I attended a Hewlett-Packard Enterprise (HPE) event in California a couple of weeks ago. Eduardo Duarte, a Senior Product Manager for media in HPE's 3PAR group, gave an excellent talk on the role he expects storage class memory to take in systems.
Storage Class Memory
Storage Class Memory (SCM) is a concept that has been talked about for years, and for good reason: it will radically change how we architect systems. It has two essential features: 1) contents are persistent (non-volatile), like storage; and, it is byte addressable, like memory.
Since it has the characteristics of both, which market will it invade? A first approximation analysis: it replaces the costlier technology, i.e. DRAM.
Mr. Duarte began by postulating that each revolution in storage has been driven by new media. Disks wiped out a slew of media - magneto-resistive delay lines, magnetic drums, primary tape - in the then tiny computer market. 3.5" disks killed 5.25" disks. Thumb drives killed floppies, QIC, Zip drives, and others. SSDs have destroyed 15k disks and, soon, 10k disks.
As he looked into the future the big change is the arrival of low-cost TLC (tri-level or 8 bit) and, later, QLC (quad-level or 16 bit) flash memory.
TLC will be the mainstream flash technology in 2020, with QLC winning by 2025. QLC's superpower? It's cheaper than MLC flash - which is good, because almost every other spec is worse.
In Eduardo's opinion, 3D XPoint will find its niche as a DRAM replacement. Intel will ensure it is cheaper, and engineering will ensure reasonable performance.
The Storage Bits take
I think replacement in some cases, but lower-cost extender in most cases.
Why? DRAM's performance isn't easily faked.
As flash prices start to drop again next year, due to new fabs coming on line and improvements in 3D and TLC production, the trend towards broader use of TLC will gain momentum. At the same time 3D XPoint and - I trust - ReRAM (resistance RAM) products from Nantero and Crossbar will start giving architects more options.
2nd gen TLC flash will be 30% cheaper than today's TLC, which means ReRAM won't be cost competitive with flash any time soon. But as Intel has promised that 3D XPoint will be cheaper than DRAM, we can be confident that it will be cost-competitive with DRAM.
But does cost matter when DRAM is faster and more durable? Yes, it does. When NAND flash became cheaper than DRAM around 2005, it soon became the Next Big Thing.
Ditto when ReRAM cost - and availability - is as good as DRAM.
Courteous comments welcome, of course. Disclosure: HPE paid my expenses to attend the event.