube NRAM licensing and joint development
Carbon-Nanotube NRAM
2016 licensing and 55nm joint-development announcementFujitsu's official archive confirms that its two semiconductor businesses licensed Nantero carbon-nanotube NRAM in 2016 and began joint development toward a 55nm product.
ScopeThis historical development evidence establishes neither 2026 production and availability nor program termination; NRAM should be classified separately from oxide RRAM.
Source date 2016-08-31
OTP / MTP · Hardware Root of Trust & Security IP Provider
NeoPUF + AntiFuse HRoT (PUFcc / PUFiot / PUFker)
Commercial Production; Qualified on TSMC / UMC nodes (5nm–55nm)Leverages zero-mask NeoPUF and AntiFuse OTP to convert microscopic gate-oxide quantum tunneling variations into chip-unique fingerprints; integrates NIST SP 800-90B TRNG, secure key storage, and crypto accelerators (AES/ECC/RSA) for end-to-end hardware root-of-trust and secure boot.
ScopeRelies on standard logic CMOS tunneling physics; zero physical charge storage prevents static TEM/SEM key extraction, but peripheral digital controllers still require layered DPA counter-measures and active metal mesh.
2025–2026 Official Whitepaper & Production Qualification
NOR / eFlash / SONOS · Specialty Analog & High-Voltage BCD Foundry
Y-Flash 0-Mask eFlash / MTP
0.18µm and 65nm BCD platforms are public; vendor briefs mention Grade 0 retention narratives, but ambient Ta and junction Tj test conditions must be cited separately — do not equate 175°C with blanket certification.Single-poly floating gate; per Tower's public NVM brief and arXiv:2202.10228, program uses channel hot-electron injection (CHE) and erase uses band-to-band-tunneling (BBT) holes — not FN/FN. Zero mask adders integrate with Tower HV BCD/power platforms; public briefs cite 1K–10K endurance and high-temperature retention; name Ta, Tj, and AEC-Q100 qualification per product.
ScopeSingle-poly footprint yields larger cell size, targeting 1Kb–512Kb high-voltage PMIC trim, battery management (BMS), and gate drivers rather than high-density code storage.
2024–2026 Volume Production Specification
ReRAM / OxRAM · Independent embedded ReRAM IP and neuromorphic AI inference provider
Embedded ReRAM (OxRAM) & analog CIM research
DB HiTek 130nm qualified IP; SkyWater 130nm and GF 22FDX are platform/evaluation stages — customer production needs named evidence.Public OxRAM route centers on SiOx active layers (e.g., IMW 2019 samples with TiN bottom / Ti top electrodes). DB HiTek 130nm offers a qualified IP macro (2 masks, 10K cycles, 125°C retention under named conditions). SkyWater 130nm and GF 22FDX are distinct platform stages; analog CIM studies and customer production must not be merged into one spec sheet.
ScopeCommercial production focused on 130nm mature and specialty BCD nodes; advanced FD-SOI nodes (22nm) undergoing silicon tapeout; analog CIM subject to thermal drift requiring digital calibration.
2024–2026 foundry qualification & whitepaper
MRAM · Discrete & embedded MRAM silicon and IP supplier
Enterprise STT-MRAM & Data Center Power Loss Protection (PLP)
Volume production; foundry manufacturing at GlobalFoundries (22FDX / 12LP) and TSMCUtilizes perpendicular MTJ (pMTJ) with DDR4, DDR3, and xSPI interfaces for nanosecond persistent write and 10^10~10^12 endurance; replaces fragile supercapacitors in enterprise NVMe SSDs and AI accelerators for zero-latency journaling and capacitor-free PLP.
ScopeHigher per-bit cost than DRAM and NAND; optimized as persistent cache and write-buffer rather than primary mass storage.
2024–2026 production specification
PUF · Physical Unclonable Function (PUF) & Security IP Provider
SRAM PUF Root of Trust & Key Provisioning-Free Enclave
Over 500 million devices deployed globally; supports TSMC, UMC, GF, Intel, and Samsung advanced & mature nodesLeverages native 6T SRAM power-up mismatch as a hardware fingerprint; reconstructs 256-bit root keys dynamically via Fuzzy Extractor and public Activation Code; eliminates factory key provisioning costs and leak risks.
ScopeRequires volatile SRAM reconstruction at boot; working keys require runtime DPA masking and single-cycle zeroization.
2024–2026 Commercial Manual
AntiFuse · Silicon Security Subsystem & Physical IP Provider
Hardware Secure Module & AntiFuse Integrated Enclave
Widely deployed narrative; Synopsys public pages list the Automotive HSM as ASIL-B and OTP NVM separately as ASIL-D — do not merge into one “tRoot ASIL-D” or CC/PSA bundle certification.HSM subsystem integrating a security processor with AntiFuse OTP; public features include secure boot, key wrapping, and rollback interfaces — named crypto modes, timing, and PSA/CC bundles require product documentation; do not port portfolio peak tiers.
ScopeSubsystem-level IP requiring dedicated silicon area, memory protection units, and secure debug infrastructure.
2024–2026 Product Manual
OTP · High-Speed Interconnect Security & Silicon Lifecycle Key Provisioning Leader
PCIe/CXL SPDM Attestation & Silicon Lifecycle Root of Trust
Standard adoption in data center AI accelerators (GPU/NPU), CXL expanders, and server SoCsImplements DMTF SPDM 1.2/1.3 device attestation and line-rate PCIe/CXL IDE encryption; anchors silicon identity in foundry AntiFuse OTP across Foundry, OSAT, and CSP data centers.
ScopeEngineered for enterprise servers and data centers; rarely adopted in ultra-constrained consumer edge devices.
2024–2026 Product Manual
No matching entries. Clear the filters or try another technology name.
NVM · INDUSTRY & RESEARCH
Trace named implementations from purchasable memories and integrable IP/processes to materials and arrays still under validation. Align mechanisms, measurement levels and maturity before assessing application value.
RRAM filaments, STT tunneling-current writes and SOT separate paths impose different process, endurance and peripheral requirements. Products, foundry platforms and research papers answer different questions; preserve their conditions and attribution.
| Organization / Role | Technology and Integration | Public Maturity | Evidence Boundary |
|---|---|---|---|
| Everspin Standalone MRAM Product Supplier | Standalone Toggle MRAM, 1Gb persistent STT memory and EMxxLX xSPI. | Commercial Products; Qualification Is Version-Specific | DDR4-like does not mean a drop-in replacement for every DDR4 device; HR 64Mb qualification does not cover the whole family. |
| UMC Foundry and Ecosystem Integration | 22nm ULP/ULL RRAM; separately, Avalanche 22nm pMTJ STT-MRAM. | Qualified RRAM IP; Named MRAM Products Available | A standalone MRAM product does not establish an available embedded macro. |
| Panasonic ReRAM Technology and Product Lineage | Tantalum-oxide ReRAM; integration in foundry-standard 40nm CMOS. | Historical Production, Test Macros and Named Products | Sampling targets and test macros do not establish production of every 40nm product. |
| IBM Research Device and Integration Research | Perpendicular STT-MRAM; distinguish 14nm CMOS integration from smaller-MTJ research. | Device and CMOS Integration Demonstrations | IBM papers are not a list of IBM production foundry services. |
| ITRI Collaborative Research, Prototyping and Transfer | SOT-MRAM and BEOL integration; separately, 1S1R RRAM transfer. | Research Arrays, Prototypes and Trial Production | Collaborative research must not be represented as an ITRI standalone production product. |
Everspin · Standalone MRAM Product Supplier
Everspin is a core commercial MRAM reference. Its three product lines differ in write physics, interfaces, retention and qualification; compare exact ordering codes and document versions.
Product-route diagram: Toggle and STT use different write mechanisms; DDR/xSPI describe interfaces and application configurations, not new storage physics.
Read the adjacent operation descriptions together with this diagram.
The current PERSYST catalog marks MR3A16ACYS35 as mass production, with 8Mb, asynchronous x16, 35ns, 3.3V and −40°C to 85°C. These are named-part conditions, not universal MRAM density, speed or retention specifications.
The 2025 filing confirms continuing 1Gb STT-MRAM shipments; the technology page describes a DDR4-like persistent-DRAM interface. Retention and controller requirements need the relevant datasheet, rather than values copied from Toggle or xSPI.
The March 5, 2026 release confirms HR 64Mb AEC-Q100 Grade 1 production qualification and order availability. Its HR 128Mb/256Mb dates are announcement-time plans; later completion is not established by the primary text reviewed here. A September 2 Teledyne partnership starts with 256Mb and anticipates partner customer availability in Q4 2026.
Engineering interpretation: Toggle, persistent DDR buffers and low-pin-count xSPI serve different controllers and data lifecycles. Compare power-loss protection, update frequency, latency, qualification and board-integration cost.
Before adoption: Specify the complete part number, datasheet revision, retention temperature, endurance conditions, controller support and exact qualification scope.
UMC · Foundry and Ecosystem Integration
UMC evidence spans process availability, qualified RRAM IP, a SoC development platform and standalone MRAM products. These are distinct delivery levels.
Relationship diagram, not a cross-section. RRAM IP and standalone MRAM require separate qualification and specifications.
Read the adjacent operation descriptions together with this diagram.
The 2017 Panasonic collaboration establishes a 40nm history; eMemory announced qualification of a 22nm version in 2023. Public evidence does not establish an identical stack or IP lineage. Low-temperature backend integration and logic-platform compatibility matter to SoC adoption.
The 2018 agreement began with 28nm CMOS. The 2022 immediate-availability evidence concerns Avalanche third-generation 22nm P-SRAM. This establishes a named MRAM product route, not a universally available embedded macro.
Faraday FlashKit-22RRAM adds silicon-validated controller, BIST and processor integration in 2025. The 16Mb automotive and 0.8V/1.8V variants targeted in 2023 still require completion evidence. The Infineon/UMC 40nm automotive agreement identifies proprietary eNVM without identifying RRAM.
Engineering interpretation: evaluate RRAM for low-standby-power MCUs, AIoT and embedded code storage. Value depends on the selected IP retention conditions, update budget, test cost and controller integration.
Before adoption: Confirm the version, PDK, macro capacity, ECC, update granularity, automotive qualification and tapeout availability.
Panasonic · ReRAM Technology and Product Lineage
Panasonic is an important early commercial ReRAM developer. Trace 180nm production, 40nm test macros and Fujitsu products separately to connect the mechanism with reliability and products.
Functional layers, not exact thickness, electrode materials or bias polarity. SET/RESET change conduction through oxygen-ion and defect redistribution.
Read the adjacent operation descriptions together with this diagram.
The 2018 8Mbit test macro separately reports 100k-cycle endurance and over ten years at 85°C after 10k cycles. These do not establish ten-year retention after 100k cycles. Engineering evaluation also needs HRS/LRS distribution tails, not only average resistance.
The 2017 announcement confirms 180nm production since 2013 and describes future 40nm sampling. The 2019 Fujitsu/Panasonic MB85AS8MT release names an SPI product without identifying its node or foundry. It cannot by itself establish UMC 40nm product production.
Engineering interpretation: this lineage informs low-power persistent storage. Product selection must separate standalone serial-interface system costs from embedded-macro process costs.
Before adoption: Confirm current supplier, ordering code, interface and combined cycling/retention conditions rather than relying on historical partnerships.
IBM Research · Device and Integration Research
IBM contributes traceable device physics and CMOS integration results. MTJ size, process node and write-error rate from different studies must not be combined into an imaginary best-specification product.
Generic single-MTJ schematic; IBM studies need not share this stack or junction count. Arrows indicate opposite update directions only.
Read the adjacent operation descriptions together with this diagram.
The 2020 14nm study places MTJs between M1 and M2 with three added masks and one electrode module; 400°C describes process compatibility. The 2017 11nm result describes a junction size for low-current switching, not an 11nm CMOS platform.
The 2024 ordered-alloy study combines a low magnetic moment with strong perpendicular anisotropy, demonstrating a high energy barrier and 2ns operation. This advances material and switching design; retention guarantees still require temperature and statistical conditions.
IEDM 2025 reports a 4-kbit DS-MTJ array in which all devices switch with 2ns pulses, with an approximately 60kT energy barrier at about 40nm critical dimension. This extends the evidence to a research array while preserving the distinction between device geometry and CMOS node.
Engineering interpretation: use these studies to evaluate the physics and integration of embedded MRAM or cache candidates. Procurement still requires a named supplier and a deliverable macro, controller and production contract.
Before adoption: Align definitions of node versus MTJ size, pulse versus access time, device versus array WER, retention temperature and thermal budget.
ITRI · Collaborative Research, Prototyping and Transfer
ITRI evidence spans distinct collaborations and versions: SOT with TSMC, cryogenic STT with NYCU, a joint β-W array, an 8-inch prototyping service and RRAM technology transfer.
Generic bottom-channel SOT mechanism, not the exact 2023 top-SOT or high-RA dual-MTJ cross-section.
Read the adjacent operation descriptions together with this diagram.
The 2022 0.4ns and seven-trillion-cycle claim concerns SOT with TSMC; the same release attributes −269°C to 127°C operation to cryogenic STT with NYCU. The IEDM 2023 result released in 2024 addresses 10ns devices and computing. Its 1% power claim lacks a fully disclosed comparison workload.
A joint NYCU, TSMC, ITRI, NSRRC, Stanford and NCHU paper uses Co insertion to stabilize β-W. The film maintains phase stability at 400°C for ten hours, and the memory demonstrates 1ns switching. This is integration evidence, not a production-platform announcement.
The 8-inch BEOL platform supports Kb-to-Mb research, prototypes and initial trial production. A separate 1S1R RRAM transfer page lists selector, current and geometry specifications. Its undated status cannot establish current production or the maturity of the overall ReRAM market.
Engineering interpretation: consider joint research, process prototyping and compute-architecture validation. SOT speed and barrier-reliability potential must be weighed against channel area, selectors, routing and drive current.
Before adoption: Request version-specific capacity, WER, field or assist requirements, retention, process compatibility and deliverable service scope.
A switching pulse excludes some or all decoding, verification, ECC and serial-transfer overhead. Engineering interpretation: establish the measurement level before comparing access time or update energy.
Thermal budget describes processing, retention temperature describes storage and operating range is a separate specification. Best values from different experiments do not form a product guarantee.
Engineering interpretation: separate components, IP, PDKs and joint research. Power, endurance and density become adoption advantages only when testing, yield, controllers, licensing and supply meet the product requirements.
The 2016 Fujitsu archive identifies carbon nanotubes as the NRAM storage technology and a 55nm joint-development plan with Nantero. That evidence does not identify an oxide-filament ReRAM device, nor does it establish a currently available product. The detailed write, reverse-update and read waveforms require a specific device publication or datasheet; no voltage or endurance specification is inferred from the licensing announcement.
Reviewed through 2026-09-10 using official releases, product documents, author papers and public abstracts. Claims based on abstracts remain limited to those abstracts. Engineering interpretations are comparative analysis, not supplier guarantees.
Physics Background 01 · Irreversible Structures
eFuse is suited to permanently storing small amounts of on-chip configuration, such as trim codes, repair addresses, and identification data. One-time programming means that each physical location supports an effective transition in only one irreversible direction; separate locations can be programmed in batches. Multiple programming commands to a macro do not make an individual fuse reversibly erasable and rewritable.
A bit is stored as a difference in the resistance of a conductive path. An unprogrammed fuse typically has low resistance; a controlled current causes material migration or a break in a designated region, producing higher resistance. Logic 0/1 is defined by sensing and encoding. High resistance does not intrinsically correspond to a particular bit value, and the programmed state must not be assumed to be an ideal open circuit.
Current and Local Heating Permanently Change Resistance in a Designed Conductor Region.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
EM
Electromigration opens a silicide gap at the neck; underlying polysilicon can remain, so high resistance does not imply complete physical separation.
The intact conductor provides a low-resistance path.
Current crowds at the constriction, producing local heat and electromigration.
Material migration leaves a local void or gap.
After program stress is removed, read current identifies the permanent state.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Joule
Shows current crowding, local heating, and melt separation near the via; a silicide electromigration path is not substituted.
The intact conductor provides a low-resistance path.
Current crowding generates local heat near the via.
Material near the via heats, melts, and separates, leaving a high-resistance gap.
After program stress is removed, read current identifies the permanent state.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Carriers travel along the conductor, while local current density and Joule heating increase the rate of material migration. US7417300B2 uses a narrowed geometry to control current crowding and address material backflow; US8847350B2 shapes current concentration through via contact placement. The teaching focus is permanent material redistribution and the final resistance distribution, rather than describing every mechanism as metal rupture.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
EM
Electromigration opens a silicide gap at the neck; underlying polysilicon can remain, so high resistance does not imply complete physical separation.
Keep the unprogrammed structure as a reference.
Removing normal power does not restore the original material.
No qualified electrical erase path exists; read still detects the changed state.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Joule
Shows current crowding, local heating, and melt separation near the via; a silicide electromigration path is not substituted.
Keep the unprogrammed structure as a reference.
Removing normal power does not restore the original material.
