What do you do when your product doesn’t really work and you have already provided excuses that beggar belief? If you are Intel and the product is Xpoint, you get only semi-tame press together and feed them BS. [We suspect some of the press present have as much disbelief as Semiaccurate. However, unlike Semiaccurate they are of the extraordinarily polite type and will tone down their disbelief in public.]
SemiAccurate is here to tell you about Intel’s top-secret Xpoint briefing next week in Folsom, CA. On the 15th Intel will be bringing in a bunch of press to talk about Xpoint memory, you know the one that was badly broken. And still is. The idea is to get them all in a room and show pretty slides with the talking points from the last round of public backpedaling offered as ‘truth’. If you only get the (hopefully) tame press there, they won’t ask the obvious questions and when the embargo lifts, you get a ‘new truth’.
Don’t believe the hype, Xpoint is still woefully broken, broken to the point that we question the business acumen of a company that would release such a product now. Before you think this is idle chatter and backstabbing, lets look at the history of Xpoint. It was announced in June of 2015 as we linked above and SemiAccurate was at the launch. It sounded good. During their Q2/2016 analyst call, Intel made the claim that the product was sampling at that time, something we know to be true because several sources that SemiAccurate talked to had said samples. What Intel didn’t go on to explain was that those samples didn’t actually work, they were broken in one of the most important functions of the technology. They did not have the write endurance of a first gen SSD with no wear leveling. But technically they did ship.
Intel went on to say that the DIMM format version, internally code-named Apache Pass, would ship next year, IE 2017. What they didn’t say is that the roadmaps given to OEMs months prior had all mention of Apache Pass were removed bar a few things like, ‘firmware supporting Apache Pass will be ready on date XYZ’. Previous roadmaps had slides about capabilities, layouts, and the rest, but they went poof. All that was removed before Intel’s promise and replaced with, “For eligible customers, see customer update on Apache Pass and Coldstream SSD for detailed schedule.” That doesn’t seem to exude confidence in a near-shipping product. Xpoint has now been removed from the first gen Purley just like SemiAccurate exclusively told you last year.
Then there was IDF 2016 as we detailed in our story about Xpoint being broken. During that show there were no bare boards shown because people might notice the overprovisioning needed to keep Xpoint devices running for ~3 days and ask reasonable questions. You know those questions Intel really doesn’t want the press asking next week. They also promised developers access but that meant use of a remote server with Xpoint SSDs on them, no chance of actually getting a device to test yourself. Why? Performance and endurance, cloud drives can be regularly swapped without user awareness.
Intel’s own IDF 2016 Xpoint slide
At IDF the performance claims went from “1000X the endurance of NAND” to “Endurance 3X the drive writes per day”, a 333x performance drop on the most key metric for the technology. Please note these numbers were taken directly from Intel slides publicly presented and often backed up by executives on stage. Performance similarly went down by 100x and density went down by 2.5x, again all taken directly from Intel’s own presentations. This wasn’t just a re-adjustment based on final bins, it is a devastating indictment that Xpoint wouldn’t even be in the same class as promised a year earlier. Worse yet SemiAccurate’s sources at the time who were involved with the technology told us that those targets were still blue sky promises, Xpoint was nowhere near 3x flash’s endurance at that time. Xpoint was and is still broken.
So what should Intel do at this point? You might look to Micron who is the co-developer of Xpoint and gets their silicon off the same line as Intel. For some strange reason they have yet to even announce an Xpoint product, or even hint that something purchasable is coming. Intel on the other hand is promising the moon and repeatedly, and silently, missing ship dates. Unfortunately SemiAccurate can’t explain this disparity, if you can, please let us know why, there has to be a sensible reason lurking somewhere, right?
Back to what Intel has to do, the answer is simple, mislead. We use that word because SemiAccurate’s editor doesn’t want us to call Intel liars, that would be accurate, not SemiAccurate, behavior. At IDF when this was pointed out to Intel, they took a curious route to explaining it. The official explanation was that the numbers presented were for devices, not for chips, Having been at the Xpoint launch and having talked to many Intel executives and engineers there, the author can definitively say this is BS. It begins to look like Intel is trying to find a scapegoat and hope the press are too technically unaware to ask the right questions. Why?
And another IDF 2016 Xpoint slide
Once again the slide above is from IDF 2016. This is not a 1000x performance difference, if you count pixels on the bars, it is hard to come up with anything but 10x gain they presented on earlier slides. At that point Intel tries to confuse the issue by claiming “Faster” is not “latency” or other items, this is grasping at straws. In many conversations with Intel personnel the metrics were made clear, and in many conversations with developers working on Xpoint, it is even more clear that the performance targets are not met. Many of the developers of enterprise Xpoint products say that the delivered devices are orders of magnitude slower that what they were promised and line up with latest downwardly revised numbers the IDF slides. One even laughed when we told him Intel’s excuses. Being a family publication, we won’t repeat what he said after he stopped laughing.
