AI News

New Memories: How MRAM, PCRAM, ReRAM, and FRAM Break Through the Memory Wall

Semiconductor Industry WatchUpdated Nov 16 19:54

AI summary

The four types of emerging non-volatile memories—MRAM, PCRAM, ReRAM, and FRAM—are driving the memory industry to shift from architectural iteration to a revolution in underlying media. In 2025, the global market size has exceeded US$7 billion and is expected to reach US$18.8 billion by 2031, with a compound growth rate of 17.72% from 2026 to 2031. Each of the four technologies has its own shortcomings: MRAM struggles to improve the yield of high-density arrays; PCRAM has high write power consumption and material fatigue; the randomness of device resistance in ReRAM affects consistency; and FRAM is limited by three-dimensional stacking processes and can only be mass-produced in small capacities. Software and hardware supporting optimizations lag behind technological iteration, which has also delayed the process of large-scale commercialization.

Body · AI translation

How much do you know about emerging memory? Original source: Semiconductor Horizons (Ban Dao Zong Heng) Published: 2026-10-26 10:30 Author: Cai Lu Memory Computing Power Generation Poster. Emerging memory is driving the memory industry from "architectural involution and iteration" into a new cycle of "underlying medium revolution."

In 2026, the global memory industry is in a super upcycle seen once every fifteen years. The explosion of AI computing power has driven supply-demand shortages across the entire track, but market enthusiasm is concentrated on iterative products such as HBM, high-end DDR5, and enterprise-grade SSDs, which achieve performance upgrades through process miniaturization, three-dimensional stacking, and bandwidth expansion. However, such optimizations are all incremental improvements within the traditional memory system—merely patching up existing architectures—and cannot break through the physical limits of materials.

As AI inference, high-frequency caching, and massive real-time data processing scenarios are deployed, the underlying flaws of traditional memory architectures continue to amplify, and the industry's growth bottleneck is visible to the naked eye. At this critical juncture of industry transformation, emerging non-volatile memories represented by MRAM, PCRAM, ReRAM, and FRAM rely on entirely new physical mechanisms—magnetic moments, phase transitions, resistive switching, and ferroelectric polarization—to reconstruct the logic of data storage, fundamentally solving the problems of computing power loss and latency. They have become the core underlying support for adapting to next-generation AI computing power and reconstructing new computing architectures, and are also driving the memory industry to officially move from "architectural involution and iteration" into a new cycle of "underlying medium revolution."

Traditional memory is gradually "becoming insufficient," and emerging memory breaks through

The core driver of this industry-level medium revolution stems from the thorough pressure that current high-end computing scenarios are placing on the traditional memory system.

Today, emerging applications such as real-time inference of AI large models, high-frequency edge computing, high-reliability automotive operating conditions, and ultra-low-power IoT networking are being fully deployed. Market requirements for memory have long shifted from the past "large capacity, low cost" to comprehensive performance featuring low latency, high stability, low power consumption, and high endurance. Traditional memory follows a four-tier hierarchical architecture of "SRAM high-speed cache—DRAM main memory—Flash memory—HDD hard drive." The tiers are distinct, but there are inherent performance gaps, which the industry calls the "memory wall" that constrains the release of computing power.

SRAM relies on transistor flip-flops to latch data, requires no refresh, and achieves read/write speeds at the sub-nanosecond level. It is currently the fastest general-purpose memory and is mainly used as on-chip L1, L2, and L3 high-speed cache for CPUs and AI accelerators. However, its cell structure is complex, its chip footprint is large, its integration density is extremely low, and its cost is high, so it is only suitable for small-capacity high-speed cache scenarios. DRAM relies on capacitor charging and discharging to store data. It has a simple structure, higher integration density, and far lower cost than SRAM, and is the main memory of mainstream systems. However, it requires continuous power to refresh and maintain data, and its read/write speed and response latency are weaker than those of SRAM. Flash memory is divided into two types: NOR Flash and NAND Flash. NOR has fast read speeds and is suitable for storing device firmware and program code. NAND has high storage density and excellent cost-effectiveness and is the core medium for consumer SSDs and mobile storage. Both are non-volatile memories that can retain data for the long term, but they generally suffer from inherent drawbacks brought by the block erase-write mechanism, such as high read/write latency, limited erase-write endurance, and rising power consumption under high-load conditions. Traditional mechanical hard drives occupy the cold data archiving market with their ultra-large capacity and extremely low unit storage cost, but their read/write rates and response latency are poor, making them unable to adapt to high-frequency, high-speed computing scenarios.