No qualified electrical erase path exists; read still detects the changed state.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Changing the data encoding or activating spare locations can change the information read by the system, but it does not restore the original fuse material. The number of available updates depends on reserved capacity and the protocol, not on the program/erase endurance of one bit. Test flows must also reserve sacrificial test cells rather than treating irreversible programming as a repeatedly recoverable functional test.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
EM
Electromigration opens a silicide gap at the neck; underlying polysilicon can remain, so high resistance does not imply complete physical separation.
Start with permanently changed material; reading does not repair it.
Observe the existing path under read conditions below program stress.
The gap reduces read current; high resistance does not mean an ideal open circuit.
Compare two pre-existing states at equal V_R; reading did not change the material.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Joule
Shows current crowding, local heating, and melt separation near the via; a silicide electromigration path is not substituted.
Start with permanently changed material; reading does not repair it.
Observe the existing path under read conditions below program stress.
The gap reduces read current; high resistance does not mean an ideal open circuit.
Compare two pre-existing states at equal V_R; reading did not change the material.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
The required guarantee is that the high- and low-resistance distributions remain distinguishable across temperature, process variation, and time in service. Programmed resistance is not infinite, and unprogrammed resistance is not zero; series resistance from the select transistor and wiring affects the result. The readout must therefore be analyzed with its reference path, rather than judging read reliability solely from a gap in a microscope image.
A transistor typically selects the cell's programming and read paths. The programming driver must withstand the pulse current, and supply or wiring drops must not leave distant cells underprogrammed; unselected cells must avoid additional stress on shared lines. A small fuse link therefore does not guarantee a small complete macro once drivers, sensing, and wiring are included.
Fuse width, silicide or metal thickness, via overlap, and the thermal environment change local current density. Programming pulses then translate those differences into a post-program resistance distribution. Measure the distributions before programming, after programming, after thermal treatment, and after read-life stress; examine material backflow, residual paths, and reference-sensing drift rather than reporting only an average programming success rate.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
The physical limit is the window for reproducible material change: too little stress produces overlapping resistance distributions, while too much may damage neighboring structures. Any local geometry improvement must be demonstrated through sensing margin across process corners and after aging. No single minimum linewidth or programming current applies to every material.
Larger arrays require more decoding, programming-drive, and sensing resources. An individual resistor may shrink, but line drop, selected-cell isolation, reference matching, and test redundancy do not disappear with it. Effective density should be calculated as the number of reliable deliverable bits divided by the area of the complete macro.
Fuse materials and contacts must comply with the baseline process's thermal budget, interconnect reliability requirements, and layout rules. Deliberately induced local migration must not propagate into ordinary wiring failure. Via placement and neighboring structures are part of process integration and cannot be signed off through logic simulation alone.
System limits include remaining programmable locations, the cost of irreversible misprogramming, and the supply's pulse capability. If many data updates are required, additional encoding and redundancy progressively consume capacity. Permanent records still require access permissions, fault handling, and update-transaction integrity.
Suitable for chip identification, analog trimming, manufacturing repair, and permanent options when programming is infrequent and the data lifecycle can be defined in advance. Before selecting it, establish the latest programming point, whether field appends are needed, the reserved bit count for each field, and whether the platform can provide the specified programming pulses and verification.
Not suitable as a direct store for frequently rewritten counters, logs, or user settings that must be freely reversible. Low-cost, large-capacity updatable storage requires comparison of complete macro and management costs. The word Fuse also does not make a resettable electronic fuse IC for power protection equivalent to this memory cell.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US7417300B2
Compare the terminal and neck widths, then trace current crowding and the local thermal gradient. Geometry and material distribution are the design variables.
Compare the eFuse write sequence: current path → material redistribution → high resistance. Different fuse stacks need different physical failure models.
Compare All Drawings, Numerals and Claims →US8847350B2
Follow the top-view metal link into the two via cross-sections. A smaller contact area concentrates current and local heating.
Compare localized heating and separation in a metal-via fuse. Its material explanation differs from the silicide electromigration example.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
If an OTP macro allows ten data appends, does each eFuse have ten program/erase cycles of endurance?
No. The appends may use ten groups of previously unprogrammed locations and version encoding; each fuse still supports only one irreversible transition. Record the system's update count, the capacity consumed per update, and the rewritability of an individual physical cell separately.
Physics Background 02 · Irreversible Structures
Antifuse design aims to establish acceptable conduction first in a designated storage dielectric while keeping selectors, peripheral circuits, and half-selected cells functional; it does not rely on uncontrolled breakdown across the chip. Permanent data can support identification, trimming, code, or key storage. Security is separately determined by the readout interface, access controls, and protective design.
Before programming, the storage dielectric separates two electrodes and permits only very small leakage. Programming creates a permanent, detectable conduction path through a high electric field. Information resides in the conduction difference before and after breakdown, opposite to the typical eFuse transition from low to high resistance. Both can provide OTP, but their storage materials and programming conditions differ.
MOS Antifuse Example: Local Dielectric Breakdown Changes a High-Resistance State to a Low-Resistance State.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
BD
Storage element only: gate connects to column C, silicon to an internal node. The series select MOS and array periphery are omitted, not a complete commercial OTP cell.
Intact dielectric blocks low-field DC.
A high field stresses the thin dielectric; periphery limits stress.
A local conducting path lets electrons cross the former dielectric region.
After program stress is removed, read current identifies the permanent state.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
The selected MOS provides a path that establishes a storage-layer field between the high-potential column line and a low-potential internal node. Programming is dielectric breakdown of the core-device gate oxide: Vpgm is set by that process oxide thickness and the allowed time-to-breakdown, typically several times core Vdd, not a node-independent 2.8–3.5 V. Public Kilopass teaching places program voltage near 8–9 V for a ~32 Å oxide and 5–6 V for ~20 Å. A named foundry example separately shows that a 1.8 V core must survive program stress well above Vdd; that is evidence of the stress class, not a universal Vpgm. This path is not I/O floating-gate hot-carrier injection, whose program window follows the I/O device (about 6.5 V PGM on 3.3 V PMOS and 7.5 V on 5 V PMOS; an NMOS floating-gate cell at the same node needs a higher Vpgm). That I/O FG HCI pairing is public literature / architecture-class and must not be back-filled onto oxide-breakdown AntiFuse. Current must be controlled to protect the selector. Evolution from soft breakdown to stronger conduction is a distributed process, not the formation of an ideal metal wire with identical dimensions and resistance in every cell.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
BD
Storage element only: gate connects to column C, silicon to an internal node. The series select MOS and array periphery are omitted, not a complete commercial OTP cell.
Keep the unprogrammed structure as a reference.
Removing normal power does not restore the original material.
No qualified electrical erase path exists; read still detects the changed state.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Resistive memories that also contain conductive paths may perform SET/RESET through controlled ion migration; antifuse OTP does not provide a product operating contract for reversible recovery. The ability to draw a conductive filament in both devices does not make their endurance, operating directions, or array requirements equivalent.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
BD
Storage element only: gate connects to column C, silicon to an internal node. The series select MOS and array periphery are omitted, not a complete commercial OTP cell.
Start with permanently changed material; reading does not repair it.
Observe the existing path under read conditions below program stress.
The broken-down region conducts more strongly; its geometry is unchanged.
Compare two pre-existing states at equal V_R; reading did not change the material.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
The read criterion is the current distribution of each state under the actual read bias. Selector threshold, series resistance, dielectric leakage, and variation in programmed conduction all affect margin. Validate these over the specified temperature and time conditions rather than treating one high-current measurement immediately after programming as sufficient evidence of lifetime read reliability.
A selected cell needs both a high column potential and a select path that establishes the low-side potential. Cells in the same row but another column, the same column but another row, and fully unselected cells see different fields, so each category needs its own bias table. Isolation requires more than turning off a row line. High-voltage level shifting, well isolation, and current limiting together allow the storage layer to break down before other devices are damaged.
Dielectric thickness, local defects, area, and field distribution produce statistical distributions of breakdown time and post-breakdown current. Research should report pulse amplitude/duration, current compliance, initial leakage, and the post-program verify threshold together, and examine accumulated half-select stress. Comparing typical programming voltage alone cannot establish yield or sensing margin.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
The usable window lies between reliably establishing sufficient conduction and avoiding unintended damage. Reducing storage area or changing the dielectric does not necessarily improve breakdown statistics and programmed current proportionally. Limits should be described by complete distributions, specified temperatures, and long-term sensing margin.
Density is jointly limited by selectors, high-voltage decoding, half-select isolation, wiring drops, and sensors. Theoretical cell area is not a substitute for complete macro efficiency. As an array grows, additional unselected leakage and accumulated programming stress must enter its failure-rate budget.
Availability in standard CMOS does not remove the need for qualification. A core-device gate-oxide AntiFuse works only when the storage oxide can form a controlled conduction path before the source/drain junction avalanches. That window typically opens on 1.8 V-class and thinner core oxides and continues into advanced logic; a 2.5 V-class thicker oxide often reaches junction breakdown first, so the bitcell fails before a usable antifuse path forms. Controlled breakdown of the storage dielectric, selector voltage tolerance, thick/thin oxide choices, and reliability tests must be reconfirmed for the foundry process version. Success at one node does not establish usable pulses at another.
System bottlenecks often include the irreversible programming flow, remaining blank bits, and supply conditions during programming. For keys or boot settings, programming permissions, read isolation, lock state, and fault recovery must be validated separately from the storage physics. The OTP label does not replace system design.
Suitable for parameters determined after manufacturing and intended to remain permanent, chip IDs, code with a defined lifecycle, and security settings. Selection starts with the required update count, latest programming location, supply, read timing, and capacity, followed by verification of the identified macro's process, test modes, and qualifications.
Not suitable for unrestricted repeated overwrites without a redundancy plan. A system that must directly overwrite the same address or restore an old value should use electrically erasable storage or an upper-layer protocol for a bounded number of appends. High-field test results must not be equated with production yield or resistance to invasive attacks.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US6667902B2
Locate the storage element and select transistor in Figure 3, then compare selected and unselected biases in Figure 8. The listed voltages belong to this embodiment.
Compare antifuse operation: intact dielectric → selected high field → permanent conduction path, followed by low-stress sensing.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
Why is reducing the gate voltage of an unselected row insufficient to prove that an entire antifuse array is immune to half-select misprogramming?
Cells sharing a column or row, and internal floating nodes, can develop different storage-layer fields. Turning off a select MOS does not hold every node at zero potential. Each category's bias, leakage, and accumulated stress must be evaluated and then validated with the specified pulse sequence.
Physics Background 03 · Charge Storage
This topic covers conventional standalone EEPROM: the array, voltage boosting, controller, and interface are packaged as a separate device accessed through a serial or parallel interface. Microchip 24LC256 is an identified I2C serial example. Foundry EEPROM macros and third-party MTP IP integrated within a chip are compared in the separate Embedded MTP IP topic.
The same storage cell supports programming, electrical erase and subsequent programming. ERS restores a window suitable for another program operation; it does not require every carrier to disappear. Whether the host issues a separate erase command depends on the macro or component interface.
PGM → ERS → PGM
A floating gate is a conductive island surrounded by insulating layers, with no direct DC metal connection to it. Retained charge changes how the control gate acts on the channel, shifting the MOS threshold voltage. In a typical n-channel example, adding electrons makes conduction more difficult. Reading measures the channel; normal read operation does not require draining the stored electrons.
Electrons Enter or Leave an Insulated Floating Gate through Tunneling or an Implementation-Specific Injection Path.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
The n-channel branch expressly permitted by US4115914A is redrawn with n+ contacts and p-type silicon. Most original process figures use a p-channel example; this is not a literal reproduction of that process section. This sequence uses tunneling only.
Isolated storage starts within its programmable window.
Terminal conditions belong only to the named variant.
Electrons tunnel through the local barrier into isolated storage.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
FN tunneling occurs at the thin dielectric's high-field region. The control gate acts through capacitive coupling and has no metal connection to the floating gate. Byte/page write commands are interface behavior and do not directly describe bare-cell bias voltages.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
The n-channel branch expressly permitted by US4115914A is redrawn with n+ contacts and p-type silicon. Most original process figures use a p-channel example; this is not a literal reproduction of that process section. This sequence uses tunneling only.
Locate programmed charge and this variant’s exit.
Terminal conditions belong only to the named variant.
Electrons leave storage through this variant’s specified exit.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Separate physical erase from host commands. A 24LC256 write includes an internally timed erase/write cycle; the host does not directly apply bare-cell erase biases. Selection lines, shared terminals, and page organization determine the affected region, which the EEPROM label alone does not specify.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
The n-channel branch expressly permitted by US4115914A is redrawn with n+ contacts and p-type silicon. Most original process figures use a p-channel example; this is not a literal reproduction of that process section. This sequence uses tunneling only.
Drawn stored carriers represent data, not the source of read current.
Select the measured path and apply low-field read conditions.
The stored state determines sense current under the same read bias.
Compare stored states under equal read conditions; logic encoding is not assigned.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
The source-to-drain channel carries the read current; floating-gate electrons need not leave. Interface clock rate, serial transfer, and random/sequential access are device timing conditions, separate from bare-cell sensing time.
The local window confines charge transfer while selection controls which storage path receives program, erase, or read biases. A standalone device also contains address decoding, page latches, voltage boosting, and command control. Byte-update costs require checking internal update granularity and page-boundary rules.
Floating-gate charge, coupling ratio, tunnel-oxide thickness, and interface defects determine threshold voltage and its drift. Traps accumulated through cycling change program/erase speed and retention. Studies should measure cycle count, temperature, retention time, and error criteria together. One product's typical cycle count cannot be combined with another product's best retention time to form a common limit.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
The localized thin window must support tunneling while preserving long-term insulation. Defect accumulation, coupling ratio, and charge distributions determine the threshold window remaining after repeated operation.
Fine updates require selection, decoding, page latches, and high-voltage distribution. Page boundaries, internal erase/write units, and verification jointly limit update speed; one floating gate's area is an incomplete comparison.
A dedicated memory process controls the tunneling window, inter-gate dielectric, and high-voltage devices. Public teaching cross-sections explain principles; the stack and process conditions of a current part require separate manufacturer evidence.
Serial transfer, internal erase/write busy time, write protection, and power-loss handling define system behavior. Budget lifetime cycles at the actual updated locations; important settings can use versions, checksums, and a controlled switchover.
Suitable for settings, calibration, product identification, and moderately updated state stored outside the host chip. Establish capacity, interface, page rules, write latency, and lifetime cycle demand before selecting a part and temperature grade.
Frequent high-throughput logging, large sequential data, and very low latency workloads may be constrained by interface and erase/write timing. For integration within the same chip, use the Embedded MTP IP topic's process routes.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US4115914A
Identify the floating gate, localized thin region and upper control gate in the late process cross-sections. Thin-region placement and the second dielectric determine coupling and tunneling paths.
Compare EEPROM FN injection and removal: both directions must pass through the actual thin dielectric region.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
Does 24LC256 page-write support prove its polysilicon count and local-window cross-section?
No. The datasheet establishes device interface behavior, update rules, and ratings. A bitcell cross-section needs separate implementation evidence. This topic uses US4115914A to explain a local window without assigning that patent to the 24LC256.
Physics Background 04 · Charge Storage
MTP IP provides rewritable nonvolatile storage inside the host chip, so selection centers on process and macro integration. Foundry double-poly EEPROM can be supplied through a dedicated NVM option; identified public evidence for third-party single-poly alternatives includes Synopsys MTP EEPROM and eMemory NeoEE/NeoMTP. This category is separate from packaged standalone EEPROM and does not merge SONOS Flash or antifuse OTP into floating-gate MTP.
The same storage cell supports programming, electrical erase and subsequent programming. ERS restores a window suitable for another program operation; it does not require every carrier to disappear. Whether the host issues a separate erase command depends on the macro or component interface.
