Why haven't SSDs built on 3D XPoint taken off yet? Don't blame the technology; blame the marketing behind Intel Optane. New Optane NVDIMMs could fare better.
The semiconductor industry has been teasing us for years with the concept of storage class memory, or SCM. This magical new material will be faster than flash, cheaper than DRAM and persistent without the pesky endurance issues that plague flash.
Over the last few years, it seemed as if 3D XPoint would be the unobtainium that would move us into a new golden age of persistent memory computing. That was until Intel and Micron's revelation last month that they will terminate their joint development deal.
I've been fooled before. Phase-change memory, resistive memory, magnetic RAM, memristors and electron spin-transfer torque each looked like they might get me as excited in my 60s as bubble did in my 20s. But 3D XPoint looked different and still does. It has the density and persistence to really be SCM if Intel Optane SSDs and Micron products can deliver on the companies' claims of 3D XPoint being 1,000 times faster than flash and if they can sell it at prices closer to the cost of enterprise SSDs than registered DIMMs.
Profitability and marketing issues
It appears it's the issue of profitability that's caused Intel and Micron to decide to end their joint development of 3D XPoint after the second generation, which is currently in development. Intel uses 3D XPoint in its Optane products. Demand for Optane products has apparently been so soft that Micron has been selling its share of the Lehi, Utah, joint fab's production to Intel at a loss rather than introduce its long-delayed QuantX-branded 3D XPoint devices.
It may be true that 3D XPoint is 1,000 times faster than flash when writing to an individual cell. But the only 3D XPoint products out so far are Optane SSDs that stick that hyperfast technology behind an x4 PCIe slot and the block storage-oriented NVMe software interface.
If Intel Optane SSDs haven't sold as fast as the company's management may have liked, I place a good part of the blame on Intel marketing. It may be true that 3D XPoint is 1,000 times faster than flash when writing to an individual cell. But the only 3D XPoint products out so far are Intel Optane SSDs that stick that hyperfast technology behind an x4 Peripheral Component Interconnect Express (PCIe) slot and the block storage-oriented nonvolatile memory express (NVMe) software interface.
Instead of that 1,000-time improvement, Optane SSDs deliver four to 10 times higher IOPS, or lower latency, than high-end NVMe SSDs using NAND flash, such as Intel's DC P3700. Since Intel Optane SSDs cost about five times as much as those flash NVMe devices per gigabyte of storage, architects have to choose between a relatively little bit of lightning-fast Optane or five times as much pretty darn fast NVMe flash.
Using a 750 GB Intel Optane DC P4800X as the performance tier in a two-tier storage system will provide lightning-fast access to the hottest 750 GB of your data -- write I/Os that go to the performance layer first. The same $3,000 or so that the P4800X costs will buy you 4 TB in the performance tier on a flash drive.
Intel's Optane SSD DC P4800X
Which configuration will make any given set of applications run faster depends as much on the I/O heat maps and long-tail sensitivity of the applications as the latency hit incurred when accessing the storage layer.
Performance-to-cost ratio
It seems to me a new media tier has to deliver a performance-to-cost ratio of at least 2-to-1. That is, the performance boost compared to the next layer down has to be at least twice as big as the cost premium for that media. Optane costs about five times as much as flash and will accelerate access to that hottest data about five times. That ratio of 1-to-1 makes Optane less than compelling than spending the same money on more fast flash, which is fast enough for today's applications.
Intel's recently launched Optane DC nonvolatile DIMMs that should deliver 10 times the performance of Intel Optane SSDs. More importantly, a 512 GB Optane DC DIMM simply appears as nonvolatile memory (NVM) in the address space. That frees applications from having to treat NVM as inconceivably fast hard disks, so an in-memory database can make a remote direct memory access call into persistent of another node to protect data.
Clearly, Intel and Micron disagree about when this brave new world of persistent memory computing will come to pass. I hope it's while I'm still reporting on this business, because it looks pretty cool.