This "memory wall" effect caused by architectural fragmentation is infinitely magnified in high-density AI computing scenarios. Problems such as insufficient bandwidth, high read/write latency, and large amounts of wasted power continue to drag down the overall release of computing power, becoming a technical shackle that the industry urgently needs to break. Against this backdrop, emerging non-volatile memory technologies that break away from traditional architectural improvement approaches and rely on entirely new materials and physical mechanisms have risen in response.

At the industry definition level, a strict distinction must be made: popular technologies such as HBM and HBF are merely advanced packaging optimization solutions and are still essentially based on iteration of traditional DRAM and NAND media; they are not revolutionary memory media. The truly new-generation emerging memories specifically refer to the four core media that operate on the principles of magnetic moment changes, phase transition effects, resistive state switching, and ferroelectric polarization: MRAM, PCRAM, ReRAM, and FRAM.

Data from Mordor Intelligence shows that in 2025 the global market size of emerging non-volatile memory has exceeded US$7 billion and is expected to reach US$18.8 billion by 2031, with a compound annual growth rate of 17.72% from 2026 to 2031, far exceeding the growth rates of the traditional flash memory and DRAM markets.

The four major emerging memory technologies have already penetrated various industries

At present, the four mainstream emerging memory technologies each perform their own roles and complement each other in differentiated ways. Leading global semiconductor companies have all made in-depth arrangements, and industrialization is advancing steadily.

MRAM (magnetoresistive memory) is positioned around ultra-high reliability, ultra-long service life, and wide-temperature-range operation, making it the optimal replacement solution for demanding scenarios in industrial control, automotive electronics, and aerospace. The technology relies on electron spin magnetic moments and changes in the magnetic moment of magnetic tunnel junctions to read and write data; after power loss, the magnetic state remains stable, and it combines extreme read/write speeds close to SRAM with the long-term storage characteristics of flash memory, while also offering the core advantages of high-temperature resistance, radiation resistance, near-unlimited erase/write endurance, and ultra-low operating power consumption. The global lineup of companies involved is comprehensive: Everspin is the industry benchmark vendor, having achieved large-scale mass production of Toggle-MRAM and STT-MRAM and deeply cultivating the automotive, industrial control, and aerospace markets; TSMC focuses on developing 22nm and 16nm embedded eMRAM processes; Samsung is deploying MRAM embedded in advanced-process chips and continuously advancing process iteration; NXP and Renesas Electronics, together with foundries, integrate eMRAM into automotive-grade microcontrollers to suit in-vehicle OTA fast flashing scenarios; IBM, Qualcomm, and GlobalFoundries are deeply engaged in R&D on magnetic tunnel junction materials and array architectures. Data from Yole shows that MRAM is the fastest-commercializing technology route among the four types of emerging memory, and embedded MRAM has gradually entered volume adoption in the MCU and sensor chip markets. At present, an edge AI SoC based on Samsung's 8nm eMRAM process has completed tape-out verification.

PCRAM (phase-change memory) has high technological maturity and strong process compatibility. It relies on the reversible phase transition of chalcogenide materials between crystalline and amorphous states, forming a high-low resistance difference to record data. Its read/write speed far exceeds NAND Flash, device consistency is excellent, and the cell structure is simple, enabling high-density three-dimensional integration. It can bridge the performance gap between DRAM and NAND flash and is suited to high-end scenarios such as high-speed persistent caching in servers, high-frequency trading, and real-time computing. Leading global companies continue to push forward: Samsung achieved small-scale commercial mass production of consumer-grade PCRAM in earlier years and is deeply engaged in phase-change material improvement, multi-level storage cells, and 3D stacking technology; IBM is deeply engaged in basic PCRAM research, holds a large number of core patents, and focuses on multi-state storage and AI compute-in-memory applications; TSMC focuses on embedded PCRAM process R&D to suit near-memory computing scenarios; Micron, STMicroelectronics, and Renesas Electronics continue to iterate automotive-grade PCRAM to suit complex high-temperature in-vehicle operating conditions.