PGM → ERS → PGM
Separate the embedded macro from a packaged EEPROM component, then identify the NVM stack and its integration contract.
| IP and Provider | Poly Layers and Control | Process and Operation Contract |
|---|---|---|
| Foundry Double-Poly EEPROM Option X-FAB XC06 (historical 2003 example) | 2: double-poly NVM stack Separate polysilicon layers can form floating and control gates; the XC06 brief does not disclose the complete EEPROM cross-section. | Use the target foundry's EEPROM/NVM option and check the actual mask combination and memory specification. The EEPROM macro defines program/erase mechanism, biases, and granularity; neither layer count nor a process brief is sufficient. |
| Third-Party Single-Poly MTP IP Synopsys MTP EEPROM; eMemory NeoEE/NeoMTP | 1: explicitly stated for the identified products Capacitive coupling controls a floating node; control capacitors, storage transistors, and erase regions are vendor-specific. | Integrate on a specified logic, analog, or BCD platform. Zero added masks applies only to explicitly supported versions. NeoEE uses FN program/erase; NeoMTP uses p-type storage and an erase gate. The reviewed Synopsys source does not disclose carrier paths. |
| Named MTP IP | Poly and Storage | Program / Erase and Integration |
|---|---|---|
| Synopsys · MTP EEPROM IP | Single-poly Floating gate | Program: PGM is supported; the reviewed product page does not disclose carrier paths or terminal biases. Erase: Electrical ERS is supported; the reviewed product page does not disclose the microscopic erase path or terminal biases. Specified standard-CMOS platforms; zero added masks; integrated high-voltage circuitry Do not assign US5844271A or another vendor's carrier paths to this product. |
| eMemory · NeoEE | Single-poly Floating gate; capacitive-coupling MOS devices and selectors | Program: FN transfer stores charge on the floating gate Erase: FN transfer removes charge from the floating gate Specified logic processes; zero added masks; integrated high-voltage and control circuits Complete terminal biases and macro update units require the specified version. |
| eMemory · NeoMTP | Single-poly p-type floating-gate MOSFET; additional erase gate | Program: Channel-hot-hole-induced hot-electron injection (manufacturer label: CHEI) Erase: FN electron transfer from floating gate to erase gate Specified logic/BCD platforms; zero-added-mask versions Do not substitute an n-type storage transistor or source-erase diagram; use the macro's terminal specifications. |
This topic focuses on floating-gate embedded MTP/EEPROM IP. Charge remains on an insulated conductive floating node and alters channel conduction through capacitive coupling. n-type and p-type storage transistors have different read-state behavior: adding electrons makes the n-channel US5844271A example harder to turn on, while eMemory describes its p-type NeoMTP device as turning on after electron injection.
Two Sections from US5844271A Show Coupling and Read Regions; the Brown Link Indicates One Floating Node, Not a Literal Layout Route. Control Is in Silicon. Current MTP IP Cells, Carrier Paths and Periphery Follow Their Own Documentation.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
Uses the control/floating gates and local window in US4115914A to explain the double-poly EEPROM route. This is not a named foundry macro cross-section; process details, tunneling terminals and operating conditions remain vendor-specific.
Isolated storage starts within its programmable window.
Terminal conditions belong only to the named variant.
Electrons tunnel through the local barrier into isolated storage.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
CHE / FN
Figures 4 and 5 of US5844271A share one FG conductor, with control buried in silicon. This teaching example does not define current vendors’ MTP cells or carrier paths. Erase coupling remains symbolic; inconsistent read entries in Table 2 are not reproduced.
Isolated storage starts within its programmable window.
Terminal conditions belong only to the named variant.
Channel electrons accelerate before local injection into storage.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Single-poly does not specify an injection mechanism. NeoMTP uses channel-hot-hole-induced hot-electron injection in a p-type device, distinct from the n-channel teaching example. Floadia ZT separately discloses FN programming. A low core supply does not eliminate internal voltage boosting or high fields.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
Uses the control/floating gates and local window in US4115914A to explain the double-poly EEPROM route. This is not a named foundry macro cross-section; process details, tunneling terminals and operating conditions remain vendor-specific.
Locate programmed charge and this variant’s exit.
Terminal conditions belong only to the named variant.
Electrons leave storage through this variant’s specified exit.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
CHE / FN
Figures 4 and 5 of US5844271A share one FG conductor, with control buried in silicon. This teaching example does not define current vendors’ MTP cells or carrier paths. Erase coupling remains symbolic; inconsistent read entries in Table 2 are not reproduced.
Locate programmed charge and this variant’s exit.
Terminal conditions belong only to the named variant.
Electrons leave storage through this variant’s specified exit.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Polysilicon count cannot establish erase destination, polarity, or granularity. Capacitive-control regions and tunneling exits may differ. Changing a control voltage is not a direct metal connection that drains the floating gate. Dielectric defects still accumulate through cycling. Electron removal by FN and net-charge reduction by hot-hole injection are different mechanisms. The separate BBHH teaching model is not evidence that a current YMC product uses BBHH.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
Uses the control/floating gates and local window in US4115914A to explain the double-poly EEPROM route. This is not a named foundry macro cross-section; process details, tunneling terminals and operating conditions remain vendor-specific.
Drawn stored carriers represent data, not the source of read current.
Select the measured path and apply low-field read conditions.
The stored state determines sense current under the same read bias.
Compare stored states under equal read conditions; logic encoding is not assigned.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
CHE / FN
Figures 4 and 5 of US5844271A share one FG conductor, with control buried in silicon. This teaching example does not define current vendors’ MTP cells or carrier paths. Erase coupling remains symbolic; inconsistent read entries in Table 2 are not reproduced.
Drawn stored carriers represent data, not the source of read current.
Select the measured path and apply low-field read conditions.
The stored state determines sense current under the same read bias.
Compare stored states under equal read conditions; logic encoding is not assigned.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Macro read behavior depends on the cell, coupling ratio, selectors, bitlines, references, and ECC. A public cross-section explains storage principles but does not replace macro timing, output protocol, or usable read margin.
Single-poly shifts control requirements into capacitive coupling and layout while retaining selectors, well isolation, and controlled high-voltage distribution. Both foundry NVM options and third-party IP require selected, half-selected, and unselected operating conditions, update granularity, and disturb limits. Total area includes boosting, control, sensing, ECC, and redundancy.
Coupling ratio, oxide quality, channel type, charge-transfer mechanism, and cycling defects jointly govern distribution drift. FN and hot-carrier paths have different energy and stress conditions. Compare vendors at fixed capacity, temperature, cycle count, retention time, and error criteria instead of combining the best figures from unrelated macros.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
Coupling, tunneling or hot-carrier efficiency, and dielectric reliability limit the usable read window after repeated updates. n/p channel types and different erase destinations cannot share one universal operating limit.
Boosting, selectors, references, ECC, redundancy, and verification determine total macro area and usable throughput. Fine updates can reduce array efficiency; shared lines and half-select disturb constrain parallel program/erase.
Double-poly NVM options and single-poly logic integration have distinct masks, oxides, wells, thermal budgets, and model requirements. Zero added masks still requires memory qualification on the target process. Base-process poly count cannot substitute for the NVM-option stack.
Update frequency, program/erase stalls, power budgets, and power-loss consistency jointly define usability. The host must honor macro busy states, locks, error handling, and test modes while budgeting cycles over product lifetime.
Suitable for analog trimming, PMIC settings, sensor parameters, and device configuration that require updates within the same chip. Fix the foundry process and use lifetime first, then compare foundry EEPROM options, third-party MTP macros, total area, and qualification conditions.
Do not substitute the MTP label for physical and macro documentation or treat it as RAM with unlimited rewrites. For a separate memory device, return to conventional EEPROM. Classify SONOS, embedded Flash, and antifuse by their respective storage mechanisms.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US5844271A
Find the overlap between the buried control region and floating gate, then use the equivalent circuit to distinguish coupling, storage and channel conduction.
Compare the MTP IP study’s single-poly teaching variant: this patent’s CHE injection and FN removal use a buried control node, with no second control-poly layer above the floating gate. Current product mechanisms require their own documentation.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
If two third-party MTP products are both single-poly, can both use the same n-channel hot-electron program and source-FN erase diagram?
No. NeoEE publicly describes FN charge transfer in both directions. NeoMTP describes a p-type floating-gate device, channel-hot-hole-induced hot-electron injection, and an erase-gate exit. Single-poly establishes layer count; a complete diagram must still match the identified cell, terminals, and operating conditions.
Physics Background 05 · Charge Storage
NOR is commonly used for code storage requiring direct, predictable reads. Stacked-gate cells place storage and selection responsibilities under the cell's gate control; split-gate cells add a selection channel that helps block unselected leakage from overerased cells. Execute-in-place support also depends on the interface, controller, and cache timing and cannot be guaranteed by the NOR name alone.
The same storage cell supports programming, electrical erase and subsequent programming. ERS restores a window suitable for another program operation; it does not require every carrier to disappear. Whether the host issues a separate erase command depends on the macro or component interface.
PGM → ERS → PGM
NOR describes array connectivity and access organization, not a unique storage material. This topic uses floating-gate NOR: charge changes cell threshold voltage, and the selected cell is sensed through the bitline and source path. Both stacked-gate and split-gate cells can serve NOR arrays, but their selection channels, programming efficiency, and erase control differ.
Split-Gate Variant: Local Injection Transfers Charge into the Floating Gate.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
CHE / FN
Corresponds to conventional stacked-gate mechanisms in the background of US6232180B1, not to its proposed split-gate invention.
Isolated storage starts within its programmable window.
Terminal conditions belong only to the named variant.
Channel electrons accelerate before local injection into storage.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
SSI / FN
WL-poly erase exit of the first/second generation; third-generation dedicated erase gates and another patent’s well-erase conditions are not substituted.
Isolated storage starts within its programmable window.
Terminal conditions belong only to the named variant.
Channel electrons accelerate before local injection into storage.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
SSI / FN
SG is below an overlapping FG; source is raised for program, while SG/S/D float and the well is raised for erase.
Isolated storage starts within its programmable window.
Terminal conditions belong only to the named variant.
Channel electrons accelerate before local injection into storage.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Source-side injection uses the potential drop at the boundary between the selection and storage regions to energize channel electrons, which are collected by the field toward the floating gate. Source/drain labels and biases must follow the original drawing; injection locations in different structures cannot be interchanged arbitrarily. Higher injection efficiency also does not eliminate the macro's need for voltage boosting, current limiting, or verification.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
CHE / FN
Corresponds to conventional stacked-gate mechanisms in the background of US6232180B1, not to its proposed split-gate invention.
Locate programmed charge and this variant’s exit.
Terminal conditions belong only to the named variant.
Electrons leave storage through this variant’s specified exit.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
SSI / FN
WL-poly erase exit of the first/second generation; third-generation dedicated erase gates and another patent’s well-erase conditions are not substituted.
Locate programmed charge and this variant’s exit.
Terminal conditions belong only to the named variant.
Electrons leave storage through this variant’s specified exit.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
SSI / FN
SG is below an overlapping FG; source is raised for program, while SG/S/D float and the well is raised for erase.
Locate programmed charge and this variant’s exit.
Terminal conditions belong only to the named variant.
Electrons leave storage through this variant’s specified exit.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Two split-gate implementations may have different erase barriers and charge exits, so their terminal voltages cannot be combined into one bias set. Erase scope depends on shared gates, wells, and array control. Updating a byte still requires the product's block-erase and data-preservation flow; fine-grained reading is not equivalent to fine-grained erase.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
CHE / FN
Corresponds to conventional stacked-gate mechanisms in the background of US6232180B1, not to its proposed split-gate invention.
Drawn stored carriers represent data, not the source of read current.
Select the measured path and apply low-field read conditions.
The stored state determines sense current under the same read bias.
Compare stored states under equal read conditions; logic encoding is not assigned.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
SSI / FN
WL-poly erase exit of the first/second generation; third-generation dedicated erase gates and another patent’s well-erase conditions are not substituted.
Drawn stored carriers represent data, not the source of read current.
Select the measured path and apply low-field read conditions.
The stored state determines sense current under the same read bias.
Compare stored states under equal read conditions; logic encoding is not assigned.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
SSI / FN
SG is below an overlapping FG; source is raised for program, while SG/S/D float and the well is raised for erase.
Drawn stored carriers represent data, not the source of read current.
Select the measured path and apply low-field read conditions.
The stored state determines sense current under the same read bias.
Compare stored states under equal read conditions; logic encoding is not assigned.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
NOR does not require an entire string of storage cells to conduct before one cell can be read, which favors random access. An overerased stacked-gate cell that conducts while unselected can corrupt the shared bitline; a separate selection function helps shut off that path. Actual latency still includes wordlines, bitlines, sensing, and the external interface, not only individual-cell speed.
NOR selects cells through wordlines and bitlines, and leakage from unselected rows directly affects sensing. A split gate adds a channel segment that can be turned off independently, addressing the selection problem associated with overerase. Programming, erase, and reading use different terminal combinations, so selected and unselected paths must be redrawn for each operation rather than inferring program/erase safety from a read circuit alone.
Floating-gate coupling, oxide thickness, and injection location change the threshold-voltage distribution. Cycling defects and erase nonuniformity widen it further. Split gates reduce certain array-leakage risks but do not eliminate charge-retention requirements, read disturb, or program/erase stress in the storage region. Comparisons must hold product generation, temperature, and cycling conditions constant.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
Injection efficiency, tunnel-dielectric quality, and the threshold-voltage window jointly limit program/erase speed and lifetime. Shrinking a cell can increase sensitivity to coupling, short-channel effects, and electron count. A generation's typical endurance must not be treated as the physical limit of the NOR family.
Bitline contacts and wiring overhead constrain density, while long-line capacitance and leakage constrain sensing speed. Split gates can improve unselected-cell isolation, but erase control, references, and redundancy remain necessary. Comparisons should report capacity, array efficiency, and peripheral area.
Floating gates, inter-gate dielectric, and injection-region geometry require dedicated process control. Different split-gate generations and erase exits are not interchangeable modules. Integration into a logic process requires checking added masks, thermal budget, voltage-tolerant devices, and qualification data.
Code-read performance depends on the interface, controller, cache, and package, and execute-in-place requires a complete timing contract. Data updates are constrained by block erase, read availability during programming, power-loss handling, and cycle budgets. The fastest quoted read time alone cannot determine the solution.
Suitable for boot code, firmware requiring predictable random reads, and resident code of moderate capacity. Execute-in-place must be matched to the processor bus, controller, cache, and actual worst-case access time. If firmware must be updated, additionally validate dual-image handling, erase granularity, and power-loss recovery.
For workloads prioritizing large sequential datasets and low cost per bit, compare against NAND's management cost and complete system performance. Frequent small overwrites should not be mapped directly to NOR block erases without first estimating write amplification, movement of retained data, and cycle consumption.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US6232180B1
Follow floating-gate formation in Figure 5 into the source/drain structure in Figure 6. Locate the select gate and tunnel oxide; the nested wells support separately controlled erase bias.
Compare the third NOR variant: source-side injection and well/channel-side FN erase, separately from implementations that tunnel toward a select gate.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
If two cells are both called split-gate NOR, can they share the same erase-bias table?
No. The SuperFlash brochure's example uses inter-gate FN tunneling, while US6232180B1 provides a channel-erase embodiment. Identify each electron exit, well, and floating terminal before citing its biases. A shared family name does not establish identical operation.
Physics Background 06 · Charge Storage
SONOS/MONOS describe material stacks or gate materials, while NROM refers to an implementation lineage that uses localized trapping and read direction. They are not directly interchangeable product names. Infineon's SONOS has production platforms, so the entire charge-trap family must not be labeled emerging. Its FN program/erase mechanism and reliability figures also must not be transferred to every NROM implementation.
The same storage cell supports programming, electrical erase and subsequent programming. ERS restores a window suitable for another program operation; it does not require every carrier to disappear. Whether the host issues a separate erase command depends on the macro or component interface.
PGM → ERS → PGM
Electrons remain in trapping centers within insulating materials such as silicon nitride, changing the potential seen by the channel and its threshold voltage. The trapping layer is not a conductive floating gate, and charge can have a spatial distribution. Channel-wide SONOS program/erase examples and localized NROM charge-trapping examples therefore require different operating and sensing explanations; simply recoloring a floating gate is not sufficient.
Electrons Occupy Localized Traps in an Insulator, Unlike a Conductive Floating Gate.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
Storage-transistor detail; Infineon’s 2T cell also has a series selector. Carrier paths follow the public Cypress patent without asserting a current macro’s complete stack.
Isolated storage starts within its programmable window.
Terminal conditions belong only to the named variant.
Electrons tunnel through the local barrier into isolated storage.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
CHE
S/D keep their programming-time names; reverse read changes bias and current direction without silently renaming terminals.
Isolated storage starts within its programmable window.
Terminal conditions belong only to the named variant.
Channel electrons accelerate before local injection into storage.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
In a tunneling example, charge crosses the thin dielectric from the channel side and is then trapped. In a hot-electron example, carriers first gain energy near the drain, producing localized stored charge. An insulating trapping layer does not automatically equalize potentials and charge distributions as a conductor does. The product page does not disclose the complete film thicknesses and terminal biases, so this explanation remains at the verified mechanism level.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
Storage-transistor detail; Infineon’s 2T cell also has a series selector. Carrier paths follow the public Cypress patent without asserting a current macro’s complete stack.
Locate programmed charge and this variant’s exit.
Terminal conditions belong only to the named variant.
Holes enter the trap layer and reduce net stored negative charge.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
BBT / HHI
Follows the one-sided pocket of Figures 8A and 9; hole injection must overlap the stored-electron region. This separate erase example is not attributed to US5768192A.
Locate programmed charge and this variant’s exit.
Negative gate and positive drain establish BBT and a local field at the pocket junction.
Holes enter the trap layer and reduce net stored negative charge.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Read US6664588B2 Figures 8A and 9–11 as a named pocket-implant erase example. Do not attribute that structure to US5768192A or substitute the SONOS FN diagram for NROM hot-hole erase. The cited implementation defines selection, pulse verification and erase granularity.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
Storage-transistor detail; Infineon’s 2T cell also has a series selector. Carrier paths follow the public Cypress patent without asserting a current macro’s complete stack.
Drawn stored carriers represent data, not the source of read current.
Select the measured path and apply low-field read conditions.
The stored state determines sense current under the same read bias.
