Today Intel releases Xpoint M.2 SSDs, a product that could have been great. SemiAccurate thinks this application would be perfect if the technology wasn’t crippled by a lack of endurance.
Intel officially calls these devices Optane Memory and they come in two PCIe3 2x M.2 form factor SSDs, 16 and 32GB capacities. Since the Xpoint chips are 128Gb/16GB each, the single sided sticks have one or two die each. Both are rated for 100GB written per day which is approximately 3DWPD for the 32GB model and 6DWPD for the 16GB. The 16GB model will retail for $44 and the 32GB will cost $77, both prices far lower than SemiAccurate expected.
The stick on a board
Intel is claiming that these drives will be used in conjunction with other storage media will make your computer fly, a claim belied by the fact that they would not publish a single speed rating during their briefing. All of their comparisons were done against magnetic HDDs which aren’t known for speed or random R/W workload latencies.
Even these sparse and peculiar numbers that were presented to the press were pulled from the slide deck when it was given out days later. This is flat-out unacceptable and unethical. Worse yet all of the specs and configurations that showed these tests were against HDDs, something purposefully scrubbed as well. Intel is unquestionably trying to mislead the press and tech sites here, this isn’t accidental.
Driver layers for obfuscation
What was left in is that Intel’s pseudo-RAID driver called “Intel Rapid Storage Technology” is necessary for this combination to work. It bonds the Xpoint SSD and whatever other primary storage medium you have in the system into one large transparent block. Then it determines what needs accelerating and what can go on your ‘legacy’ storage system, all behind the scenes and devoid of user control. SemiAccurate has never seen a scheme like this work but since Intel did not allow independent testing of Xpoint, I guess we won’t know for a while how it works.
When SemiAccurate asked about the intended purpose of these Xpoint SSDs, our moles were both unanimous and clear, it is a sleep/wake accelerator like we told you about weeks ago. This is the killer app for Xpoint, in theory it will do the job far better than Flash and potentially at a lower power cost. Intel’s lack of sequential R/W numbers and independent testing cast a dark shadow on the efficacy of the technology but on paper it is a clean kill.
Back to the real world, SemiAccurate thinks these Xpoint SSDs should be avoided at all costs. Why? Endurance. The drives are rated at 182.5TB written for their lifetime, (.1 * 365 * 5) or 6DWPD for the 16GB and 3DWPD for the 32GB. Compared to the 375GB P4800X there is a distinct lack of overprovisioning. You would be safe to assume that the consumer Xpoint SSDs also use a far lower grade of Xpoint than the enterprise models as well. Both of these factors lead to a 10x reduction in DWPD, about what you would expect.
Since the Xpoint SSDs are restricted to Kaby Lake CPUs on a Z270 based motherboard, it is pretty safe to assume these will be aimed at higher end systems, aka systems with large memory complements. The drivers for Xpoint are also Windows 10 only so you effectively need a brand new high-end PC to run these SSDs. It is not unreasonable to assume that these PCs will come with large memory complements, 16GB for mobile and possibly much more for desktops. Currently 8GB DDR3 sticks cost about $44, exactly the cost of a 16GB Xpoint SSD.
In their removed slides, Intel provides statistics about average users which says both have 2 power cycles, defined as power off or sleep, per day. [Intel, when asked, paraphrased, “Those slides were intentionally removed.”] Think lunch end going home. Almost every laptop SemiAccurate or the people the author associates with sees far more power cycles than that per day. If you imagine students going to classes, home, coffee shops, etc, we can’t see how the 2-cycles average comes about, we think this number should be much higher.
Back to the tech, if you assume a 16GB laptop, you can sleep it 6x per day before you over-stress the drive and exceed its limits. This assumes the magic driver under the hood does nothing else but clocks out space for a sleep/hibernation file. Does this sound like an unreasonable use case? Watching videos or editing photos will chew up 100GB writes in no time, go look up you photoshop scratch file size for a good example of how uncompressed images can eat storage. It should be easy enough to blow through the maximum writes of Xpoint SSDs just by closing your laptop lid every few hours.
Why is this 3/6DWPD rating a problem in Xpoint SSDs when it is the same as most NAND based consumer SSDs? Simple, size. If you have a 240GB SSD with a 3DWPD rating, each laptop ‘lid close’ writes 16GB or so. That translates to 1/15th of a DWPD on a 240GB drive, 1/8th DWPD on a 128GB drive, and 1/32nd on a 512GB SSD. That same number is 1 or .5 on Xpoint due to their meager size.
Worse yet if you write 100% of the drive capacity at once there is no chance for wear leveling. On the 32GB variant, you can sort of do it but nowhere near what is common practice behind the scenes of every modern NAND SSD controller. Intel has also stated that there are no spare blocks on an Xpoint die, nor is there any overprovisioning possible on these drives. One dead sector could be very bad for M.2 Xpoint SSDs.
This is why SemiAccurate is wary of Intel’s new Optane Xpoint SSDs. Endurance isn’t there, period. The technology is not up to the job and it is very unlikely that the drives will meet their rated lifetime, something that usually means catastrophic user data loss. Throw in ‘magic’ software and it makes the potential to recover critical data all the more unlikely and you don’t have a winner. Performance too is probably not there because Intel went to great lengths to not produce any numbers, then unethically pulled the few somewhat related ones they showed off. Avoid Xpoint M.2 sticks like the plague people.S|A
Latest posts by Charlie Demerjian (see all)
- SemiAccurate has Skylake-SP die shots, sizes, and more - Jul 21, 2017
- Skylake-SP has a diverged core - Jul 19, 2017
- Intel’s Purley platform architecture is a step forward - Jul 13, 2017
- Intel launches Purley aka metal Xeons - Jul 11, 2017
- AMD’s Epyc has lots of connections - Jun 26, 2017