ReRAM (resistive memory) is the core carrier for edge-side AI compute-in-memory and has outstanding computing-power optimization advantages. It stores data through voltage-controllable high and low resistance state changes in oxide thin films, with an extremely simple structure, extremely high integration density, fast read/write speed, and extremely low power consumption. Its most core innovation is that it naturally suits compute-in-memory architectures, eliminating the need for frequent data movement as in traditional architectures, and can implement analog-domain matrix multiplication under a crossbar array architecture, reducing computing power loss and improving operational efficiency. At the industrial deployment level, Crossbar is a leading global ReRAM IP vendor, focusing on crosspoint array and 3D stacked ReRAM technology and providing mature IP licensing externally; Weebit Nano is deeply engaged in embedded ReRAM and has reached an IP licensing cooperation with Texas Instruments, with technology transfer underway, targeting general-purpose MCU scenarios; TSMC and UMC have built dedicated process platforms to support embedded ReRAM tape-out and mass production. Data from Dataintelo shows that in 2025 ReRAM accounted for 31.2% of the technology market share in the emerging non-volatile memory track, making it the medium with the highest attention and fastest implementation in the compute-in-memory field, and it can be widely adapted to devices such as smart wearables, autonomous driving terminals, edge computing, and AI sensors.

FRAM (ferroelectric memory) is positioned around ultra-low power consumption and high interference resistance, suiting low-power IoT and precision industrial scenarios. The technology uses the polarization switching effect of ferroelectric materials to store data; its read/write speed far exceeds traditional Flash, its power consumption is only a few tenths of that of traditional memory, its erase/write endurance is excellent, it does not require complex erase-verify processes, and its stability is extremely strong. Core global vendors have mature technology and complete product lines: RAMXEED (formerly Fujitsu Semiconductor), as the earliest vendor to achieve large-scale commercial FRAM, is the world's largest supplier, with mass-produced products fully covering industrial instruments, smart meters, and automotive electronics; after Infineon acquired Cypress, it gained an industrial-grade FRAM product line, but on September 16, Winbond Electronics announced the signing of an agreement to acquire 100% equity of Infineon's NOR flash and FRAM businesses for US$1.12 billion, with closing expected in the second half of 2027; Texas Instruments embeds FRAM in its MSP430 series MCUs, targeting ultra-low-power portable devices; ROHM focuses on automotive-grade FRAM R&D to suit scenarios such as automotive sensors and battery management systems, while Samsung focuses on R&D of 3D stacked FeRAM, targeting the high-end AI cache market

Popularity is maxed out, but challenges remain: the growth dilemma of emerging memory

However, behind the high industry enthusiasm and rapid technology implementation, the four mainstream emerging memory technologies remain mired in inherent technical shortcomings and industrial ecosystem constraints. Various core difficulties have not yet been fundamentally resolved, becoming the core bottlenecks restricting large-scale industry adoption and high-end breakthroughs.

From a technical perspective, MRAM, with its advantages of high speed, low power consumption, and non-volatility, has become a popular solution for high-end storage, but the industrialization of high-density arrays poses prominent challenges. In the process of large-scale mass production, controlling magnetic domains in high-density storage arrays is extremely difficult; after process miniaturization, the magnetic stability of unit devices is easily affected by the external environment and process errors, making it difficult to steadily improve array yield. At the same time, limited by the physical structural characteristics of magnetic tunnel junctions, device unit miniaturization has hit a bottleneck and cannot, like traditional storage, continuously increase storage density through process iteration, greatly limiting its implementation in large-capacity data storage scenarios; currently it can only be applied in a small range of niche scenarios such as cache and embedded storage.