Compare stored states under equal read conditions; logic encoding is not assigned.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
CHE
S/D keep their programming-time names; reverse read changes bias and current direction without silently renaming terminals.
Drawn stored carriers represent data, not the source of read current.
Bias original S and ground original D, reversing the program direction.
The stored state determines sense current under the same read bias.
Compare stored states under equal read conditions; logic encoding is not assigned.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Read current travels through the channel; it does not empty all electrons from the nitride. A channel-wide threshold model explains many SONOS operations, but ignoring charge position in a localized cell can obscure why reverse reading works. Read direction must be labeled against the original programming terminals, rather than silently exchanging source and drain names in a drawing.
In Infineon's 2T SONOS, a MOS selector is in series with the charge-trap storage transistor, separating isolation and storage functions. High-field program/erase also requires consideration of shared wordlines and wells. Localized-charge arrays must additionally manage direction, half-selection, and neighboring data states. Successful reverse reading of one cell does not establish reliable storage of two independent bits in a large array.
Trap density, energy levels, interface quality, and film thickness affect programming speed, detrapping, and retention. Localized charge also varies in lateral position and distribution width. Compare windows initially, after cycling, and after high-temperature retention. Extending the concept to storage at both ends requires measuring their interaction rather than simply counting an additional possible storage position.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
Program/erase speed and retention depend on the tunneling barrier, trap depth, and blocking layer. Charge that is easier to inject may also escape more readily. Cycling-induced defects change the window, while localized charge is additionally affected by short-channel behavior and lateral distribution. The SONOS name alone does not establish a fixed endurance ceiling.
Selectors and shared biases determine isolation, granularity, and effective density. Localized multibit implementations must keep spatially distinct states distinguishable while accounting for programming direction, half-selection, and neighbor interference. Two charge clusters in a cell illustration do not establish two reliable deliverable bits.
The upper/lower oxides, nitride, and interfaces must retain appropriate traps and barriers within the thermal budget. Changing the metal gate or blocking layer also changes field distribution. Production qualification for an embedded process belongs to a particular stack, node, and macro combination and cannot be extended to arbitrary charge-trap devices.
The system must still budget program/erase latency, supply requirements, post-cycling retention, and ECC conditions. A macro family listing fast reads and long retention does not establish simultaneous operation at maximum temperature, maximum capacity, and maximum accumulated cycles. Obtain the fully paired conditions before committing to an application lifetime.
Suitable for embedded code and parameter storage where the identified process has qualification evidence. When selecting a solution such as Infineon SONOS, obtain the corresponding macro data for the node, capacity, temperature, and read/write requirements. NROM is also useful for understanding the principles and costs of localized charge, reverse reading, and spatial multibit encoding.
Do not treat SONOS, MONOS, NROM, and 3D NAND as one interchangeable cell. Requirements involving two-bit storage, very-high-temperature retention, or a specific cycling guarantee need complete product or measurement data for that implementation. The best figures from a different charge-trap family must not be used to fill missing evidence.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
WO1981000790A1
Read upward from silicon: thin memory oxide, silicon nitride, interfacial oxide and polysilicon gate. The drawing explains the dielectric stack.
For SONOS operation, distinguish the lower tunnel oxide from the upper blocking oxide. Stored charge resides in the nitride.
Compare All Drawings, Numerals and Claims →US5768192A
Compare the prior-art A panels with embodiment B panels. Trace the localized charge region and READ arrow; reading direction changes which end of the channel barrier controls current.
Compare the localized NROM sequence: CHE stores electrons near one end, reverse read senses from the opposite direction, and BBHH erase is explained with its separate source.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
If a SONOS nitride layer is drawn as one continuous film, does that mean all its charge can freely equalize as it would on a metallic floating gate?
No. Nitride is an insulating trapping material, so charge can remain at different positions. Readout of a localized-charge cell consequently depends on charge position and read direction. Distinguish continuity of the material, localization of charge, and the way the full channel is sensed.
Physics Background 07 · Charge Storage
Series connection increases density by sharing contact overhead across cells, but reading one cell requires the other cells in its string to provide a conduction path. Planar feature shrink, additional 3D layers, and more bits per cell are distinct density axes, each with charge-window, process, and reliability costs. Page programming and block erase also make the controller an important part of usable storage.
The same storage cell supports programming, electrical erase and subsequent programming. ERS restores a window suitable for another program operation; it does not require every carrier to disappear. Whether the host issues a separate erase command depends on the macro or component interface.
PGM → ERS → PGM
NAND data can still be retained by charge in a floating gate or dielectric trapping layer that changes threshold voltage. NAND itself describes an array organization with multiple cells connected in series. Storing N bits per cell requires 2 to the Nth power distinguishable states, such as eight for TLC and sixteen for QLC. More bits do not provide additional windows of unchanged width for free.
Stacked Word Lines Share a Vertical Channel. The Front Cutaway Exposes the Radial Material Sequence.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
A string topology with an unfolded local film section. Erase follows source-line raising and electron release, not an asserted GIDL hole mechanism.
WL* identifies the target level; SGD/SGS control terminal access.
BL=0 keeps the selected channel low; target WL receives V_PGM and neighbors V_PASS.
The local section traces CH through tunnel dielectric into CTL, where electrons are trapped.
The low channel at left programs; BL=V_DD at right precharges a floating, boosted channel and reduces tunneling field.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
The early US7696559B2 example uses Vpgm on the selected wordline and lower pass potentials on the others, with channel conditions inhibiting unwanted programming. Both planar and 3D implementations depend on the effective field across the storage layer, not merely the wordline-to-ground voltage. Specific channel boosting, pulse stepping, and selector timing require verification for each implementation.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
A string topology with an unfolded local film section. Erase follows source-line raising and electron release, not an asserted GIDL hole mechanism.
Locate programmed charge and this variant’s exit.
Terminal conditions belong only to the named variant.
Electrons leave storage through this variant’s specified exit.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
GIDL / FN
A positive terminal above the select-gate potential generates electron–hole pairs. Only the upper supply is expanded; the other end depends on the example. Separate from the older floating-terminal erase.
Locate programmed charge and this variant’s exit.
A positive terminal above the select gate separates electron–hole pairs and supplies channel holes.
Holes enter the trap layer and reduce net stored negative charge.
The state shifts toward its target window; residual charge and defects are not assumed absent.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
Erase granularity follows shared wells, channels, and wordline-bias control, and differs from a read page or a host address. Planar floating-gate cells and modern 3D stacks may use different charge-removal or neutralization paths. This material preserves the boundary of the early embodiment and does not present one common-source bias as the erase specification of all NAND.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
FN
A string topology with an unfolded local film section. Erase follows source-line raising and electron release, not an asserted GIDL hole mechanism.
Drawn stored carriers represent data, not the source of read current.
Select the measured path and apply low-field read conditions.
A low-threshold selected cell and pass-biased neighbors permit BL discharge; electrons travel SL to BL.
Comparison branch: a high-threshold selected cell blocks the string; reading did not change its charge.
A principle drawing redrawn from public sources, not a process cross-section to scale. Arrows represent the stated carrier or field; numerical design requires device-specific specifications.
When one cell is read, the other storage cells still experience pass bias, so the absence of a program command does not mean zero read disturb. String resistance, wordline/bitline loading, and threshold-voltage distributions jointly affect readout. In US7696559B2, the selected 0 V condition belongs only to that early example; one fixed potential cannot distinguish all multilevel states.
NAND selection has three levels: the block defines the shared control scope, the wordline identifies a position in the string, and bitlines with string selectors determine which strings are programmed or sensed. Half-select inhibition requires management of channel potential and pass stress. Unselected wordline potentials must enable the path without accumulating excessive disturb, so reporting only the selected wordline voltage is insufficient.
Planar scaling increases sensitivity to small charge losses and neighboring-cell coupling. 3D introduces hole-profile, layer-to-layer process, and channel differences. Abstracts of original measurement studies also identify early retention loss and retention interference. Narrower multilevel windows make these distribution differences more likely to cause errors, so analysis must include cycling, temperature, retention time, and ECC conditions together.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
The usable threshold-voltage range is finite; more bits divide that same range into more states. Charge loss, trapping/detrapping, cycling defects, and interference broaden distributions. Physical limits must be expressed through distinguishable state count and specified retention/cycling conditions, not minimum cell area alone.
String length, selectors, pass bias, and wordline loading affect read/program timing and disturb. Layer variation in 3D and multilevel distributions also require reference adjustment, verification, and error correction. Effective density must account for redundancy and management overhead rather than treating layer count multiplied by bits per cell as usable capacity.
High-aspect-ratio 3D memory holes require balanced control of profile, uniformity, etch time, and subsequent film quality. More layers also increase manufacturing time and cost risk. Kioxia's deep-hole process research identifies productivity as a key bottleneck; a layer-count record alone does not demonstrate lower cost per bit.
Host-visible performance and lifetime depend on ECC, data movement, redundancy, the controller, and workload. Cycle count, temperature, and retention time jointly determine reliability. Performance after the cache fills, worst-case tail latency, and write amplification require measurement; a NAND die's peak transfer rate alone is insufficient.
Suitable for SSDs, managed flash, and systems requiring large-capacity data storage. Selection must distinguish raw NAND from products with a controller and evaluate lifetime against actual read/write mix, retention time, operating temperature, and written data volume. Capacity, bits per cell, and ECC conditions must be tied to an identified product.
Raw NAND should not be treated as directly and arbitrarily overwritable byte-addressable memory. Individual-device speed or performance with an empty cache is also not representative of application performance. When the primary requirements are small capacity, direct random reads, or highly predictable latency, compare alternatives with controller and management costs included.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US7696559B2
In Figure 2, trace the silicon pillar from the common source to the bit line. Figure 6 unfolds the same structure into a string circuit, with select gates at both ends.
Compare selected-word-line injection, pass biases and program inhibit. This patent’s source-side electron removal is shown separately from later GIDL hole-assisted erase.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
Does moving from TLC to QLC while increasing the layer count necessarily improve speed and endurance and reduce cost at the same time?
No. QLC must distinguish sixteen states, while TLC needs only eight; narrower windows increase verification and error-management demands. More layers add deep-hole process and layer-variation challenges. Evaluate an identified product's usable capacity, manufacturing and controller costs, workload, and reliability conditions together.
Physics Background 08 · Magnetism
Toggle MRAM assigns data retention to the magnetic energy barrier and data modification to precisely sequenced magnetic fields. The controller first determines whether the existing and requested values differ, then toggles only when needed. It does not require the block-erase sequence of Flash, but it adds read, comparison, and toggle control. This complete sequence is essential to a valid comparison of write latency and energy.
MRAM overwrites existing data by changing magnetic state, without a Flash-style erase-before-program step. Clearing to all zeros or ones is a series of target-state writes; the P/AP-to-data mapping is product-specific.
P ⇄ AP
The bit is stored in the magnetization direction of the free magnetic layer. Parallel and antiparallel alignment relative to the reference layer produce different resistance levels in the magnetic tunnel junction. A magnetic energy barrier maintains the direction after power is removed. Toggle specifically denotes a write method that reverses data through the rotation of coupled magnetic moments; it is not a general name for all field-written MRAM.
In-Plane Magnetization Example: Orthogonal Conductors Supply Switching Fields; Junction Resistance Is Read.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Compare data before following H1, overlapping H1/H2, H2, and field removal; reading senses junction resistance.
M1 starts parallel at low R to REF. A read/compare decision precedes toggling.
H1 alone starts spin-flop rotation of the SAF moments while they remain largely antiparallel.
H2 rises before H1 ends. The resultant field changes direction and both moments continue along the same rotational sense.
H1 falls first. H2 continues driving the moments beyond the hard-axis instability.
After H2 falls, the moments return to the easy axis. M1 has reversed about 180°, reaching antiparallel high R.
M1 is the free sublayer next to the barrier; its angle to REF determines P/AP. Intermediate angles and pulse heights are not measured. Rewriting the opposite data uses another toggle sequence, with no block erase.
One Toggle sequence reverses the existing state. If the target and existing values match, toggling must be skipped. Products may define their own logic encoding; low resistance does not necessarily represent 0.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Compare data before following H1, overlapping H1/H2, H2, and field removal; reading senses junction resistance.
M1 starts antiparallel at high R to REF. A read/compare decision precedes toggling.
H1 alone starts spin-flop rotation of the SAF moments while they remain largely antiparallel.
H2 rises before H1 ends. The resultant field changes direction and both moments continue along the same rotational sense.
H1 falls first. H2 continues driving the moments beyond the hard-axis instability.
After H2 falls, the moments return to the easy axis. M1 has reversed about 180°, reaching parallel low R.
M1 is the free sublayer next to the barrier; its angle to REF determines P/AP. Intermediate angles and pulse heights are not measured. Rewriting the opposite data uses another toggle sequence, with no block erase.
Restoration here is performed through individual-bit rewriting, not a physical erase that clears the whole array. The result of the same toggle sequence depends on the current state, so the initial read and decision cannot be omitted.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Compare retained P/AP states at the same small bias, latch the current difference, then remove bias while retaining each alignment.
The two diagrams represent possible P and AP states. Their access paths are open and carry no read current.
Close the access paths at the same small bias. P carries more current and AP less current.
The sense circuit latches the current difference and removes read bias. Each magnetic alignment is retained.
M1 is the free sublayer next to the barrier; its angle to REF determines P/AP. Intermediate angles and pulse heights are not measured. Rewriting the opposite data uses another toggle sequence, with no block erase.
Reading requires separation between resistance distributions while keeping disturbance and stress within acceptable limits. Reference cells, temperature, and process variation affect the read margin.
For reading, the word line normally enables the access transistor and the bit line connects the cell to the sensing circuit. Writing selects the target through the sequence and intersection of two lines; cells that are not fully selected may still experience partial magnetic fields. Verification must therefore examine magnetic-field trajectories and data retention in half-selected cells, as well as the logical address and successful switching of the target cell.
Free-layer dimensions, coupling strength, magnetic anisotropy, and line current change the safe switching region. Circuits must keep process and temperature distributions within the qualified timing window, using write verification, redundancy, and error correction to control tail failures. If these variations are used for a PUF, reproducibility must also be demonstrated; occasional write errors are not automatically a usable fingerprint.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
The magnetic energy barrier must be high enough to retain data yet permit reliable reversal by a field generated with reasonable current. An improperly designed coupled free layer can follow an incorrect rotation trajectory or lose stable states.
The spacing, current distribution, and half-select disturbance of two write-line sets constrain array scaling. If the lines cannot scale with the bit, a smaller cell does not produce a proportional improvement in effective density.
Magnetic-layer deposition, coupling-layer thickness, and junction etching affect magnetic uniformity. Thermal processing must accommodate the magnetic layers, tunnel barrier, and CMOS interconnects; package reflow conditions also require separate validation.
Pre-write reading, comparison, and worst-case sequencing belong in the access time. Power-failure atomicity, bus transfers, and capacity cost determine system benefit; magnetic reversal time cannot substitute for complete write latency.
Suitable for small records, industrial control, and equipment-state retention when available products meet capacity requirements, updates are frequent, and data must survive power loss. Selection should check actual capacity, interface, operating temperature, and magnetic-field conditions, considering the availability of mature products together with their lifetime conditions.
For maximum bit density, low-cost bulk storage, or very small embedded caches at advanced nodes, commercial Toggle availability alone does not establish the best technology choice. Write-line area, drive current, and comparison operations may dominate cost; compare concrete STT implementations and other options with suitable interfaces.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US6545906B1
Start with the two pulse waveforms in Figure 4, then follow the moments through Figures 5 and 6. The overlap interval and turn-off order are part of the operation.
Compare the five-frame Toggle sequence and initial-state check: apply a toggle when the stored bit needs to change.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
If the stored data already matches the target value, why must the controller avoid executing another Toggle sequence?
A Toggle write reverses the current state. Executing it once would invert data that is already correct. The controller must read and compare first, toggling only when the existing and target values differ.
Physics Background 09 · Magnetism
STT concentrates write current in the selected junction and improves on the scaling limitations of magnetic-field write lines, making it an important route for commercial discrete and embedded MRAM. Increasing current can shorten switching time, but also raises access-transistor requirements and barrier stress. Reducing current can lengthen latency and worsen the error-rate tail. The best speed, lifetime, and density values from separate conditions cannot be combined into one product specification.
MRAM overwrites existing data by changing magnetic state, without a Flash-style erase-before-program step. Clearing to all zeros or ones is a series of target-state writes; the P/AP-to-data mapping is product-specific.
P ⇄ AP
The stored quantity remains the orientation of the free magnetic layer relative to the reference layer, and reading relies on the resistance difference of the magnetic tunnel junction. The principal difference from Toggle is writing: current passing through the magnetic stack carries spin angular momentum and exerts torque on the free layer, changing its magnetic state. Nonvolatility comes from the magnetic energy barrier, not from keeping current inside the device.
Perpendicular-Magnetization Example: Current through the Barrier Supplies Spin-Transfer Torque; Read the P/AP Resistance Difference.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Separate conventional current from electron flow, then follow spin torque, free-layer reversal, and low-bias sensing.
mF starts parallel to REF; no current crosses the junction.
