Tuesday, January 18, 2022

Progress On Storage Class Memory

Storage Class Memory (SCM) is fast enough to be used as RAM, but persistent through power loss like flash. The idea of returning to the days of core storage, when main memory was in effect a storage layer, is attractive in theory. There have been many proposed SCM technologies, such as restive RAM, magnetoresistive RAM and phase-change memory. Intel and Micron got Optane resistive RAM into production, but it wasn't a great success. It wasn't as fast as RAM, but it was a lot more expensive than flash. The main barrier to adoption was that exploiting its properties required major system software changes.

Now, Mark Tyson's UltraRAM Breakthrough Brings New Memory and Storage Tech to Silicon reports on a promising SCM development published by Peter Hodgson and a team from Lancaster University in ULTRARAM: A Low-Energy, High-Endurance, Compound-Semiconductor Memory on Silicon. Below the fold, I comment briefly.

Their abstract reads in part:
ULTRARAM is a nonvolatile memory with the potential to achieve fast, ultralow-energy electron storage in a floating gate accessed through a triple-barrier resonant tunneling heterostructure. Here its implementation is reported on a Si substrate; a vital step toward cost-effective mass production. ... Fabricated single-cell memories show clear 0/1 logic-state contrast after ≤10 ms duration program/erase pulses of ≈2.5 V, a remarkably fast switching speed for 10 and 20 ┬Ám devices. Furthermore, the combination of low voltage and small device capacitance per unit area results in a switching energy that is orders of magnitude lower than dynamic random access memory and flash, for a given cell size. Extended testing of devices reveals retention in excess of 1000 years and degradation-free endurance of over 107 program/erase cycles, surpassing very recent results for similar devices on GaAs substrates.
Note that the advance here isn't the design of the cell, the same team had earlier published similar cells on gallium arsenide. It is that it has been implemented in silicon with promising performance. Hodgson et al write:
Incorporation of ULTRARAM onto Si substrates is a vital step toward realizing low-cost, high-volume production. Si substrates offer several advantages over III–Vs, including mechanical strength and large wafer sizes, thereby allowing fabrication of more devices in parallel and reducing production cost. Moreover, Si is the preferred material for digital logic and has a highly mature fabrication route. In contrast, III–V substrates are fragile, expensive, and generally only available in much smaller wafer sizes, making them less suitable for high-volume production. But III–V semiconductors do provide advantages such as high electron mobilities, superior optoelectronic properties and a greater degree of bandgap engineering, making them the preferred material for LEDs, laser diodes, infrared detectors and for power, radio frequency and high electron mobility transistors.
This work is at a very early stage, testing large single cells. The cell design is complex, but appears to be manufacturable, and the initial performance is encouraging. Nevertheless, any SCM technology faces the same barriers as Optane; it has to be as fast as RAM and cheap enough compared to flash to motivate making the substantial system software changes it requires. Given high manufturing costs before high volumes can be achived, this is difficult.

1 comment:

Blissex2 said...

«any SCM technology faces the same barriers as Optane; it has to be as fast as RAM and cheap enough compared to flash to motivate making the substantial system software changes it requires>»

That depends on whether it is meant to be used for main storage (then it has to be fast but not cheap) or to replace storage (then it has to be cheap but not fast). The main limitations of current technologies are:

* CMOS RAM is fast but volatile and expensive, and needs both refresh and complicated software to ensure resumption of state after power loss.

* Flash is much slower and hotter with writing than reading.

* Disks are cheap but have long positioning times that become a bigger issue as their capacity (now slowly, as out blogger here point out) increases.

My impression is that the main problem with non-volatile memory is that fast and easy resumption of state after power loss is a niche requirement that cannot sustain a large investment in manufacturing capacity, especially if it is incompatible with current CMOS manufacturing techniques; most practical applications, both client and server side, can just be rebooted from scratch.

BTW I have recently discovered the existence of CeRAM, which is interesting.