Micron Abandons 3D XPoint Memory Technology
by Billy Tallis on March 16, 2021 6:00 PM ESTIn a sudden but perhaps not too surprising announcement, Micron has stated that they are ceasing all R&D of 3D XPoint memory technology. Intel and Micron co-developed 3D XPoint memory, revealed in 2015 as a non-volatile memory technology with higher performance and endurance than NAND flash memory.
Intel has been responsible for almost all of the commercial volume of 3D XPoint-based products, under their Optane brand for both NVMe SSDs and persistent memory modules in the DIMM form factor. Micron in 2016 announced their QuantX brand for 3D XPoint products, but never shipped anything under that brand. Their first and only real product based on 3D XPoint was the X100 high-end enterprise SSD which saw very limited release to close partners. Micron has now decided that further work to commercialize 3D XPoint memory isn't worth the investment.
Micron currently owns the only fab equipped to mass-produce 3D XPoint memory: the Lehi, Utah fab that was formerly the home of the Intel-Micron flash and 3D XPoint joint venture IMFT. Intel and Micron began splitting up their partnership in 2018, first parting ways for 3D NAND flash memory development, followed by dissolving the 3D XPoint partnership after completing development on the second generation 3D XPoint. In 2019, Micron exercised their rights to buy out Intel's share of the IMFT fab, leaving Micron as the sole owner of the fab and Intel in the position of buying 3D XPoint wafers from Micron to use in Optane products. Intel's Optane products have not been enough to fully utilize the capacity of that fab, and Micron's non-GAAP operating profits have been taking a hit of over $400 million per year in underutilization charges.
Micron is now putting that 3D XPoint fab up for sale, and is currently engaged in discussions with several potential buyers. Intel is the most obvious potential buyer, having recently begun the long process of selling their NAND flash and flash-based SSD business to SK hynix while keeping their Optane products. Intel has already moved their 3D XPoint R&D to Rio Rancho, NM but has not built up any 3D XPoint mass production capacity of their own; buying the Lehi, UT fab would save them the trouble of equipping eg. their NAND fab in Dalian, China to also manufacture 3D XPoint.
However, Intel is not guaranteed to be the buyer of the Lehi, UT fab. They've doubtless had opportunities to do so before as Intel and Micron unwound their partnership. Micron states that the Lehi, UT fab could be used to produce analog or logic ICs, not just memory—and that converting it to large-scale manufacturing of DRAM or NAND flash memory would not be as appealing to Micron as simply expanding capacity at their other existing fabs. With widespread semiconductor shortages affecting almost all corners of the industry, this fab is likely to sell quickly even if the buyer needs to put substantial effort into retooling.
Micron does not have a direct replacement lined up for 3D XPoint memory technology, but continues R&D into new memory and storage technologies. Micron's announcement is emphasizing a pivot toward developing memory products that will use the Compute Express Link (CXL) interface, which promises to be a vendor-neutral interface for DRAM and non-volatile memories such as 3D XPoint.
Related Reading:
Source: Micron
60 Comments
View All Comments
Chaser - Tuesday, March 16, 2021 - link
I agree. PCI 4.0 was the nail in the coffin at least for Desktop usage. The upcoming Direct-Storage API makes it terminal. Funny thing is I still own the Star Citizen ship that was included in the 900P Optane drive.tuxRoller - Wednesday, March 17, 2021 - link
Direct Storage api (gpudirect) only reduces overhead by eliminating the requirement of goiny through the os's block layer but not the problem of media access latency times.It's the latter one xpoint is attempting to address.
Eliadbu - Wednesday, March 17, 2021 - link
I think that Micron just gave up too quickly. 3d Xpoint is where NAND was 10 years ago expensive and limited capacity but like NAND then it has advantages over other technologies. It's disadvantage is NAND is good enough for most uses so it's not compelling to go the expensive route. Micron wants to invest its money and resources somewhere else and it is understandable, the question is how further does Intel is willing to push the technology.Slash3 - Wednesday, March 17, 2021 - link
While it's true that there are Gen4 (and upcoming Gen4) NAND based SSDs which have incredible transfer rates, for something like 4K Q1T1 random read/write, NAND is still much, MUCH slower than Optane.A high end current SSD like the SN850 or 980 Pro may deliver 60-70MB/s under such conditions, while Optane will deliver ~300MB/s. The same goes for queue latency (extremely low) and IOPS (extremely high, across all queue depths).
That we're not going to see a successor to the 905P at the enthusiast desktop level is genuinely sad. For a lot of workloads, the jump from an SSD to Optane is just as significant as the jump from HDD to SSD was.
I hope they, or someone else, revisit it outside of the datacenter.
romrunning - Wednesday, March 17, 2021 - link
Yes, Optane was easily the best thing to have for a boot/system drive, and the technical benefits at lower queue depths (which is the typical I/O load on desktops) were great.But neither Intel nor Micron (especially Micron) put more effort into lowering the costs. I'd easily buy an Optane drive over a premium SSD if it was 50-100% higher cost, but they couldn't even get it down into that range.
Very sad for storage as the industry keeps going down, instead of up, the performance ladder with QLC, PLC, etc.
Jon Tseng - Wednesday, March 17, 2021 - link
I suspect the differentiation is more in memory DIMMs rather than storage SSD. Direct access let's you have vast pools of (albeit slow) system memory. Effectively this lets you combine the worlds of "big data" and "in memory" compute.FunBunny2 - Wednesday, March 17, 2021 - link
"Effectively this lets you combine the worlds of "big data" and "in memory" compute."makes one wonder whether Oracle TimesTen and SAP HANA might run at least OK with a true NVM and not have to worry about power and writing to storage and such. cutting out more than just some of the OS routines...
Lolimaster - Wednesday, March 17, 2021 - link
It's not about brute speed but endurance and a near order of magnitude less latency.Kamen Rider Blade - Tuesday, March 16, 2021 - link
I really wish they would try to make these take off as basic OS drives.They have the right latency factor and latency once the drive is filled to function as ultra low latency OS drives.
Billy Tallis - Tuesday, March 16, 2021 - link
They're excellent as OS drives and cache drives. But SLC is pretty good for that too, especially for more read-oriented workloads. And pseudo-SLC caching doesn't require a new SSD controller, and is a lot more flexible than having to decide how many dies of 3DXP vs dies of 3D NAND to include.If Intel had done their own NVMe SSD controller that could interface to both 3D NAND and 3DXP and allowed them to implement caching in the SSD rather than be limited by the hassle of custom storage drivers for motherboards and Windows, then products like the Optane Memory H10/H20 could have been a lot more interesting. But Intel doesn't think it's worth their effort to design their own consumer-focused SSD controller.