Enable WL. Electrons flow from upper free layer to lower REF. Orange Ic points oppositely while the free moment deflects.
After current stops, mF settles in AP at high R while REF retains its direction.
Here positive conventional current is upward and electron flow is free-to-reference, illustrating P→AP; reverse flow illustrates AP→P. Verify polarity for the actual stack. Switching probability and read disturbance are not quantified.
Switching has a probability distribution. Pulse amplitude and duration must cover process, temperature, and the target error rate. A teaching diagram's current arrow represents one stack convention; actual direction must be checked against the electrodes and current convention.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Separate conventional current from electron flow, then follow spin torque, free-layer reversal, and low-bias sensing.
mF starts antiparallel to REF; no current crosses the junction.
Enable WL. Electrons flow from lower REF to upper free layer. Orange Ic points oppositely while the free moment deflects.
After current stops, mF settles in P at low R while REF retains its direction.
Here positive conventional current is upward and electron flow is free-to-reference, illustrating P→AP; reverse flow illustrates AP→P. Verify polarity for the actual stack. Switching probability and read disturbance are not quantified.
STT can directly rewrite either magnetic state without first erasing a whole block. A clear command in a particular product may be a controller or interface function; it does not establish a Flash-like physical erase mechanism.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Enable low-bias sensing, compare P/AP junction currents, then latch and isolate the paths while retaining free-layer magnetization.
The two diagrams represent possible P and AP states. Their access paths are open and carry no read current.
Close the access paths at the same small bias. P carries more current and AP less current.
The sense circuit latches the current difference and removes read bias. Each magnetic alignment is retained.
Here positive conventional current is upward and electron flow is free-to-reference, illustrating P→AP; reverse flow illustrates AP→P. Verify polarity for the actual stack. Switching probability and read disturbance are not quantified.
Reading also produces spin torque, so the signal cannot be increased without limit. The design must simultaneously satisfy sensing speed, read margin, and the long-term cumulative probability of read disturbance.
The access transistor in a 1T1MTJ cell isolates unselected cells and supplies write current. The word line and bit/source lines jointly determine the address and current direction. The transistor must drive the MTJ at worst-case voltage and temperature, so its area may become a bottleneck before the magnetic device itself. Reference cells, sense amplifiers, and spare rows also count toward macro density.
MTJ diameter, barrier thickness, magnetic anisotropy, and reference-layer properties create distributions of resistance and critical current. Thermal fluctuations add a probabilistic switching tail, so average write latency alone is insufficient. Grouped pulse settings, write verification, ECC, redundancy, and read-reference design can reduce errors, but they add energy, time, and peripheral area.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
Shrinking the free layer reduces its magnetic energy barrier and may weaken high-temperature retention. Increasing anisotropy to restore stability can increase write current. Barrier breakdown lifetime and stochastic switching tails jointly constrain the operating window.
Access-transistor size, bit-line voltage drop, and write-polarity asymmetry affect writability at the worst-case location. Large arrays must also account for reference drift, tail bits, redundancy, and the actual area overhead of ECC.
MTJ etch redeposition, sidewall damage, and barrier uniformity directly affect shorts and the resistance ratio. The thermal budget must preserve magnetic properties while supporting interconnects and package reflow. Demonstrating room-temperature switching in a cell is insufficient.
Complete latency includes the bus, macro access, possible write verification, and ECC. Higher-level power-failure consistency, cache writeback, and update frequency determine endurance requirements. A DDR-derived interface does not make every behavior equivalent to DRAM.
Suitable for code, equipment state, data logging, and some persistent working memory that require power-off retention and frequent updates. MCU or SoC integration must jointly confirm available capacity, process options, write latency, temperature-dependent lifetime, and software update strategy. Select a concrete macro rather than the MRAM name alone.
Generic STT figures are insufficient for decisions involving extremely high density, the lowest bit cost for long-term bulk storage, or unlimited high-frequency writes without workload conditions. Cache replacement also requires worst-case write error rate, energy, and read-disturb checks, rather than average switching speed alone.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US5695864A
Figure 1 is a five-layer metallic-conductor model. Follow A→F1→B→F2→C to locate fixed and variable moments, then relate current to torque on F2.
Use this patent for spin-transfer physics, then the modern STT-MRAM plates for P/AP resistance sensing and the tunnel barrier.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
Why cannot STT-MRAM read current simply be increased indefinitely to accelerate sensing?
Read current also passes through the MTJ, producing spin torque and electrical stress. Higher current increases the signal but may also increase read disturbance and barrier stress. Sensing margin and error rate must be considered together.
Physics Background 10 · Magnetism
SOT seeks short write pulses and lower barrier stress by separating the read and write paths, making it a focus of last-level-cache research. However, the third terminal and extra line consume area, and deterministic field-free switching, large-array yield, and process integration must also be established. Low-energy or high-cycle-count cell demonstrations satisfy only part of that validation.
MRAM overwrites existing data by changing magnetic state, without a Flash-style erase-before-program step. Clearing to all zeros or ones is a series of target-state writes; the P/AP-to-data mapping is product-specific.
P ⇄ AP
SOT-MRAM also retains data in the magnetization direction of an MTJ free layer and senses it through magnetoresistance. Its distinguishing feature is that write angular momentum is generated primarily by a spin-orbit material beside or beneath the free layer and injected into it, rather than by sending the main write current through the tunnel barrier. The stored physical quantity is therefore similar to STT, while the write structure and array cost differ.
Three-Terminal Topology: Lateral Write Current and Vertical Tunneling Read Current Use Separate Paths.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
A lateral write line supplies spin injection; an independent upper terminal and MTJ branch provide sensing.
The free layer retains its initial state. R is isolated, so write current need not cross the barrier.
A calibrated W1/W2 pulse injects Js into the free layer. Explicit Hassist supplies symmetry breaking for this example.
Turn off the lateral pulse. The free moment relaxes toward the target equilibrium under effective fields and damping.
The free layer settles in the opposite state. The write line carries no current, and the MTJ can be read independently later.
Hassist is explicit; no deterministic field-free switching is assumed for a lone ideal line. I+/I− are opposite calibrated write directions; their state mapping depends on spin-Hall sign, stack orientation, and assist field.
A perpendicularly magnetized system generally needs an additional structure or mechanism to select the final direction reliably. A schematic must not assume that one ideal line alone provides deterministic writing with zero external magnetic field.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
A lateral write line supplies spin injection; an independent upper terminal and MTJ branch provide sensing.
The free layer retains its initial state. R is isolated, so write current need not cross the barrier.
A calibrated W1/W2 pulse injects Js into the free layer. Explicit Hassist supplies symmetry breaking for this example.
Turn off the lateral pulse. The free moment relaxes toward the target equilibrium under effective fields and damping.
The free layer settles in the opposite state. The write line carries no current, and the MTJ can be read independently later.
Hassist is explicit; no deterministic field-free switching is assumed for a lone ideal line. I+/I− are opposite calibrated write directions; their state mapping depends on spin-Hall sign, stack orientation, and assist field.
This is direct rewriting, with no inherent block-erase requirement. The relationship between current polarity and magnetic state depends on the SOT material, stack orientation, and field-free switching method.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Sense from R through the MTJ and return via W2 while W1 stays isolated; latch and remove read current while retaining magnetization.
The illustrated P state is retained with R and W1 isolated; no holding current is required.
Read current passes from R through REF/barrier/free layer and returns via W2. W1 is isolated, so no W1→W2 write drive is applied.
After latching the sensed value, open the R branch and stop read current; the free-layer direction is retained.
This sequence illustrates low-bias sensing of retained P; AP can be sensed through the same path. Set actual read bias and duration according to the MTJ stack and read-disturb constraints.
Separate paths can reduce the barrier stress from the main write current, but reading must still satisfy disturbance, leakage, resistance-distribution, and reference-circuit requirements.
Selection must manage both the lateral SOT line and the vertical MTJ read branch, using different transistor and shared-line arrangements as appropriate. Sharing can reduce area but introduces current through unselected devices, line-resistance effects, and current-distribution concerns. Density assessment must show all terminals, access transistors, and lines rather than comparing only MTJ diameters.
Differences in SOT conversion efficiency, line thickness, magnetic-layer dimensions, interface roughness, and field-free switching structures change the required current and the probability of the final magnetic state. Array validation must examine worst-case tails, thermal conditions, and mutual disturbance rather than only representative devices. Using stochastic switching for probabilistic computing requires retention specifications and statistical validation to be redefined.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
Spin-torque efficiency, magnetic stability, and deterministic switching must be achieved together. A material that lowers switching current may introduce high line resistance. Device pulse energy must include actual resistance and switching success probability.
Three terminals and separate paths add selection-area cost, while shared lines introduce mutual disturbance and current shunting. Large arrays require validation of writability, sensing margin, and error-rate tails at every location.
Alignment, etching, sidewall cleaning, and interface quality of the SOT line and MTJ affect both torque and read quality. Integration with advanced logic and subsequent thermal processing can also damage materials, requiring validation of the complete process flow.
Last-level-cache requirements include capacity, standby leakage, worst-case write latency, and bandwidth, as well as ECC and coherence costs. Cell-level fJ energy and high cycle counts do not directly establish the energy per access or lifetime of a complete cache.
Suitable for high-speed memory research combining frequent updates, short latency, and power-off retention, particularly where more complex selection and material integration are acceptable. Research programs should develop device, functional-array, and macro models together so that the benefits of separate read and write paths can be verified in effective density and system energy.
A SOT research record alone does not establish production readiness for near-term products needing standard off-the-shelf chips, complete automotive qualification data, or a process replacement without integration risk. If area is tightly constrained or external magnetic fields and complex write assists are unacceptable, first verify the specific field-free approach and its array-selection cost.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US10930843B2
Trace the first horizontal wire through the magnetic stack to wiring in the other direction. Access transistors and crossing interconnect explain array integration beyond one MTJ.
Compare three-terminal SOT read/write separation. Preserve the control terminals rather than reducing them to the STT current path.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
Why can a SOT macro still be larger than an STT macro even when the main write current is moved out of the MTJ?
SOT typically needs a third terminal, a SOT line, and additional selection paths. Shrinking the MTJ does not mean the entire bit and its peripheral circuits shrink together. Effective macro density must be calculated.
Physics Background 11 · Resistive Switching
VCM operation centers on controlling reversible local changes without driving the oxide into permanent breakdown. SET commonly lowers resistance, while RESET raises it. Some stacks require initial current-limited forming to activate a conductive path. Rebuilding the path may differ slightly on each cycle, making the relationship among forming, write verification, cycling distributions, and retention more important than one attractive typical I–V curve.
RESET is the reverse data-state update from low to high resistance; a later SET restores low resistance. This is reversible resistance switching rather than Flash-style block erase. The circuit defines logical zero/one encoding.
SET → RESET → SET
VCM stores data in the ionic distribution, local redox state, or interfacial barrier of an oxide, producing distinguishable resistance states. A typical filament model explains conduction and rupture through redistribution of oxygen ions/oxygen vacancies, but not every device has a single clearly defined filament. Materials, electrodes, and measurement evidence determine the mechanism; a hysteretic I–V curve alone is insufficient to identify VCM.
Defect Redistribution Alters a Local Conduction Path; This Mechanism Is Not Disclosed for Every Commercial ReRAM.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Follow oxygen-ion exchange, oxygen-vacancy distribution, and a local gap through SET, RESET, and low-bias read.
The plate begins in a formed, RESET high-R state. A locally oxidized gap interrupts the oxygen-deficient path.
With the selected positive TE bias, O²− moves toward the upper exchange interface, leaving oxygen-deficient sites along the path.
The local vacancy-rich path connects and resistance falls. Ilim limits excessive path growth and Joule heating.
After SET bias is removed, the oxygen-deficient path remains connected without a holding voltage.
Positive upper-electrode bias drives O²− toward the upper interface in this selected convention. Vacancies are not metal ions; actual VCM can switch at interfaces or broader channels. Forming is distinct from recurring SET.
Current compliance controls path growth. Excessive SET can make RESET difficult or even cause hard breakdown. Forming is not required for every write, and not every process requires it.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Follow oxygen-ion exchange, oxygen-vacancy distribution, and a local gap through SET, RESET, and low-bias read.
After SET, an oxygen-deficient region connects the electrodes and retains low resistance without bias.
Reversing TE bias in this model returns O²− from the exchange region toward the local path; arrows indicate oxygen-ion motion.
Oxygen reincorporation interrupts the narrowest path segment. The entire pre-existing path need not disappear.
After RESET bias is removed, the local gap and residual oxygen-deficient regions remain for a later high-R read.
Positive upper-electrode bias drives O²− toward the upper interface in this selected convention. Vacancies are not metal ions; actual VCM can switch at interfaces or broader channels. Forming is distinct from recurring SET.
RESET is a local resistance-state transition, not a Flash-like block erase. Polarity, thermal effects, and path morphology depend on the material. One bipolar operating mode cannot define the entire family.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Follow oxygen-ion exchange, oxygen-vacancy distribution, and a local gap through SET, RESET, and low-bias read.
Left and right are the low- and high-resistance alternatives for one cell. Read bias has not yet been applied.
Compare currents at the same small read bias, chosen to avoid appreciable ionic redistribution.
After the sense circuit latches the difference, current stops. The connected low-R path and local high-R gap remain retained.
Positive upper-electrode bias drives O²− toward the upper interface in this selected convention. Vacancies are not metal ions; actual VCM can switch at interfaces or broader channels. Forming is distinct from recurring SET.
Read bias and accumulated read count may cause disturbance. Check worst-case overlap of high- and low-resistance distributions, temperature variation, line resistance, and sensing noise.
In 1T1R, the word line controls the access transistor, while bit and source lines apply pulses; the transistor also limits current. In 1S1R, strong selector nonlinearity distinguishes fully selected from half-selected biases. The approaches differ in area, forming capability, and sneak paths. A device that switches under a probe does not necessarily have a usable array-selection window.
Device-to-device variation must be distinguished from cycle-to-cycle variation within the same device. Path location, width, and defect distribution cause SET/RESET voltages and resistance values to vary. Multipulse verification, differential encoding, ECC, and calibration can reduce the impact. Analog weights additionally require validation of update linearity, symmetry, effective state count, and read noise; demonstrating several resistance values is insufficient.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
The conductive path must switch with bounded pulses while remaining stable at high temperature and low read bias over time. An overly thick path is difficult to reset; an overly thin path may relax. Defect generation can also accumulate into permanent breakdown.
Sneak current and line resistance alter the bias actually reaching the target device, and forming may exceed the selector operating window. Large arrays require read/write reliability validation across position, half-select count, and the tails of resistance-state distributions.
Oxygen content, electrode oxygen affinity, interfacial layers, and deposition uniformity govern reversible switching. Interconnect thermal budgets and contamination requirements must both be met. Low-temperature material deposition does not establish reliability qualification of the complete process.
Write verification, ECC, redundancy, and wear management add time and energy. In-memory computing may be dominated by ADCs, DACs, line-resistance compensation, and model-level error tolerance; device conductance alone does not establish system cost.
Suitable for embedded NVM when a concrete process provides design and qualification support and macro specifications cover the required capacity and endurance. Research applications include analog weights and small crosspoint computing arrays that can tolerate calibration and error compensation, provided array and peripheral costs are counted alongside device energy.
VCM variation and selection requirements may be unsuitable when every write must reach exactly the same analog resistance, when a fixed short write time is required without verification, or when two-terminal devices are assumed to scale directly into an arbitrarily large array. A best-case pulse from one device or an image showing successful forming cannot replace high-capacity yield and post-cycling retention data.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US8331131B2
Follow the Figure 5 cycle through mobile-species, barrier and conductive-region changes. The intermediate state has a physical location; it is not a single-step SET/RESET diagram.
Compare species redistribution in VCM operation; this patent additionally shows how pulse sequencing introduces an intermediate state.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
Why does a hysteretic current–voltage curve alone not establish that a device uses oxygen-vacancy VCM?
Several ionic, interfacial, thermal, or electronic mechanisms can produce resistance hysteresis. Materials and electrodes, polarity, time and temperature response, and structural evidence must be considered together to identify the dominant switching mechanism reasonably.
Physics Background 12 · Resistive Switching
During SET, the active metal oxidizes into ions, moves through the medium under an electric field, and is reduced to progressively establish a metal bridge. RESET dissolves part of that bridge. A thin bridge can reduce switching energy but may be destabilized by heat and surface energy. Fast formation and long-term retention must be checked under the same conditions rather than taken from separate best-case experiments.
RESET is the reverse data-state update from low to high resistance; a later SET restores low resistance. This is reversible resistance switching rather than Flash-style block erase. The circuit defines logical zero/one encoding.
SET → RESET → SET
ECM changes a conductive path through the motion and redox reactions of active-metal ions. An existing conductive bridge commonly produces a low-resistance state; dissolving a critical part of the bridge produces a high-resistance state. Both ECM and oxygen-vacancy VCM exhibit resistive switching, but their ion sources and path materials differ. CBRAM is a common commercial name for this conductive-bridge memory. Sharing the ReRAM label does not justify combining their physical models.