So Xpoint is still broken. It is orders of magnitude slower than promised and may now barely reach the write endurance of flash. This assumes there was some magic that happened since the last time SemiAccurate talked to Xpoint developers and engineers. That magic would mean the endurance backpedal was only 333x, not far worse as our sources have been indicating. Fingers crossed people.
What is Intel going to demo next week to the press? Two things from what we hear, the M.2 form factor SSD for Intel’s Z270 boards and the DC P4800X. DC P4800X is a datacenter SSD with the curious capacity of 375GB and performance figures that lose to Intel’s own flash on read but beat them in writes. The bugaboo however is that in the device level specs SemiAccurate has seen, the endurance is not at all what Intel has led customers to expect. On a device level Intel’s flash drives have twice the rated endurance of Xpoint according to the latest specs.
There is more to it than the overall device numbers. Compared to Intel’s workhorse DC P3600 series flash SSDs with a write endurance of 22PB written, the DC P4800X is sitting at ~12PB. The smallest P3600 is ~4x the size of the P4800X so if you look at writes per GB of capacity, it is ~5.5PB/400GB vs ~12PB/375GB. That isn’t the 3x difference promised. If you do the same math for the largest P3600 drive at 4TB, you end up with 2.5x less endurance. That means Xpoint beats Intel flash by 5-6x in endurance but based on years of looking at how Intel rates drives, the rating is almost assuredly based on the smallest device in the family. Feel free to do the writes/cell math for yourself if you are bored.
So comparing Intel’s own best flash SSD to their own best Xpoint SSD in the same market segment, with the same interface, and the same almost everything else, things line up nicely with the dire numbers from IDF 2016. Performance is in some cases a bit better, worse in others, and endurance is barely better than flash on paper. To put a bright spin on things, latencies are about half that of their flash counterpart, but that is most likely down to controller engineering tradeoffs. Also bear in mind that Intel’s SSD lines are middle of the road for enterprise SSDs, comparing the DC P4800X to high-end providers would not be pretty.
The one spec where the DC P4800X shines is writes per day where it beats flash by 10x, 3 vs 30. Unfortunately that number isn’t real, it is just a marketing guideline which roughly equates to PB writeable divided by expected (rated) lifetime. Without knowing the rated lifetime, this ‘crushing win’ is nothing more than a way to use technical sounding numbers to mislead the ignorant. Any guesses as to what Intel is going to lead with for the press next week? No points if you said writes per day and latency. Any guesses as to what they won’t say? Sorry, no cookie, too easy.
That brings us to the other form factor, the Xpoint sleep/wake accelerator for the Z270 chipset. The idea is the same as with flash, to put in a small SSD with enough space to cover maximum main memory so you can safely dump DRAM when you shut the lid of a laptop. If Xpoint worked right, it would be a great match for this use case. Painfully expensive, sure, but bang on perfect for the job at hand. If it worked. A full write every time you open and close the laptop lid is expected, and it doesn’t take much to hit that 30WPD (writes per day) metric of the DC P4800X.
That drive is vastly overprovisioned and come with enterprise pricing, not consumer pricing. Intel isn’t going to use the same quality components as the DC line, nor are they going to vastly overprovision it to meet the enterprise ratings. So how long do you expect this Xpoint Z270 accelerator to live? The warranty for most consumer parts is 1 year or less, if you actively use your device a lot, say the workload of an average student, it is going to be quite a trick to keep these turkeys alive for that warranty period. Time will tell but if you buy one, don’t blame us. As we mentioned earlier, SemiAccurate doesn’t feel a company would be making a good business decision to launch this product.
So what is Intel to do? Well if history is any guide they will take a journalist who they know will do their bidding despite potential ethical issues, and pick an outlet that needs Intel ad revenue to survive. Even if that journalist was critical of Xpoint in the past, they know who to lean on. 37 days later you can get a hit piece that conveniently avoids fact-checking or attempting to contact the victims if your patsy is Chris Mellor. This may sound cynical but do note the tone from a piece a mere 7 days later, sans Intel guidance and potential other background factors. These three articles are not conclusive though, so lets add a fourth that hit 27 days later. Not enough? How about another a few months after? Any guesses which ones were “honestly not influenced”?
It is going to be interesting to see what comes from next week’s press session on Xpoint. Intel seems to only be inviting non-hardcore techs to the briefing for all the reasons you would expect. The idea is to feed them selective and purposefully misleading questions and hope they won’t ask the obvious questions. Golden samples will abound and all will look right, the embargo will lift long before endurance testing can be done. Intel, what happened to you?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