Two emerging non-volatile memory types (PCM and MRAM) are expected to lead the stand-alone memory in the next decade.
Traditional memory chips reach technology nodes
The memory industry has now formed three relatively independent markets for DRAM chips, NADA Flash chips, and special memories.
However, with the extension of Moore’s Law, the technical demand is getting higher and higher, and the drawbacks of traditional memory chips have gradually begun to appear.
As chip technology nodes approach their physical limits, the reduction in the number of electrons in capacitors makes DRAM memory more susceptible to external charges; Flash faces serious crosstalk problems during operation, thereby shortening its lifespan; SRAM suffers from poor signal-to-noise ratio and There are also problems with soft faults.
In addition, when the chip process is smaller than 28nm, these problems will become more serious.
Changes in structural design in the AI era affect logic and storage. Machine learning algorithms make extensive use of matrix multiplication operations that are cumbersome in general-purpose logic, driving the development of accelerators and memories.
Performance and power consumption A major disadvantage of SRAM and DRAM as “working memory” in cloud computing and edge computing application scenarios is that they are volatile and require continuous power to hold data (such as weights).
The main new memory candidates are Magnetic Random Access Memory (MRAM), Phase Change Memory (PCRAM).
Both of these memories use new materials that can be designed with high and low resistivity, which in turn represent 0 and 1, respectively.
MRAM controls resistivity by changing the magnetic orientation; PCRAM exploits the change in the arrangement of materials from amorphous to crystalline.
PCRAM: Leading Candidate for Cloud Computing Architecture
Because PCRAM offers lower power consumption and cost than DRAM, and has higher performance than SSDs and HDDs.
PCRAM and even ferroelectric field effect transistors (FeFETs) are good choices because they have the potential to store multiple bits per cell.
In recent years, non-volatile storage technology has made some significant progress in many aspects, which has brought new opportunities for the improvement of storage energy efficiency of computer systems. The use of new non-volatile storage technology to replace traditional storage technology can adapt to computer Technological development requires high storage energy efficiency.
PCRAM is like all emerging memories, but if you are talking to someone who is bullish on MRAM (magnetic random access memory), of course you will say that MRAM is suitable for various purposes, while PCRAM is not, and vice versa.
Compared with the more commercialized MRAM, the expansion potential of PCRAM is one of its most attractive features. The area of memory cells in MRAM is about 10 times that of PCRAM.
This means that the former has much fewer units for the same byte size, so the scalability of MRAM is very questionable.
Although PCM memory technology seems to be about to mature, it may take some time to popularize, but in order to cope with the development of emerging technologies, high-speed storage devices are still an indispensable technology and the focus of the market.
At present, only Samsung and Intel have launched related products in the world, but most of them are non-embedded phase change memory products.
In August this year, Era Core released the 2-megabit programmable read-only phase change memory product “Puyuan 611” based on phase change materials.
This is the first commercial mass-produced phase change memory product in China, and its release marks that AMT has become one of the few companies in the world after Micron and Samsung that has mastered the research and development, production process and independent intellectual property rights of phase change memory.
MRAM shows advantages at the edge
As an alternative, MRAM promises to increase transistor density several times, enabling higher storage densities or smaller chip sizes.
Another key feature of MRAM is that it can be designed as a back-end interconnect layer for embedded SoC products. MRAM can be used to store the SOC’s operating system and applications, eliminating the need to use embedded flash memory chips for this purpose, reducing the total number of SoCs and cost.
High-performance “near-edge” application scenarios, such as defect detection and medical screening, require higher performance. A variant of MRAM called spin-orbit torque MRAM (SOT-MRAM) may prove to be faster and less power-hungry than spin-torque transfer MRAM (STT-MRAM).
Today’s edge devices primarily use SRAM memory, which uses up to six transistors per cell and can suffer from high active leakage power, which affects efficiency.
As an alternative, MRAM can increase transistor density several times, enabling higher storage densities or smaller chip sizes.
More capacity, more compact chips, lower power consumption, sounds like a victory for all processors at the edge.
This is close to the performance ‘touted’ by SRAM, making MRAM an attractive alternative to almost all volatile memories today.
Compared with traditional DRAM and flash memory, a clear gap of MRAM is in its capacity. Such as Everspin recently released a 32Mb device.
But by comparison, the largest NAND parts with 4 bits per cell offer a density of 4Tb. But MRAM is all the more reason to stand out in IoT and industrial applications because its performance, durability, and unlimited battery life more than make up for its lack of capacity.
Possibilities like these allow us to foresee the bright future of new, greatly improved single-chip systems in the future.
MRAM will not become mainstream under the current computing architecture
Under the current computing architecture, logic and storage are separated. Existing transistor technology has been able to achieve several nanometer processes, logic circuits containing billions of transistors, and existing storage can be done at a sufficiently low cost. terabytes.
In terms of logic or storage, spin chips cannot replace the existing mainstream chips, and can only be used in certain fields with specific needs.
Although some have found embedded technologies to achieve a degree of commercial success, they also lag behind cost-effective alternatives to discrete memory. Despite higher performance, durability and retention, or reduced power consumption.
Although MRAM has had some success in the embedded market for discrete applications, even proving that it can handle the extreme environments of automotive applications, MRAM is still a niche memory.
It’s unclear which current or next-generation memory technology is the winner. Maybe all technologies have a place.
All in all, there is no consensus on which type of next-generation memory is more suitable for AI edge applications. The industry continues to explore current and future options.