Metal-Ion Migration and Reduction Create a Bridge; Reverse Operation Can Dissolve the Filament.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
The active Ag electrode releases Ag+ by oxidation; ions drift to the cathode for reduction and nucleation, then the growing bridge connects before bias removal.
Upper Ag is the oxidizable metal source and lower BE is inert. Initially no metallic bridge spans the medium.
The Ag anode releases Ag+ and electrons. Ag+ drifts toward the cathode, where electrons reduce ions and initiate nucleation.
Cathodic metal deposition extends toward the Ag electrode. This selected growth direction is not universal across ECM.
Current compliance limits bridge thickening. After connection and bias removal, the retained bridge supplies a low-R electronic path.
A representative cathode-nucleated embodiment is selected; other media and kinetics can change growth direction. RESET interrupts a critical path and does not remove all metal.
Growth direction and nucleation location depend on the material and kinetics. A single path in a schematic is not the morphology of every real device. Current compliance prevents an excessively thick bridge that would be difficult to RESET.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Apply reverse bias to a connected silver bridge, oxidize and dissolve its neck locally, then remove bias while retaining a high-R gap and residual metal.
A continuous silver bridge forms a low-R electronic path. Ionic transport and electronic conduction are distinct.
Reverse bias oxidizes Ag at the bridge neck into Ag+. Released cations move toward the active electrode, now cathodic.
A critical neck gap interrupts the metallic connection between electrodes while residual deposits can remain.
After bias removal, the high-R gap remains. A later SET can use residual nucleation sites.
A representative cathode-nucleated embodiment is selected; other media and kinetics can change growth direction. RESET interrupts a critical path and does not remove all metal.
RESET does not remove all the metal. Residual metal or nucleation sites affect the next SET, creating cycle-history effects and variation that are important in reliability analysis.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Compare low-bias currents for a retained metal bridge and a local gap; latch and remove read bias with the ionic structure approximately unchanged.
Left and right are the low- and high-resistance alternatives for one cell. Read bias has not yet been applied.
Compare currents at the same small read bias, chosen to avoid appreciable ionic redistribution.
After the sense circuit latches the difference, current stops. The connected low-R path and local high-R gap remain retained.
A representative cathode-nucleated embodiment is selected; other media and kinetics can change growth direction. RESET interrupts a critical path and does not remove all metal.
Accumulated bias and temperature can still drive metal ions. Validate whether prolonged reading and half-select operations gradually change the bridge, rather than checking only a single low-voltage read.
1T1R can limit SET current while selecting the cell. Crosspoint implementations require a sufficiently nonlinear selector or an appropriate array-bias scheme. Selector tolerance for forming and RESET pulses, half-select leakage, and reverse current must be matched to the memory device. An active-metal bridge can have very low resistance; without current-limited driving, local overcurrent can easily destroy reversibility.
Metal-ion nucleation sites, bridge width, and residual metal distributions create both cycle-to-cycle and device-to-device variation. Short pulses, current compliance, and write verification can narrow the resistance window but increase operation count and energy. When using these differences for PUFs or stochastic computing, distinguish reproducible fixed features from path noise that changes each time the bridge is rebuilt.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
Lower forming energy often implies a thinner bridge, which may break through thermal or surface-energy effects. An overly thick bridge increases RESET energy. Metal depletion, agglomeration, and residue progressively change the reversible switching window.
Low-resistance bridges, sneak paths, and half-select bias can expose non-target devices to additional current. Bidirectional selector operation, current compliance, and line resistance must be matched, and resistance-distribution tails must be tested after extensive cycling.
Active-metal diffusion barriers, medium uniformity, and interfacial nucleation control determine yield. CMOS contamination rules and subsequent thermal processing may impose stricter limits than the device-formation temperature. The full interconnect integration flow requires validation.
Write verification, error correction, and data-update strategy determine achievable lifetime. If supply-chain evidence covers only historical products, current part numbers, temperature conditions, and long-term supply must be checked again. Material potential does not establish system availability.
Suitable for embedded applications requiring low-energy updates of small data records when a specific product or process provides support, and for research into stochastic and analog functions of metal-ion paths. Practical selection must jointly consider post-cycling retention, read disturbance, temperature, and supply, rather than only one SET energy value or the initial resistance ratio.
A conceptual ECM design should not be adopted directly for long-term retention at very high temperature without post-cycling retention data, or where active-metal process-contamination constraints are unacceptable. Maturity and algorithms from oxide VCM also cannot simply be transferred to a metal bridge, because the ion sources and failure paths differ.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US5761115A
Trace the dendrite in Figure 1’s plan and cross-section, then compare the vertical geometry in Figure 4. Figure 5 adds an insulating condition that prevents direct contact.
Compare ECM metal oxidation, ion migration, reduction/deposition and reverse dissolution. The bridge is metallic, rather than an oxygen-vacancy filament.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
Why can an ECM metal bridge still influence the next write after RESET?
RESET often dissolves only part of the bridge, leaving metal and nucleation sites in the medium. These remnants alter the path and threshold of subsequent metal growth, producing cycle-history effects and variation.
Physics Background 13 · Phase Change
RESET locally melts material with a short, high-peak pulse and rapidly cools it into an amorphous state. SET uses an appropriate thermal history to crystallize the material. Reducing phase-change volume can lower energy, but retention, cycling failure, and thermal crosstalk must still be considered. PCM appears in production MCUs and is also researched for storage-class memory and analog weights. Different uses do not change its phase-based storage mechanism.
PCM RESET creates a high-resistance amorphous region through local melting and rapid quenching; SET promotes crystallization with a suitable thermal profile. Data is rewritten through pulse-induced thermal histories, not merely voltage reversal or mandatory Flash-style block erase.
SET → RESET → SET
PCM stores data in the fraction and geometry of crystalline and amorphous phase-change material. In a typical electronic device, the crystalline state has lower resistance and the amorphous state higher resistance; material kinetics retain the state after power removal. The actual state is more than an abstract resistance value. It includes the location and size of the phase-change region, degree of crystallization, and evolution over time, which jointly determine reading and lifetime.
Mushroom-Cell Example: A Short, Strong RESET Pulse Melts and Quenches; a Milder SET Pulse Promotes Crystallization.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Heat the amorphous cap into a crystallization-favorable region below Tm, dwell for nucleation and growth, then cool into a retained crystalline low-R state.
The RESET-created amorphous cap A sits above the heater and increases cell resistance.
SET raises the local temperature into a crystallization-favorable region below Tm and maintains sufficient dwell time.
Thermally activated nucleation and grain growth reduce amorphous volume, governed by material and the temperature-time history.
Cooling leaves continuous crystalline material C and lower resistance. This thermal history differs from melt-quench RESET.
Tx denotes a qualitative crystallization region and Tm the melting point; curves and geometry are not measurements. Crystallization depends on temperature, time, and material; ordinary reads require neither melting nor restore.
SET is controlled jointly by time and temperature. A stronger pulse is not necessarily better: remelting followed by rapid cooling may return the material to an amorphous state.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
A short strong pulse melts a local crystalline volume above Tm; a steep pulse fall rapidly quenches it into a retained amorphous cap and high resistance.
The initial local phase-change volume is crystalline, with conduction through material above the heater.
A short strong RESET pulse produces Joule heat, taking a local volume above Tm into liquid state L.
A steep pulse fall rapidly quenches the molten volume. Insufficient time for crystal growth produces an amorphous cap.
The amorphous cap A interrupts the low-R crystalline path and retains high resistance after cooling; no material is removed.
Tx denotes a qualitative crystallization region and Tm the melting point; curves and geometry are not measurements. Crystallization depends on temperature, time, and material; ordinary reads require neither melting nor restore.
RESET requires control of peak amplitude, duration, and cooling rate. It rewrites the phase-change device state; it does not imply that a whole region must first be erased before writing.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Compare crystalline and amorphous currents at a small read bias, then latch and remove bias while retaining phase; read temperature stays below the crystallization region.
C and A are alternative stored states of one cell. The local amorphous cap increases resistance.
At a small read bias, the crystalline state carries greater current. Read energy is chosen to avoid appreciable crystallization or melting.
After latching the current difference, remove bias and retain each phase. Resistance can still drift with time, requiring sense margin.
Tx denotes a qualitative crystallization region and Tm the melting point; curves and geometry are not measurements. Crystallization depends on temperature, time, and material; ordinary reads require neither melting nor restore.
Amorphous resistance drifts with time and temperature, and multilevel reading requires greater precision than binary reading. A cumulative read method in a specific patent may be destructive; ordinary PCM reads are not therefore all destructive.
In a transistor-based cell, the word line selects the heater paths that can conduct, and the access device must supply RESET current under worst-case conditions. A crosspoint version additionally requires a selector whose nonlinearity isolates half-selected cells. Selector turn-on, line resistance, and thermal crosstalk can all alter the actual thermal history of the phase-change region. An ideal current waveform cannot simply be assumed at every array location.
Material composition, heater-contact dimensions, amorphous-region shape, and local thermal resistance create write distributions. Structural relaxation also causes resistance to drift with time. Write verification, differential storage, drift compensation, and ECC can reduce errors. Projected PCM uses an additional branch to reduce the dependence of reading on the drifting material, but introduces new tradeoffs in current shunting and interface design.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
The material must crystallize readily for fast SET while keeping the amorphous state stable over time. Repeated melting and crystallization can cause segregation, atomic migration, and voids, progressively changing resistance and the energy required for switching.
RESET current, line resistance, and heat removal create position-dependent thermal histories, and neighboring devices can be heated. Selectors and access transistors must match the pulses while preserving half-selected data and the read-resistance window.
Heater-contact dimensions, phase-change composition, interfaces, and package thermal history affect the switched volume. Confined and projected structures have different deposition, etching, and material-compatibility costs. Ideal thermal simulations alone are insufficient.
Write verification, drift tracking, ECC, and update strategies increase control overhead. Storage-class-memory use requires persistence and power-failure consistency; analog-weight use must include data-conversion and calibration energy.
Suitable for code and data storage on an available ePCM platform and for multilevel, cumulative, or analog-computing research that can accommodate verification and calibration. Selection requires post-cycling retention and read conditions for the specific product or stack at its operating temperature, with thermal management and access current included in the system assessment.
PCM nonvolatility alone does not establish suitability for drift-free long-term precision analog weights, extremely low peak current, or unconditional long-term retention at high temperature. Production status of one ePCM MCU also cannot be converted into proof of volume production for arbitrary high-capacity SCM or crosspoint arrays.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US5912839A
First inspect Figure 1’s nonmonotonic resistance versus pulse-current relationship, then the material/electrode structure in Figure 2. The plot lacks a complete measurement contract for current product specifications.
Compare PCM temperature and phase-state sequences, then examine how this patent uses repeated stimulation for data encoding.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
Why might simply increasing the peak amplitude of a SET pulse fail to produce lower resistance?
A higher peak may melt the material. Rapid cooling afterward can then form an amorphous high-resistance region. The outcome depends on the complete time–temperature history, not current amplitude alone.
Physics Background 14 · Ferroelectricity
Writing sets polarization through the electric-field direction. Reading applies an excitation and identifies the original value from the charge difference between switching and non-switching responses. If reading changes polarization, the circuit must restore the original data. Commercial FeRAM can conceal this sequence behind its interface so that the user sees an ordinary read command, but internal restoration, power-failure conditions, and timing remain part of reliability.
An opposite write stimulus changes ferroelectric polarization to another readable state, which can be rewritten again. Research labels such as ERS or RESET refer to that polarization mechanism, not necessarily floating-gate charge removal or block erase.
A → B → A
The bit is represented by the polarization direction retained in a ferroelectric material after the external electric field is removed. Unlike DRAM, which relies on temporarily stored free charge, FeRAM is based on switchable remanent polarization. Reading uses the different charge responses when polarization switches and when it does not. Distinguishing the material's retention mechanism from the circuit's sensing method explains why nonvolatile memory may still require restoration after a read.
1T1C Topology: an Access Transistor Connects to the Capacitor Storage Node; Plate Pulses Work with Sensing and Restore.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Enable WL and raise PL above BL to switch domains with an upward field; remove capacitor bias and isolate the cell to retain P↑.
The starting remanent polarization represents opposite data, with no voltage across the capacitor.
Enable access and set PL above BL. A pulse meeting the effective switching condition drives domains upward.
Remove capacitor voltage and isolate the cell. Remanent polarization is retained; opposite data is directly rewritten without a block erase.
P↑/P↓ have no fixed 0/1 mapping. Field direction follows the relative PL/BL voltage; choose actual pulses for the ferroelectric material, stack, and effective switching condition. Opposite data is directly rewritten without block erase.
Positive and negative polarization can be assigned to logic 0/1 by the design. Validation must establish voltage distribution, switching charge, and write success probability at the worst-case temperature.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Enable WL and raise BL above PL so a downward field rewrites P↑ as P↓; remove bias and isolate the cell to retain the opposite data.
The starting remanent polarization represents opposite data, with no voltage across the capacitor.
Enable access and set BL above PL. A pulse meeting the effective switching condition drives domains downward.
Remove capacitor voltage and isolate the cell. Remanent polarization is retained; opposite data is directly rewritten without a block erase.
This reverse operation directly rewrites polarization, without block erase. P↑/P↓ have no fixed 0/1 mapping; choose actual pulses for the ferroelectric material, stack, and effective switching condition.
This is direct rewriting and usually does not require a prior block erase. If a product offers a chip-clear command, its controller behavior must be distinguished from the cell's polarization physics.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Distinguish switching and non-switching charge, latch the read value, then restore polarization as PL falls while WL stays enabled.
A and B represent two possible initial states of one 1T1C cell: P↑ or P↓. BL and PL are zero and WL is off.
Enable WL and raise PL to create an upward field. A contributes Qns without switching; B switches and adds Qsw, creating distinct BL signals.
The sense amplifier compares against an external reference and latches. In this convention A drives BL to zero and B to V; both presently have P↑.
Keep WL enabled while PL falls to zero. A sees no reverse field; B retains BL at V, creating a downward field that restores its original P↓.
After restore, turn WL off and precharge BL to zero. A and B retain their respective pre-read polarization; sensing plus restore completes the read.
A/B are alternative initial states of one cell, not two additional physical cells. PL rises during the illustrated read. Qsw is the polarization-switching component after dielectric/parasitic contributions, ideally about 2PrA.
Destructive reading means that the sensing operation may change the internal storage state. It does not mean that every external read loses the data. The complete read sequence must include sense latching and restoration timing.
The word line enables the access transistor, connecting the ferroelectric capacitor to the bit line. The plate line supplies excitation, and the sense amplifier compares charge with a reference cell. A 1T1C design depends on reference precision; 2T2C allows differential sensing at greater area cost. The array must also control voltage distribution across unselected capacitors to prevent unnecessary switching or fatigue from shared plate lines.
Capacitor area, grains, ferroelectric-phase fraction, and coercive-field distribution change switching charge. Fatigue reduces switchable polarization, imprint makes switching asymmetric between directions, and reference drift reduces sensing margin. Differential circuits, reference design, and timing control can reduce the impact, but a typical polarization curve cannot replace full-array validation of writing, reading, and restoration.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
Shrinking the capacitor reduces switchable charge. Increasing the field to strengthen the signal may increase fatigue and dielectric stress. Retention, imprint, and polarization back-switching also depend on material, temperature, and cycling history.
Bit-line parasitic capacitance dilutes the sensing signal. Reference cells and plate-line distributions affect the worst-case read margin. Differential structures improve discrimination at an area cost, and restoration also consumes array timing.
Ferroelectric-film phase, electrodes, and annealing conditions must be integrated with CMOS. Traditional oxides and HfO₂ systems have different material windows. The ability to deposit a ferroelectric material does not establish array-process maturity.
Read restoration, supply droop, and interface completion conditions determine data consistency. Datasheet cycle counts, retention temperatures, and interface clocks describe different levels and must be matched to the actual update workload and environment.
Suitable for frequent data logging, metering, industrial control, and equipment-state retention when existing capacities and interfaces meet the requirements. Selection can use commercial datasheet tables for temperature, retention, and cycling, while checking minimum supply voltage, write-completion conditions, and the power-failure sequence to turn nonvolatility into verifiable system reliability.
If the primary goal is extremely high density, the lowest bit cost, or a read operation that never requires internal restoration, FeRAM's high cycle capability should not obscure its structural cost. Specifications of commercial capacitor-based products also cannot guarantee the performance of FeFET or FTJ technologies still under research.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US4873664A
Trace the 1T1C cell along the bit line into the sense/restore circuit, then compare word-line and plate-line timing. After charge sensing, the latched result restores the original polarization.
Compare both initial polarizations, switching-charge contrast, latching and write-back in the FeRAM read plates. Restoration is an explicit stage.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
If FeRAM retains data after power loss, why might reading still require a writeback?
Nonvolatility comes from remanent polarization, but some sensing sequences deliberately switch polarization to obtain a charge difference. The original direction must be restored after sensing to preserve the data held before the read.
