A couple of weeks ago Intel and Micron jointly announced their Xpoint - pronounced "cross point" - memory. Offering few details they claimed
3D XPoint technology combines the performance, density, power, non-volatility and cost advantages of all available memory technologies on the market today. The technology is up to 1,000 times faster and has up to 1,000 times greater endurance than NAND, and is 10 times denser than conventional memory.
3D Xpoint is the technology inside the Optane drive.
At last week's Flash Memory Summit, semiconductor analysts Jim Handy and Dave Eggleston, whose work I respect, offered their educated guesses on the Xpoint technology. I'm stealing freely from both.
Xpoint is a form of resistance RAM (RRAM) based on technology that Micron bought with their acquisition of Numonyx in 2010. Numonyx was formed from Intel's NOR flash and STMicro's NAND flash businesses and was the first to ship Phase Change Memory (PCM) a form of RRAM where the junction material is heated to change its resistance, unlike filamentary types of RRAM.
Mr. Eggleston's conclusion is that Xpoint is a rebranded PCM or PCMS (S for selector) memory, which finds backing in a July req from Micron for
. . . engineers to support the development of advanced Phase Change Memory based non-volatile memory products. As a Process Integration Engineer, you will contribute to . . . [PCM] technology in R&D and subsequent transfer of the process to a production Fab.
In addition, Micron recently received a patent for "Accessing memory cells in parallel in a cross-point array" which refers chalcogenide materials and PCM. It's not certain, but PCMS is the clear favorite.
The 3D part is - like 3D flash - building up multiple layers of devices to achieve high capacity in a single chip. This answers a common criticism of PCM - that cell sizes are too large to be economic - by making the vertical axis the limiting factor.
The Intel/Micron annoucement showed a 128Gb chip made of two 64Gb layers. Compare that with Samsung's recently announced 48 layer, 256Gb flash chip, and you can see that Micron has a lot of headroom to increase capacity.
The good news: Xpoint will be cheaper than DRAM. Bad news: it could easily cost more than 10x flash, because DDR4 DRAM is about $8/GB, while flash is about 30¢/GB.
A 250GB Xpoint SSD could easily cost more than $1,000, depending on how aggressive the pricing is. Which is the rub: flash didn't take off until it was cheaper than DRAM, but flash wasn't competing against something cheaper.
Volume is key for driving down chip costs. If yields are good, forward pricing will drive adoption. If not, prices will reflect relative scarcity.
No firm dates announced, other than 2016, which means this could easily slip into 2017. Don't hold your breath.
This is a server technology, not a consumer play. Since Optane SSDs will be faster than flash SSDs and lower power than DRAM, they'll be most attractive to hyper-scale data centers first, even with their higher cost.
That also simplifies Intel's distribution and support, giving them better control of any bad news that develops. Don't expect to see them at NewEgg before 2018.
Do we really need another layer in the storage pyramid? As Jim Handy sees it, semiconductor memory and storage is getting faster and cheaper, while disk and tape are only getting cheaper, which opens up a gap in the storage hierarchy over time.
Given that the current gap between DRAM and SSD access times is about 100,000x, we could use a faster non-volatile memory layer. Combining the efficiency of PCIe hardware, NVMe software and a ≈10x faster SSD might reduce that gap to 1,000x.
While Intel and Micron have said Xpoint won't replace DRAM, there's no reason it couldn't be fitted to DIMMs, like Diablo's Memory 1 product, to extend server memory capacity. Xpoint is years away from becoming a consumer technology, but we'll all benefit if our hyper-scale backend systems become faster and more efficient.
Comments welcome, as always.