Intel recently launched their 3D XPoint non-volatile memory (NVM) under the brand name of Optane. The SSD label in some of the branding might imply that it’s a different type of durable storage, but the technology is aimed at applications that would normally use RAM. This marks the beginning of the end of the compromise between in memory and persistent, as Optane is touted as offering the best of both worlds - DRAM performance and SSD durability.
The reality of Optane is that it’s the first commercial release of 3D XPoint. The hardware is still immature, and the software ecosystem to take advantage of it has yet to properly develop. Optane isn’t as fast as DRAM, so it still forces some degree of compromise between in memory performance and ‘on disc’ persistence. It’s not just SSD by name, as for the time being it’s also packaged as an SSD; with initial data center models coming in 375GB capacity as PCIe add-in-cards (AIC), to be shortly followed by U.2 SATA Express variants. Consumer models come in M.2 packaging in 16GB and 32GB capacities, and essentially provide a fast cache for SSDs in the same way that SSDs were used to cache mechanical HDDs not so long ago. Intel has previously talked about packaging 3D XPoint as DIMMs to replace DRAM modules, but that isn’t on the initial Optane roadmap.
The software components of Optane are key to understanding its capabilities at launch. The consumer version simply acts as a cache, with Intel’s Rapid Storage Technology (RST) taking advantage of 3D XPoint as a higher performance store for frequently accessed files. Using Optane as a cache requires a 7th Generation Intel Core processor, and corresponding supported Intel motherboard chipset. The data center version introduces a new set of software drivers called Intel Memory Drive Technology (MDT), which adds Optane SSD capacity to a DRAM pool. This has the effect of presenting a large pool of RAM to applications, but it should be noted that since it mixes DRAM and SSD it can’t be treated as persistent. At this stage it’s effectively cheap (but slow) RAM, so it solves for capacity rather than persistence.
Precisely how well Optane performs has become the topic of some debate. SemiAccurate accused Intel of crossing ‘an unacceptable ethical line’ in the way that performance data was presented to press and analysts. This discussion on Hacker News also shows that there’s confusion over how the technology will fit into system architectures, and hence to what degree the trade off between volatile memory and persistent storage will be eroded. Those doubts and lingering questions focus however on what’s available today, and with investment and time the technology is sure to improve and software will adapt to harness its capabilities. Similar arguments about the performance of DRAM versus static RAM took place decades ago, but it wasn’t long before DRAM became dominant, edging out static RAM to on die cache and niche embedded applications.
The potential impact on applications and architecture coming from the shift to NVM was well illustrated in Adrian Colyer’s QCon London 2016 keynote ‘Unevenly Distributed’. Persistent storage that’s fast enough versus DRAM will first blur the lines between in memory and on disk, and ultimately that choice could become meaningless. That drives through into many other conceptual anchors and design compromises like stateless versus stateful, so over time this new hardware could upend many of the principals that presently dominate software design.