Physics Background 15 · Ferroelectricity
FeFET combines ferroelectric retention with transistor current gain, providing research opportunities for nondestructive reading and density scaling. The challenge is that write voltage is divided across both the ferroelectric and interfacial layers, while polarization switching may also generate or fill traps. Polarization stability, memory window, and endurance are not three independently optimizable numbers.
An opposite write stimulus changes ferroelectric polarization to another readable state, which can be rewritten again. Research labels such as ERS or RESET refer to that polarization mechanism, not necessarily floating-gate charge removal or block erase.
A → B → A
FeFET uses ferroelectric polarization in the gate stack to change channel electrostatics, giving the transistor distinguishable high and low threshold voltages. Reading selects a gate bias between those thresholds and senses channel current. Charge trapping and detrapping also affect the actual memory window, so not every threshold-voltage change can be attributed solely to polarization.
Polarization Shifts the Channel Threshold; Interface Traps Also Affect the Read Window.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Polarization changes channel-side bound charge and threshold voltage; sense channel current between the two Vt distributions.
Initial P points away from the channel, corresponding to higher n-channel Vt.
The selected gate pulse drives P toward the channel, creating positive channel-side bound charge and lowering Vt.
After the pulse, remanent polarization retains the Vt shift. Opposite data is directly rewritten; not every trap-related Vt shift is assigned to polarization.
Here P toward the channel creates positive bound charge and lowers n-channel Vt; verify pulse polarity for the actual stack. Trapped charge also shifts Vt, so hysteresis is not attributed solely to polarization.
This schematic uses an n-channel device and one stack orientation. The actual polarity corresponding to high and low threshold voltage must be checked against the structure. The pulse may also change trapped charge, so the two effects must be separated.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Polarization changes channel-side bound charge and threshold voltage; sense channel current between the two Vt distributions.
Initial P points toward the channel, corresponding to lower n-channel Vt.
The selected gate pulse drives P away from the channel, creating negative channel-side bound charge and raising Vt.
After the pulse, remanent polarization retains the Vt shift. Opposite data is directly rewritten; not every trap-related Vt shift is assigned to polarization.
Here P toward the channel creates positive bound charge and lowers n-channel Vt; verify pulse polarity for the actual stack. Trapped charge also shifts Vt, so hysteresis is not attributed solely to polarization.
Research often calls the two directions program and erase, but the mechanism is polarization rewriting. It differs from Flash erase through charge removal or neutralization; the named stack defines its threshold direction and operating conditions.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Polarization changes channel-side bound charge and threshold voltage; sense channel current between the two Vt distributions.
The illustrated low-Vt state has polarization toward silicon and positive bound charge favoring an n-channel. The alternative high-Vt curve is also shown below.
Use a small drain bias and Vg,r between Vt,L and Vt,H. Low Vt gives larger Id and high Vt smaller Id, flowing through the source-drain channel.
Remove read bias after latching the decision. Polarization and the Vt window remain; actual read bias must control disturbance.
Here P toward the channel creates positive bound charge and lowers n-channel Vt; verify pulse polarity for the actual stack. Trapped charge also shifts Vt, so hysteresis is not attributed solely to polarization.
Read bias must avoid polarization-switching conditions and substantial trap charging or discharging. If the memory window shrinks or distributions overlap, reliable reading can fail even while a typical device still exhibits hysteresis.
The array uses word lines for gate operation and bit/source lines to sense the channel. Practical schemes may add access transistors or special biases to suppress write and read interference in unselected cells. Write voltage divides across the ferroelectric and interfacial layers. The terminal potentials of half-selected cells must be drawn explicitly; being a transistor does not automatically provide complete isolation.
A small device contains only a few ferroelectric grains and may switch in discrete steps. Grain orientation, coercive field, interface thickness, and trap distributions also create device-to-device differences. Analog weights require multilevel windows and update control. PUF research can exploit cycle variation, but reproducibility, bias, environment, and reconfiguration conditions need separate testing. Variation is not itself security.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
Polarization back-switching, depolarization fields, trap generation, and oxide lifetime jointly limit retention and endurance. Introducing trapped charge to stabilize polarization may sacrifice the initial window. Readable distributions must be compared after both cycling and retention.
Write bias may disturb half-selected gates, while threshold-voltage tails and bit-line leakage limit read margin. A small cell does not imply small high-voltage drivers, references, or redundancy overhead. Multilevel data requires particularly fine control.
Ferroelectric phase, grains, doping, annealing, and interfacial-layer thickness require stable control. The floating metal in MFMIS and the direct interface in MFIS introduce different integration problems. A material patent cannot replace complete process qualification.
Analog computing must compensate for window drift, update asymmetry, and read noise. Digital storage requires ECC and update strategies. PUFs must validate entropy, reliability, and the attack model rather than merely provide an attractive threshold-voltage histogram.
Suitable for research into dense embedded storage, transistor-based analog weights, and PUFs with explicit environmental validation. For a target process, first establish a coupled model of ferroelectric behavior and trapping, then validate the worst-case window through array operation so that small-device potential translates into usable sensing margin.
FeFET research data is insufficient to promise a mature standard chip in the near term, calibration-free multilevel precision across all temperatures, or direct reuse of the high-cycle specifications of commercial FeRAM. Hysteresis alone does not establish nonvolatility, and short room-temperature retention cannot replace post-cycling lifetime at high temperature.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US10153155B2
Figures 1 and 2 compare three- and four-layer film arrangements. Identify the material and outer conducting layers, then read the distinct dopant layers and heating requirements in the process claim.
Use this source for FeFET material/process constraints. The next transistor-stack patent and operation plates cover the electrical device.
Compare All Drawings, Numerals and Claims →US11502083B2
Read layers 31, 32, 33b, 34 and 35 upward from the substrate, then trace the junctions back to the channel. This stack includes a floating gate and is not the simplest metal/ferroelectric/silicon structure.
Compare polarization-controlled threshold shift in FeFETs. Distinguish material layers from electrical terminals to locate the applied field.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
Why cannot all FeFET threshold-voltage drift be treated as weakening polarization?
Traps in the interface and dielectric capture and release charge, also changing threshold voltage and, in turn, polarization stability. Time, bias, and material evidence are needed to separate the two effects.
Physics Background 16 · Ferroelectricity
FTJ reverses polarization with a larger pulse, then senses tunneling current at a smaller bias, pursuing two-terminal nondestructive storage and interconnect-layer integration. The barrier must be thin enough to provide readable current while retaining stable ferroelectricity and suppressing leakage. An attractive resistance ratio does not establish sufficient absolute read current, much less a reliable selection window for a large array.
An opposite write stimulus changes ferroelectric polarization to another readable state, which can be rewritten again. Research labels such as ERS or RESET refer to that polarization mechanism, not necessarily floating-gate charge removal or block erase.
A → B → A
FTJ controls tunneling current through the polarization direction of a thin ferroelectric barrier. The two directions produce different effective barrier profiles and resistance levels, commonly described through tunnel electroresistance, or TER. FTJ and FeFET both use polarization, but FTJ does not rely on threshold-voltage amplification in a semiconductor channel. Read current, barrier thickness, electrode screening, and leakage therefore become central tradeoffs.
Polarization Modulates Tunneling Current; the Barrier Is Exaggerated to Show the Material Sequence.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Compare interfacial accumulation/depletion, effective barrier, and low-bias current for opposite polarizations.
Initial P points toward Cr/Au, giving NSTO depletion and high R.
Write voltage is defined at Cr/Au relative to NSTO, driving P toward NSTO. Interfacial bound charge and electronic screening rearrange.
After write bias removal, remanent P retains interfacial accumulation and a smaller effective barrier. This polarization/resistance mapping is limited to the named research stack.
The x axis runs from Cr/Au toward NSTO. P toward NSTO corresponds to LRS in this named stack; reverse P strengthens depletion and the HRS barrier. This mapping is not universal to FTJs; HRS transport can include thermal assistance.
The polarization direction corresponding to low resistance depends on electrode and interface asymmetry. A teaching diagram must not define a fixed relationship between up/down arrows and high/low resistance for the entire family.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Compare interfacial accumulation/depletion, effective barrier, and low-bias current for opposite polarizations.
Initial P points toward NSTO, giving electron accumulation and low R.
Write voltage is defined at Cr/Au relative to NSTO, driving P toward Cr/Au. Interfacial bound charge and electronic screening rearrange.
After write bias removal, remanent P retains NSTO depletion and a higher, wider effective barrier. This polarization/resistance mapping is limited to the named research stack.
The x axis runs from Cr/Au toward NSTO. P toward NSTO corresponds to LRS in this named stack; reverse P strengthens depletion and the HRS barrier. This mapping is not universal to FTJs; HRS transport can include thermal assistance.
This is polarization rewriting, sometimes called SET/RESET or program/erase. Those labels do not imply Flash-style block erase; the electrode structure defines the relation between polarization and resistance.
Follow state, stimulus and result across the frames. Each sequence uses one specified structure and operating convention.
Specified Structure and Operating Convention
Compare interfacial accumulation/depletion, effective barrier, and low-bias current for opposite polarizations.
The coordinate runs Cr/Au→BSO→NSTO. The illustrated low-R initial state has P toward NSTO; U(x) is a qualitative barrier, not a measured band profile.
Positive interfacial bound charge for P toward NSTO attracts electron accumulation. At the same small read bias, the smaller barrier permits greater electronic current.
This frame compares the alternative high-R initial state; reading does not turn low R into high R. Reverse P depletes NSTO, adding a barrier and reducing current.
Latch after removing small read bias. The diagram returns to the low-R branch with polarization, accumulation, and barrier retained; the high-R branch likewise retains its state.
The x axis runs from Cr/Au toward NSTO. P toward NSTO corresponds to LRS in this named stack; reverse P strengthens depletion and the HRS barrier. This mapping is not universal to FTJs; HRS transport can include thermal assistance.
TER and absolute current must both be observed. A high resistance ratio with extremely small currents may still be limited by sensing noise, leakage, and read time.
A single FTJ has only two terminals and does not inherently provide complete array isolation. A 1T1FTJ implementation uses a transistor for selection and biasing; a crosspoint array needs a selector or sufficiently nonlinear current characteristics. Half-select voltages may cumulatively affect polarization, and read leakage may obscure small tunneling currents. Device TER and array readability therefore require separate validation.
Small film-thickness differences can change tunneling current exponentially. Grains, ferroelectric phase, and interfacial traps also affect polarization switching and the barrier. Multilevel research must distinguish a controllable polarization fraction from random leakage paths. Differential cells, pulse verification, and reference calibration may improve behavior, but their additional area and time must also be quantified.
A switching device must still satisfy four layers of constraints before it can serve a reliable system.
A thinner barrier improves read current but may reduce polarization stability or increase leakage. A thicker barrier reduces the available sensing signal. Electrode screening, depolarization fields, and traps jointly affect retention and TER after cycling.
Small tunneling currents are easily obscured by sneak currents and sensing noise. Half-select pulses may also change polarization cumulatively. The selector turn-on window, line resistance, and cell-current distributions determine the usable array size.
Ultrathin-layer uniformity, ferroelectric crystallization, and electrode/interface reactions are highly sensitive. Lower annealing temperature can benefit interconnect integration, but phase, polarization, and leakage must all meet requirements; deposition alone is insufficient.
With very small read current, sensing time, amplifiers, and calibration energy may offset device-level write savings. Multilevel weights also require management of drift and update distributions. TER or cell cycle count alone cannot compare complete memories.
Suitable for research into two-terminal polarization storage, interconnect-layer integration, and analog weights that tolerate calibration, particularly when electrodes, interfaces, and selectors can be designed together. Research targets should jointly specify absolute read current, TER, pulses, area, retention, and post-cycling distributions to assess suitability for larger arrays.
Available research abstracts do not guarantee standard chips deliverable in the near term, calibration-free high-capacity multilevel storage, or extremely low sensing energy without a long integration time. A low-temperature step in a patent also does not establish validation of the complete back-end process and its reliability.
These representative patents are research entry points. Bibliographic records establish priority and assignment; embodiments and claim scope require separate reading.
US20240057343A1
Read the five-layer stack in Figure 3, then compare polarization-dependent barriers in Figure 5. Figures 15–18 connect memory stacks to transistors.
Compare FTJ polarization reversal and barrier changes, keeping this stack distinct from research devices using other electrodes and ferroelectrics.
Compare All Drawings, Numerals and Claims →Illustrations synthesize the cited physical principles. Performance, production status, and patent claims remain bound to their original documents. Undisclosed biases, material recipes, and process conditions remain implementation gaps.
Why can an FTJ still read slowly despite a large high-to-low resistance ratio?
A large ratio does not mean a large absolute current. If both state currents are very small, the sense circuit needs a longer integration time and becomes more sensitive to leakage and noise. TER and current magnitude must be evaluated together.
Integration Study
Switching a device twice establishes only that it has usable storage states. A practical memory must also select its target from a large population of cells, avoid disturbing its neighbors, and read data correctly across temperature, aging, and process variation. Selectors, wires, and peripheral circuits therefore determine how much of the cell-level advantage survives.
For a resistive cell, data may be represented by high- and low-resistance states, but an external address does not directly select a laboratory probe. Row and column decoders first identify the target. Word lines, bit lines, and access devices then establish the current path. The read circuit converts a finite current or voltage difference into digital data, while the controller schedules reads, writes, and error handling.
This path gives cell switching time and complete access time different meanings. The former may record only the pulse needed to change the material state; the latter also includes decoding, wire charging and discharging, sensing, verification, and data transfer. The first step in interpreting any performance number is therefore to identify where the measurement starts and what defines completion.
Back to the Section Index ↑A 1T1R cell consists of an access transistor and a resistive storage device. The transistor gate provides an independent control terminal that can isolate unselected cells. With an appropriate circuit design, the transistor can also limit forming or programming current to prevent excessive resistive switching. A read selects the target path, applies a lower bias, and uses the sensing circuit to determine the state.
The trade-off is transistor area. The required write current and voltage affect transistor sizing and voltage tolerance. Shrinking the storage material does not mean the complete cell can shrink by the same ratio. This explains why papers may report a very small resistive device while a practical 1T1R array occupies substantially more area.
Back to the Section Index ↑A 1S1R cell places a two-terminal selector in series with a storage device. Strong nonlinearity, rectification, or threshold switching can suppress current at low bias while delivering enough current at the selected read or write bias. The two-terminal structure fits at the intersection of crossing wires and creates opportunities for vertical stacking. A volatile selector does not prevent its series-connected storage device from retaining data: the two components perform different functions.
Without sufficient isolation, current can flow through other low-resistance cells along parasitic paths, overwhelming the target read signal or altering programming conditions. A selector also consumes voltage headroom and introduces leakage and variability. IBM's array research shows why the selector and storage device must be assessed together with the wires, operating biases, and array size. A high nonlinearity ratio does not establish operation at arbitrary scale. Devices with intrinsic rectification or selection require separate evidence for their array conditions.
Back to the Section Index ↑Consider an idealized V/2 scheme: drive the selected word line to V, the selected bit line to 0, and all other word and bit lines to V/2. The target intersection sees V. Cells sharing one selected line with the target see approximately V/2. The remaining cells see no voltage difference if the wires are ideal. Cells selected at only one terminal are called half-selected cells.
Even if a half-select voltage cannot switch a cell during one pulse, repeated exposure can cause cumulative disturbance. Real arrays also have wire voltage drops and voltage division across selectors, so the voltage delivered to a distant cell may differ from the driver's setting. V/2 is therefore a biasing strategy that must be validated, not a universal solution across materials and switching polarities.
Back to the Section Index ↑The sense circuit compares a cell signal with a reference. Even when average high- and low-resistance states are widely separated, a minority of cells can approach the decision boundary because of process variation, temperature, read noise, or programming history. In a large array, these distribution tails can determine product yield more strongly than the typical curve of a well-behaved device.
Reducing read bias generally helps reduce disturbance, but it can also reduce the signal, require more settling time, or demand a more precise sense circuit. Practical macros use reference tracking, offset compensation, and measures to address wire voltage drop. The 40 nm RRAM macro studied by TSMC and collaborators treats read-channel mismatch, leakage, ADC offset, IR drop, and cell variability as coupled problems. The device resistance ratio cannot substitute for this circuit-level evidence.
Back to the Section Index ↑Program verify applies a programming pulse, reads the cell back, and checks whether it lies within the target range. If it does not, the controller adjusts or adds pulses until the target is reached or the retry limit is exhausted. This feedback helps manage device-to-device and cycle-to-cycle variation. It is particularly important for multistate storage, where each state has a narrower usable margin.
Verification has a cost and its own limitations. Every readback consumes time and energy, and the verification measurement is itself affected by noise. NIST's oxide-RRAM experiments show that read fluctuations can produce false-pass decisions and leave longer tails in the state distribution. IBM's analog-PCM work also shows that read noise and drift limit closed-loop programming precision. To assess write quality, report pulse count, verification thresholds, retries, failure fraction, and the delay after programming together.
Back to the Section Index ↑Error-correcting codes (ECC) add check information to detect or correct errors within a defined capability. Spare rows, spare columns, and defect replacement can avoid some permanent failures. These methods translate raw-cell error rates into data reliability acceptable to the system, but consume capacity, area, latency, or computation.
