Micron Finally Announces A 3D XPoint Product: Micron X100 NVMe SSD
by Billy Tallis on October 24, 2019 4:30 PM EST
Micron and Intel co-developed 3D XPoint memory as a high-performance alternative to flash, but so far only Intel has brought products to market, under their Optane brand. Despite owning the fab where 3D XPoint memory is produced, the closest Micron has come to commercializing that tech for themselves was their announcement in 2016 that upcoming Micron products using 3D XPoint memory would be branded as Micron QuantX, their counterpart to Intel's Optane brand. Years later, we finally have a concrete product announcement, and they seem to have abandoned the QuantX name.
The new Micron X100 is a high-end enterprise NVMe SSD to compete against Intel's upcoming second-generation Optane SSDs and any specialized low-latency SLC NAND their competitors can come up with (eg. Samsung Z-NAND, Toshiba XL-FLASH). Micron has not yet released full specs for the X100, but the top line performance numbers are 2.5M IOPS for 4kB random reads and around 10GB/s for sequential transfers—both likely to be new records for a single SSD if they can ship it soon enough. A preview video posted by Micron includes a graph that labels the 2.5M IOPS figure as being tested at QD1, which sounds too good to be true: almost 5x the performance of Intel's current Optane SSDs. Micron says the X100 should be good for at least 9GB/s for reads, writes, or mixed workloads, reflecting how much closer 3D XPoint is to symmetrical read/write performance than any flash memory. (And also suggesting that the controller may be the bottleneck for sequential transfers more than the 3D XPoint memory itself.) For QoS, Micron is listing both read and write latencies of 8µs or less, slightly better than the 10µs that Intel's current Optane SSDs promise.
The card Micron is showing off today is a full-height half-length PCIe x16 add-in card, so it should be able to reach full throughput even on PCIe 3.0 systems. Micron says the X100 will be in limited sampling to select customers sometime this quarter, so it's not going to be shaking up the storage market much in the immediate future but it is far enough past the vaporware stage that Micron should be able to deliver the rest of the specs soon—including the range of available capacities. Since Micron hasn't said anything about a second generation of 3D XPoint memory being ready, the density and costs of the X100 shouldn't be drastically different from Intel's Optane offerings.
Source: Micron
Facebook
Twitter
RSS
30 Comments
View All Comments
vFunct - Thursday, October 24, 2019 - link
was hoping Micron would release 3DXPoint DIMM memory modules for AMD Epyc.. could use that for database servers.Billy Tallis - Thursday, October 24, 2019 - link
Intel had to make substantial changes to their CPU's memory controller to make their 3D XPoint DIMMs possible. If Micron had collaborated with AMD to add support for something similar to Rome's memory controller, we would have heard about it earlier. I think it'll be another product cycle or two before Micron's 3D XPoint stuff can move beyond NVMe.Diogene7 - Thursday, October 24, 2019 - link
@Billy Tallis: If what you are saying should be correct, if it should happen, I would think it would mean 2020 at the earliest, but I would think it is more likely H2 2021 at the earliest / 2022.I think AMD will likely wait for the JEDEC NVDIMM-P specification to be completed, which it is hoped will be completed a few quarters after the JEDEC DDR5 specifications.
I am not an expert but my understanding is that a NVDIMM-P compatible memory controller should allow to slot DIMM using DRAM and/or different kind of Persistent Memory (PM) / Non Volatile Memory (NVM) like MRAM, RRAM, Nanotube RAM (NRAM), and so also 3D-Xpoint
I have my fingers crossed that the JEDEC NVDIMM-P standard to be completed in ~H1 2020 or H2 2020 and then it takes approximatively 1year and half / 2 years from that point to be able to purchase the 1st product that support a new standard (ex: PCI-e Gen4 standardized in 2017, and some AMD products support it in 2019) which would make ~H2 2021 / H1 2022 at earliest...
IntelUser2000 - Thursday, October 24, 2019 - link
You guys had an article where 3rd gen Optane will be DDR5 compatible and it would be an inflection point.I took that as to mean it'll work in most platforms with support with much less work needed compared to current Optane so others can use it too.
Thoughts on the X100: Other sites are reporting 8-pin auxiliary port. Getting 3D Xpoint that high bandwidth is costly on power.