PCRAM has the advantages of fast read/write speed and resistance to high and low temperatures, making it suitable for demanding scenarios such as industrial and automotive applications, but its shortcomings in power consumption and lifespan are particularly prominent. This technology relies on the transition of phase-change materials between crystalline and amorphous states to store data; write operations require instantaneous high energy to trigger the phase-change reaction, causing write power consumption to be far higher than that of traditional storage devices and making it unable to meet the long-term battery life requirements of mobile devices and low-power IoT devices. At the same time, repeated phase transitions of the material cause fatigue loss, leading to problems such as crystallization failure and resistance drift; the material fatigue effect continuously weakens device performance, severely constraining the product's long-term high-frequency read/write capability and making it difficult to support high-intensity, long-cycle application scenarios such as data centers and high-frequency computing devices.

ReRAM has a simple structure and high integration potential and is a core candidate technology for next-generation high-density storage, but device randomness has become a core obstacle to industrialization. Its storage principle relies on the resistive switching characteristics of thin-film materials, and tiny errors in material preparation and process manufacturing lead to significant differences in resistive switching parameters among different array cells; device on/off resistance and read/write thresholds are highly random. This problem directly makes it difficult to precisely control the consistency of storage array cells, resulting in insufficient read/write stability and data reliability and making problems such as data misreading and storage errors prone to occur. It greatly increases the R&D difficulty of chip design and error-correction algorithms and hinders large-scale array mass production and commercialization.

FRAM has the characteristics of ultra-low power consumption, ultra-fast read/write, and ultra-high endurance, giving it significant advantages in specialty storage fields, but large-capacity large-scale application is severely limited. The core shortcoming of this technology is concentrated in high-density three-dimensional stacking processes. During multilayer stacking and high-density integration processes, ferroelectric materials are prone to stability problems such as polarization degradation, material aging, and interface leakage; the more 3D stacking layers, the lower the overall stability and reliability of the device. Therefore, FRAM can currently only achieve mass production of small-capacity products and cannot accomplish large-capacity, high-density chip integration, making it difficult to suit mainstream scenarios such as big data storage and large-capacity embedded storage; its application track is severely limited.

At the same time, software and hardware supporting optimization for emerging memory seriously lags behind the pace of technology iteration. Current mainstream operating systems, low-level drivers, chip interface protocols, and storage control algorithms are all developed based on traditional storage architectures, and targeted optimization for the physical characteristics, read/write logic, and power consumption features of emerging memory is seriously insufficient. In actual implementation, emerging memory devices cannot fully leverage their core advantages of high performance and low power consumption, and problems such as adaptation lag, abnormal read/write, and wasted performance redundancy may even occur. In addition, the cost for enterprises to carry out scenario adaptation, technology migration, and solution transformation is extremely high, requiring targeted development of dedicated drivers, optimization of system architecture, and debugging of adaptation algorithms; upfront R&D investment, transformation cycles, and trial-and-error costs remain high, greatly delaying the implementation process of emerging memory from pilot applications to large-scale commercialization.

Looking ahead, global emerging memory technology will continue to iterate and upgrade around five core directions: "high performance, low power consumption, high density, compute-in-memory integration, and full-scenario adaptation," with a clear industrial development path and huge growth potential. In the short term, MRAM, PCRAM, ReRAM, and FRAM will continue to optimize mass-production processes and reduce manufacturing costs, continue to penetrate segmented scenarios such as industrial control, automotive electronics, the Internet of Things, and edge AI, and gradually replace traditional Flash, SRAM, and some DRAM functions in some high-end fields, becoming core storage carriers for embedded intelligent devices and highly reliable industrial equipment. In the medium and long term, compute-in-memory technology will gradually mature, effectively breaking the core bottleneck of the "memory wall" under the von Neumann architecture and becoming core support for the lightweight and efficient upgrade of edge AI.

This content is original to the platform, and the copyright belongs to the platform. Reproduction without permission is prohibited. To reprint, please contact the platform. Comments: No user comments yet. Latest News More Related Recommendations

Samsung's self-tested Exynos 2700 performance is about 20% higher than Qualcomm's flagship chip 2026-10-26 11:48

Large quantities of GPUs arrive but cannot be powered on! Amazon exposes the power bottleneck in AI computing expansion 2026-10-26 11:15

AMD Ryzen 5 5500F, can it shake the sub-$100 CPU market? 2026-10-26 11:13

Source

Semiconductor Industry Watch