Successful data retrieval does not mean that the raw bits contained no errors. Reports should distinguish error rates before and after ECC. ECC also has finite capability against overlapping state distributions, correlated errors, and failed bits. Effective density should divide actual usable data capacity by the area that includes the associated periphery, rather than simply multiplying layer count by bits per cell.
Back to the Section Index ↑The signal account asks how much read and write margin remains at the worst location and under the worst data pattern. The energy account includes not only the selected cell but also wire charging and discharging, leakage through unselected paths, and repeated verification. The area account includes decoders, drivers, sense circuits, ECC, and redundancy.
A larger array is therefore not necessarily more efficient. Partitioning it into smaller subarrays can add some peripheral overhead while improving voltage drop, speed, and reliability. A research result should identify the bottleneck under the stated material, array, and operating conditions, rather than declare that an entire technology family has reached an immutable physical endpoint. A persuasive improvement reports performance, reliability, and cost under the same conditions.
Back to the Section Index ↑A paper demonstrates that an RRAM cell switches with a 10 ns pulse. Does that establish a 10 ns write time for a 1 Gb memory? What additional categories of evidence are needed?
No. Establish whether the material and implementation are the same, whether selection uses 1T1R or 1S1R, the actual bias at the array's worst location, sense margin, pulse and verification counts, retry and failure thresholds, ECC and peripheral latency, and the product's definition of write completion. The cell pulse describes only one part of the complete path.
Integration Study
Storage-class memory (SCM) addresses the gap in requirements between DRAM and NAND storage. It is not another bitcell type, and adopting CXL does not automatically establish an SCM implementation. Understanding SCM requires distinguishing storage physics, attachment, access granularity, and which data can actually be recovered after failure.
DRAM provides low-latency working memory, while NAND supports large-scale storage through high capacity and low cost per bit. SCM is often used to discuss the space between them in performance, capacity, cost, and endurance. Different publications may include low-latency storage products or NVM with a memory interface, so any discussion of SCM should first define its scope.
This topic treats SCM as a system role rather than a material name. Technologies such as PCM and ReRAM have been explored as candidates. Persistent-memory discussions focus more specifically on properties such as nonvolatility, byte addressability, and low latency, together with how software guarantees recoverability. Not every NVM is suitable as main memory, and not every low-latency SSD provides the same load/store semantics.
Back to the Section Index ↑CXL is an interconnect protocol built on a PCIe physical link that can let a processor access memory on a device. The specification distinguishes volatile and persistent memory ranges. They can share an attachment architecture while retaining different data-persistence properties. Samsung CMM-D, for example, is a DRAM memory module using CXL.
Capacity expansion, memory sharing, and pooling therefore do not establish retention through power loss. Assessing a CXL device requires identifying its media, exposed address ranges, platform-supported capabilities, and handling of host, link, or device failures. The interface name alone answers none of these questions.
Back to the Section Index ↑One approach uses media that inherently retain data. Another lets fast but volatile DRAM handle normal reads and writes, then combines it with NAND and backup energy to save data during power loss and restore it after restart. NVDIMM-N is an example of the latter. A member technical article published by the CXL Consortium in 2026 also discusses a similar CXL architecture.
The guarantee depends on sufficient backup energy, successful completion of the save sequence, firmware identification of valid data, and platform coordination. The architecture does not make a DRAM bitcell nonvolatile; it provides data persistence through system design. An architectural diagram must also keep conceptual feasibility, product qualification, and volume-production status separate.
Back to the Section Index ↑3D XPoint and Optane are important commercial implementations in the history of SCM, but their lifecycle has changed. On 2021-03-16, Micron announced the immediate discontinuation of 3D XPoint development and stated that manufacturing would end after existing commitments were fulfilled. It completed the sale of the Lehi fab on 2021-10-22. Intel announced in July 2022 that it would discontinue further Optane product development.
Intel's customer letter dated 2023-03-21 stated that, at the time, media inventory was expected to support customer demand through 2025, with support resources planned through 2030. Inventory, warranties, and support cannot be described as continued investment in next-generation manufacturing, nor does that letter establish that every part number remains available in 2026. The history also shows why the exit of one commercial roadmap does not eliminate an entire storage-physics family such as PCM: ST SR6P6C8 provides a separate embedded-PCM implementation in volume production.
Back to the Section Index ↑After a program executes a store instruction, the new data may still reside in a CPU cache, memory controller, or device buffer. If any of those locations lack power-loss protection, the system can lose power before the nonvolatile media ever receive the data. The persistence domain is the boundary within which the platform guarantees that data are retained, or their retention is completed, under specified failure conditions.
Software must use platform-appropriate synchronization, cache flushing, and ordering mechanisms to ensure that data reach this domain. PMDK documentation explains the process in terms of cache flushing and hardware-buffer draining, and notes that some platforms can omit certain steps. The objective is not to memorize a universal instruction sequence, but to identify what the platform protects and which completion event establishes persistence.
Back to the Section Index ↑Suppose a program creates a new data record and then updates an index pointer to refer to it. If the pointer becomes persistent before the content does, recovery may find an incomplete record. Requiring both writes to become persistent eventually does not remove the need to control their order. Conversely, persisting the content before updating the index can leave unreferenced data, which requires a reclamation or recovery strategy.
This is a failure-consistency problem. A persistence operation guarantees that data reach the persistence domain; it does not automatically combine multiple fields into an indivisible transaction. Logging, commit markers, version information, and transaction mechanisms establish recognizable recovery points. The platform must guarantee the relevant atomic-write granularity; arbitrary-size writes cannot be assumed to survive as complete units. SNIA's white paper on atomics and transactions illustrates these distinctions through data structures.
Back to the Section Index ↑Byte or cache-line access can reduce some software and data-movement overheads. Block SSDs use controllers and parallelism to increase throughput. Each approach has suitable applications; one number in ns or µs cannot determine the role. Granularity, queue depth, read/write ratio, access locality, and synchronization frequency all affect user-visible performance.
For example, Intel's 2021 P5800X tests report average 512 B random-read latency of 3.5 µs and average 4 KB random-read latency below 6 µs. These are results for a specific complete SSD and host platform, not material-level pulse durations. Average latency also cannot substitute for tail latency. For databases and logging, the relevant measure is transaction latency after data become persistent, rather than merely the speed of handing data to a buffer.
Back to the Section Index ↑A technology positioned between DRAM and NAND does not necessarily sustain a commercial position between them. Media density, yield, controller and packaging costs, supply scale, software modification costs, and alternatives within existing systems all influence adoption. Micron's 2021 announcement explicitly cited insufficient market validation to support scaled investment as part of its decision. That is a commercial judgment, not proof that a single physical-performance metric failed.
Evaluate an SCM candidate through three questions: Which specific workload bottleneck does it address? What is the complete system cost under the same capacity, reliability, and failure-protection requirements? If the media or supplier change, is there still a viable path for the software and hardware? This connects bitcell principles to practical product decisions.
Back to the Section Index ↑A CXL card uses DRAM plus NAND and advertises memory expansion. Can it be called persistent memory on that basis alone? Can power be removed immediately after a program completes a store?
The interface or combination of media is insufficient. Verify whether the device exposes a persistent range, its backup energy and save/restore mechanism, protection for buffers from host to device, synchronization and ordering semantics, and the software recovery method. Store completion does not necessarily mean the data have reached the persistence domain. Even after persistence, updates to multiple fields still require failure atomicity and consistency.
Representative Patent Studies
Begin with the problem each patent addresses, then read its structure, operations, and claims. This is a seed patent research index; it does not constitute a complete family legal-status review or freedom-to-operate analysis.
Drawing, Elements and Claim Reading
Compare the terminal and neck widths, then trace current crowding and the local thermal gradient. Geometry and material distribution are the design variables.
Claim 1 combines terminal narrowing, different silicide and polysilicon footprints, and unsilicided boundaries. Figure 4A explains the narrowing; the material and boundary requirements remain part of the claim.
Open the Original ClaimsCompare the eFuse write sequence: current path → material redistribution → high resistance. Different fuse stacks need different physical failure models.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Follow the top-view metal link into the two via cross-sections. A smaller contact area concentrates current and local heating.
Claim 1 requires a via end that lands only partly on the fuse link and specifies the connection between two metal layers. Dimensions, resistance and other conditions appear in dependent claims.
Open the Original ClaimsCompare localized heating and separation in a metal-via fuse. Its material explanation differs from the silicide electromigration example.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Locate the storage element and select transistor in Figure 3, then compare selected and unselected biases in Figure 8. The listed voltages belong to this embodiment.
Claim 1 combines a MOS select transistor, a thin-dielectric storage element, and row-select, column-select and row-program lines. The breakdown cross-section alone does not capture the array connections.
Open the Original ClaimsCompare antifuse operation: intact dielectric → selected high field → permanent conduction path, followed by low-stress sensing.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Identify the floating gate, localized thin region and upper control gate in the late process cross-sections. Thin-region placement and the second dielectric determine coupling and tunneling paths.
Claim 1 is a fabrication method: active regions, a localized thin dielectric, floating gate, isolating second dielectric and second gate covering the channel. Its process requirements are more specific than a generic EEPROM sketch.
Open the Original ClaimsCompare EEPROM FN injection and removal: both directions must pass through the actual thin dielectric region.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Find the overlap between the buried control region and floating gate, then use the equivalent circuit to distinguish coupling, storage and channel conduction.
Claim 1 includes a buried control gate, coupled floating gate and a thin tunnel region spanning part of the channel and a junction, with inhibition of an unselected overerased cell. Claim 4 separately specifies a split-gate structure.
Open the Original ClaimsCompare the MTP IP study’s single-poly teaching variant: this patent’s CHE injection and FN removal use a buried control node, with no second control-poly layer above the floating gate. Current product mechanisms require their own documentation.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Follow floating-gate formation in Figure 5 into the source/drain structure in Figure 6. Locate the select gate and tunnel oxide; the nested wells support separately controlled erase bias.
Claim 1 specifies nested wells, select and floating gates, and a source acting as the control-coupling node. Claims 4–6 add particular erase and programming biases; these values do not define all split-gate NOR.
Open the Original ClaimsCompare the third NOR variant: source-side injection and well/channel-side FN erase, separately from implementations that tunnel toward a select gate.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Read upward from silicon: thin memory oxide, silicon nitride, interfacial oxide and polysilicon gate. The drawing explains the dielectric stack.
Claim 1 specifies a CVD second oxide with a thickness range and an upper limit for the first oxide. Claim 6 covers fabrication. This early SONOS stack does not specify every later engineered tunneling stack.
Open the Original ClaimsFor SONOS operation, distinguish the lower tunnel oxide from the upper blocking oxide. Stored charge resides in the nitride.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Compare the prior-art A panels with embodiment B panels. Trace the localized charge region and READ arrow; reading direction changes which end of the channel barrier controls current.
Claim 1 combines localized electron trapping near the programming drain with different thresholds for reverse and forward read. The directional behavior is essential to the reading, beyond the ONO stack alone.
Open the Original ClaimsCompare the localized NROM sequence: CHE stores electrons near one end, reverse read senses from the opposite direction, and BBHH erase is explained with its separate source.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
In Figure 2, trace the silicon pillar from the common source to the bit line. Figure 6 unfolds the same structure into a string circuit, with select gates at both ends.
Claim 1 specifies gate stacks, sidewall dielectric containing an insulating storage layer, semiconductor pillars, data lines and upper/lower select gates. It describes a particular vertical NAND structure, not every modern cylindrical-hole array.
Open the Original ClaimsCompare selected-word-line injection, pass biases and program inhibit. This patent’s source-side electron removal is shown separately from later GIDL hole-assisted erase.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Start with the two pulse waveforms in Figure 4, then follow the moments through Figures 5 and 6. The overlap interval and turn-off order are part of the operation.
Claim 1 combines at least two antiferromagnetically coupled free layers, a moment-balance condition and pulse order t₀<t₁<t₂<t₃<t₄. Arbitrary orthogonal field pulses are not equivalent.
Open the Original ClaimsCompare the five-frame Toggle sequence and initial-state check: apply a toggle when the stored bit needs to change.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Figure 1 is a five-layer metallic-conductor model. Follow A→F1→B→F2→C to locate fixed and variable moments, then relate current to torque on F2.
Claim 1 specifies fixed and changeable magnetic conducting layers, a nonmagnetic conductor between them and a perpendicular current source. This early embodiment has a metallic spacer, rather than a modern MgO tunnel barrier.
Open the Original ClaimsUse this patent for spin-transfer physics, then the modern STT-MRAM plates for P/AP resistance sensing and the tunnel barrier.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Trace the first horizontal wire through the magnetic stack to wiring in the other direction. Access transistors and crossing interconnect explain array integration beyond one MTJ.
Claim 1 is a fabrication method covering isolated first wires, common device layers formed and separated above them, and second wires in another direction. The drawing explains connectivity; the claim centers on fabrication steps.
Open the Original ClaimsCompare three-terminal SOT read/write separation. Preserve the control terminals rather than reducing them to the STT current path.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Follow the Figure 5 cycle through mobile-species, barrier and conductive-region changes. The intermediate state has a physical location; it is not a single-step SET/RESET diagram.
Claim 1 requires three intermediate layers and two pulses: accumulation in the first layer, then motion into the third to complete the change. This specific multilayer memristor scheme does not define all bipolar VCM.
Open the Original ClaimsCompare species redistribution in VCM operation; this patent additionally shows how pulse sequencing introduces an intermediate state.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Trace the dendrite in Figure 1’s plan and cross-section, then compare the vertical geometry in Figure 4. Figure 5 adds an insulating condition that prevents direct contact.
Claim 1 covers an ion-containing conductor, electrodes and dendrite growth from negative toward positive. Claim 2 adds opposite-polarity reversal. Claim 3’s blocking condition is an additional limitation, not universal to all embodiments.
Open the Original ClaimsCompare ECM metal oxidation, ion migration, reduction/deposition and reverse dissolution. The bridge is metallic, rather than an oxygen-vacancy filament.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
First inspect Figure 1’s nonmonotonic resistance versus pulse-current relationship, then the material/electrode structure in Figure 2. The plot lacks a complete measurement contract for current product specifications.
Claim 1 focuses on a program pulse insufficient for a single SET but effective cumulatively with later pulses. Claim 2 adds RESET; claim 3 adds a read method that counts extra pulses. This is more specific than generic PCM heating.
Open the Original ClaimsCompare PCM temperature and phase-state sequences, then examine how this patent uses repeated stimulation for data encoding.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Trace the 1T1C cell along the bit line into the sense/restore circuit, then compare word-line and plate-line timing. After charge sensing, the latched result restores the original polarization.
Claim 1 specifies cell connections to word, bit and separate plate lines, with one capacitor electrode connected to the bit line through a switch. Claim 2 adds a sense amplifier and dummy ferroelectric reference cell.
Open the Original ClaimsCompare both initial polarizations, switching-charge contrast, latching and write-back in the FeRAM read plates. Restoration is an explicit stage.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Figures 1 and 2 compare three- and four-layer film arrangements. Identify the material and outer conducting layers, then read the distinct dopant layers and heating requirements in the process claim.
Claim 1 is a film-formation method involving three hafnium/oxygen layers, two different dopant layers, heating and conductive layers on both sides. The figure is not a complete FeFET bitcell and does not define its read channel.
Open the Original ClaimsUse this source for FeFET material/process constraints. The next transistor-stack patent and operation plates cover the electrical device.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Read layers 31, 32, 33b, 34 and 35 upward from the substrate, then trace the junctions back to the channel. This stack includes a floating gate and is not the simplest metal/ferroelectric/silicon structure.
Claim 1 details the relative positions of the composite gate, isolation, sidewalls, source/drain and silicide. The drawing teaches one implementation, rather than a universal FeFET stack.
Open the Original ClaimsCompare polarization-controlled threshold shift in FeFETs. Distinguish material layers from electrical terminals to locate the applied field.
Return to This Technology’s Operation SequencesDrawing, Elements and Claim Reading
Read the five-layer stack in Figure 3, then compare polarization-dependent barriers in Figure 5. Figures 15–18 connect memory stacks to transistors.
This is a published application. Claim 1 combines a first electrode, ferroelectric material and contacting catalytic metal; claims 2–3 add particular electronegativity and thickness limits. The five-layer embodiment includes details beyond the independent claim.
Open the Original ClaimsCompare FTJ polarization reversal and barrier changes, keeping this stack distinct from research devices using other electrodes and ferroelectrics.
Return to This Technology’s Operation SequencesA Shared Vocabulary
Identify the physical meaning and measurement level of a number before comparing cost or suitability.
Traceable Sources and Reusable Content
Structured study records are the shared source for this site: bitcells, operations, limits, patents, comparison conditions, and annual milestones. Future presentations can reuse the same material while preserving sources and boundaries.
Research Revision: 2026-09-10. Classification, sources, maturity, and limitations remain separate data fields. Presentation layout and editorial selection can be prepared independently.
303 Source Records
No sources match. Clear the search or use a company, title or source ID.