Intel Optane, RIP
Another division of Intel dies
When Intel first announced the technology, at IDF 2015 if I remember correctly, I was sitting next to a very sceptical Dave Eggleston, the VP of Embedded Memory technology at GlobalFoundries. Lots of big promises were made - a new memory that was bit addressable, ‘almost as fast as DRAM’, but non-volatile. Sitting somewhere between DRAM and SSDs, it planned to open a path to multi-TB DRAM systems, and had a multi-year roadmap ahead.
Then came the slow downfall. Intel refused to talk about what was under the hood. I remember a very heated briefing a few months later in the UK, where one Chris Mellor (now @BlocksandFiles) ended up shouting at the VP for lying to him. Intel had already stated that it was using a ‘bulk property transition material’, and the industry had suspected that Intel was using a phase change chalco glass to do this, as only a phase change device could be that fast. The Intel VP outright lied, saying it wasn’t phase change, and so Chris shouted at him. I was surprised at the time, it seemed really unprofessional, but in retrospect, Chris was right to push the issue. In time, we found out what we had suspected was right. The only thing was that Intel had a hard time yielding and productizing.
The first big blow was that despite being advertised as a DRAM add-in, the first products were storage related. To add to the confusion, these products were called ‘Optane Memory’, despite not being a memory but a storage. Storage is already non-volatile, so this was super expensive storage, and the capacities were so low that it ended up in cheap laptops as a caching drive. Eventually we saw 375GB SSDs with it, offering amazing QD1 performance, but it was still SSDs. The big play for Optane was as a DRAM replacement, called DCPMM. It came with Skylake, which was delayed, and Intel never sampled it to the press for testing. That only came with the second generation on Ice Lake.
The Optane DRAM allowed multi-TB systems to work in two different modes, like Xeon Phi, but it had applications in databases for quick lookups, quick startups, and in HPC it offered 48TB+ in a 4 socket system when Ice Lake rolled around. But stuff got heated on the business end.
3D XPoint was a joint venture between Intel and Micron, manufactured at the co-owned fab in Lehi, Utah. Each company got 50% of the production, which Micron sold right back to Intel to use. After 5 years from the announcement, due to the agreements and Intel having management issues, Intel sold its half of the fab to Micron as per the contract, as it didn’t want to take on the risk itself. It was still buying the 3D XPoint chips from Micron, but then Micron pulled out of the business altogether. Micron sold the fab. So Intel had nowhere to manufacture it, it was doing the research in China at the NAND fab, but that was sold to SK Hynix, so research went to Arizona. But ultimately Intel were sitting on a bunch of 3D XPoint memory, and the price was so high that sales were slow.
In the past week, Intel started winding down the Optane business. It listed a $559m write-off in Optane inventory in the recent financials, and has stated that work has stopped on future lines of Optane. That’s despite Intel saying as recently as two weeks ago that it has a sustained future planned for Optane. I confirmed with Intel that ALL of its Optane business is being mothballed, however certain customers who need it for upcoming Sapphire Rapids, they’re working with them to get it to work.
I really liked Optane. It was a new technology that fit in the market, but really Intel’s productization had to crap the bed with terrible 16GB caching drives first because management wanted to see a more immediate return. Optane’s real deal was the big SSDs and the DCPMM memory modules - the idea of 512GB+ on a module was great. Really what Optane needed was some form of universal fabric and composability to expand Optane use, which would have come with CXL (originally an Intel technology), but the timing didn’t align. Optane would have been great for CXL. Now we won’t ever see it in action.
Ultimately, Optane was a generation ahead of its time, came too early, came in too expensive, and was cut before it had a chance. Financial decisions like this happen all the time (Qualcomm Centriq, anyone?) and the world can be worse off for it. Investors though, short term, might see the uptick.
This was just a short summary, but here’s my 20 minute video on the topic going into more detail.
Hey Ian, loved your comment: "Really what Optane needed was some form of universal fabric and composability to expand Optane use, which would have come with CXL (originally an Intel technology), but the timing didn’t align."
Remember your visit to the GigaIO office a couple months ago where you got to see FabreX in action? It's ironic that Optane is now dead, yet we had (part of) that solution all along - FabreX is in fact a universal fabric, operating at native PCIe but outside the box, as you saw.
The testing we conducted with Optane SSDs, both composed and over NVMe-oF, showed that for the first time a pool of Optane could be shared across servers at full performance. The transport used to kill the performance, but with FabreX the transport got out of the way. Oh well, next!
Minor typo: the city in Utah is Lehi, not Leti.
Excellent article though. I'm sad to see this one go even if the business reasons are fairly clear. Some former colleagues got great speedups using this stuff as RAM for some HPC/data analytics workloads. It lets a single machine handle large datasets that would normally require distributed computing. Avoiding distribution can be a huge win for speed and programmability.