Billy Tallis - Friday, October 25, 2019 - link
I don't think Intel has committed to making Optane DIMMs any more cross-platform or standardized than the current versions. The third generation Optane DCPMM could easily just be based on the DDR5 interface in the same way that the current generation is based on the DDR4 interface. There's certainly nothing to indicate that future versions of 3D XPoint will be any easier to attach directly to a DRAM bus without the kind of controller current Optane DCPMM uses on each module. It's possible that Intel could give up on having Optane DCPMM locked to their platform, but I would be very surprised if they made that kind of change before Micron or someone else had a serious competitor shipping in volume.As for the 8-pin power connector the X100 seems to have: the controller has to be quite a bit wider than the first-gen Optane controller, and Micron may have gone for an FPGA instead of an ASIC. That plus the sheer quantity of 3DXP media it will be accessing in parallel can plausibly push the card over 75W. But I also wouldn't be surprised if what they showed isn't the final PCB layout, and I think it's likely that they'll at least have a base capacity model that can run fine on just slot power.
IntelUser2000 - Friday, October 25, 2019 - link
Optane DIMM: Mr. Rob Crooke himself said in this article(https://www.pcworld.com/article/3105854/intel-pave... that he doesn't want to hinder adoption by limiting the technology to their own platforms.Now that's for the SSD. But consider this. They had significant problems allowing the DIMMs to work on their *own* platform, nevermind opening it up. They said "platform challenges" significantly hindered ramp of Optane. And we've yet to see a workstation supporting them despite early as May this year saying they will come with it.
Intel through the AT article said it will be an "inflection point". I don't think they said it out of coincidence. Everyone already expects they will do as they did in the past, so to really change that would suggest opening it up.
X100: Their M2 Optane P4800x varies significantly in power and performance. The 100GB version 2200/1000 sequential with 7W active power and goes all the way to the 375GB version with 2500/2200 but with 11W active power. The in-between 200GB is at 2200/2000 with 9W active.
There's no particular reason to limit the highest performing, most expensive SSD with significantly lower sequential speeds. The puzzling specs are very apparent on their Optane Memory modules. You can also see the write speeds vary tremendously with varying power use. Also their DIMMs can write at nearly 10GB/s but the write is no better at 2.3GB/s. The power is similar at 18W.
Some technical articles talk about Phase Change memory requiring more power for writes. While lot of such articles are theoretical and based on previous experiments, in this case it might be true.
Great_Scott - Friday, October 25, 2019 - link
3DXP "RAM" is so damn slow compared to DDR RAM that it's a miracle this works at all. I'm not surprised that there are numerous issues.I don't see the advantage in using the RAM interfaces as opposed to a PCIe-bus interface for adding a tier between memory and disk.
Diogene7 - Friday, October 25, 2019 - link
I think that one possibe advantage of using 3D-Xpoint on the RAM interface could be to used it as a Non-Volatile RAMdisk : in theory, it would lower access latency compare to accessing 3D-Xpoint SSD through NVMe interface and would make the system more responsive.Even though not as responsive as a volatile DDR4 RAMdisk, in theory, access latency should be better than any NVMe SSD (Optane SSD or Flash SSD) by a factor ~5 to 10.
It is not the best use of 3D-Xpoint, but as a 1step, as a consumer, I am dreaming to replace a 512GB NVMe Flash SSD by a bootable 512GB Storage Class Memory (SCM) (that could be 3D-Xpoint, RRAM, Carbon Nanotube (NRAM),...) on DRAM interface : It would make launching softwares, games,... more responsive which visibly improve the user experience much more than a faster CPU (a bit like going from HDD to SSD make your system much more responsive, but not as dramatic improvement).
Azethoth - Saturday, October 26, 2019 - link
I just launched Diablo 3 from scratch for 18 seconds total. Part one is the Blizzard launcher and its authentication which was about half the time. The rest is D3 and its authentication and launch.Exactly how much faster do you need it to be after you hit play? 10 seconds goes to ... 8 seconds? Yeah it is not nothing but its insignificant compared to the hundreds of hours of playtime. You can even be semi logged out, wake computer from sleep, and it logs in and continues in < 2 seconds.
Don't get me wrong, im gonna buy this, but mostly because YOLO, not for any practical benefit. Maybe one day if game state is saved and almost instantly restored from non volatile memory there will be a benefit. But that needs OS architecture changes.
Diogene7 - Sunday, October 27, 2019 - link
I am interested by any usage of Persistent Memory (PM) that improve overall system responsiveness by a factor of at least x10 or x100 because, in my opinion, responsiveness is much more valuable than CPU raw speed : I much prefer that when I click to launch a software, it launch instantly with minimum load time / boot time but that the computer speed would be slower.Ex: As an average consumer, I would like that the computer boot time is as close to zero as possible, and that when I click to launch a software to convert a video, the software opens instantly and is immediately ready to use because I can be immediately productive, and launch a set of tasks —> That is responsiveness
However, I don’t mind that it takes 45mn to convert a video instead of 20mn with a faster CPU, because if I know in advance how much time it will take, after I launched the tasks, I can do other tasks —> That’s why speed is less important to me than responsiveness
As a 1st step, using