Summary
In computer architecture, the computer memory hierarchy describes computer storage based on response time. There are four main storage levels: internal storage with process registers and cache; main memory with system RAM and controller cards; disk storage with SSDs, and backup with tape. The top of the hierarchy is small, fast, expensive and volatile and as you move down memory gets larger, slower, cheaper and less volatile. Persistent memory (PMEM) is a technology that sits between system memory like RAM and disk storage. It is known as a “storage-class memory (SCM)” meaning it has DRAM-like speed with the nonvolatility of NAND flash.
Viability (5)
The first PMEM or “storage-class memory” was the 3D XPoint introduced in 2015 by Intel and Micron. It was claimed to deliver 1000x the performance and 1000x the endurance of NAND storage, and 10x the density of DRAM. We have seen a variety of products from all leading memory players. However, in 2021 Micron exited the market and in 2022 Intel followed. Both noted “insufficient market validation” and an industry shift to Compute Express Link (CXL) interconnect memory products. This technology looks to have come and gone with other technologies in particular ReRam and other NVM types as well as CXL memory.
Drivers (3)
Global data storage requirements are growing 20% every year and 30% of the data generated will be real time by 2025. So there is an every growing need to access data at quickly as possible and therefore scale memory. The gap PMEM was filling in the market was the expensive cost of DRAM and the slow speed of SSDs. The cost of DRAM has stayed stubbornly high over the the past 10 years after major declines from 2006 to 2012 after which is basically stayed around $3 dollars per gigabyte. So there has long been a need for a cheaper way to scale fast memory.
Novelty (1)
PMEM is a technology for scaling memory. The value proposition of PMEM versus alternative memory and storage technologies is that it is faster than flash and cheaper than DRAM. It is attempting to be a new layer in the computer memory hierarchy above flash and below DRAM. Squeezing in a new layer requires winning market share from DRAM in the memory market and flash on the storage market. Both of which are hard to achieve in terms of competition and economies of scale in those mature and entrenched products. Other memory scaling solutions like CXL interconnect built on top of the PCIe standard offer a way to efficiently share memory between CXL and main memory to allow a server to expand its memory capacity and increase bandwidth. CXL is not bound by the same size and power limitations of a DIMM unit.
Diffusion (2)
The biggest restraint is that because PMEM is slower than DRAM and has novel persistence features, server software cannot, in DIMM form, treat is as DRAM. Deploying PMEM requires rewriting software in a way just adding more DRAM or SSDs does not. All the major memory players from Samsung to Intel to Western Digital had or have PMEM products in market, but at this point it seems clear the industry has moved to CXL memory. The restraints to adoption for PMEM are limited, the questions were commercial in terms of performance and cost, and it seems CXL delivers much more scalable performance and PMEM costs are not materially cheaper. Creating a new market between two established markets with DRAM and flash was always going to be a tough sell.
Impact (1) Medium certainty
It already seems that PMEM is dead on arrival after the Micron and Intel exits noting a focus on CXL interconnects which can scale memory on a more affordable on a $/TB/server basis. The concept of fast and non-volatile memory is still compelling, but can be achieved in more cost-effective ways either by making DRAM more persistent with power-fail protection or flash with a caching layer. The future will be disaggregated memory pools that are not constrained by rack space efficiently orchestrated over the network.
Sources