The Holy Grail of computing is to have terabyte-sized persistent DRAM in servers. While 3D XPoint memory looked...
to be the strongest contender in that quest, recent delays may allow the competition to catch up.
Persistent dynamic RAM (DRAM) could lead to applications that no longer need to store data in 4 KB blocks using a slow, complicated file I/O stack. The difference is simple: Persistent DRAM takes 20 nanoseconds to write a byte, while block I/O takes a microsecond or two with ultra-fast flash and 10 milliseconds with hard drives.
The quest is still on, but in the past two years, we have found a credible alternative. A nonvolatile dual inline memory module (NVDIMM) using flash memory goes some of the way toward speeding data access and becoming a persistent DRAM surrogate. There are currently two types of NVDIMM:
- NVDIMM-F is a flash-only DIMM that can drop into a memory slot in a server and act as a fast SSD. It uses block I/O, so transfers are done in 4 KB blocks through the file I/O system and a special driver. Even so, NVDIMM-Fs are faster than any other SSD or flash card, but considerably slower than DRAM.
- NVDIMM-N uses DRAM for normal transfers and only resorts to the flash element onboard the DIMM when power fails. This means transfers are done at CPU operation speeds, which cuts latency for an I/O operation into the nanosecond range. This is almost what persistent DRAM should be, but the limitations of DRAM power use and packaging causes capacities in the gigabyte range, not terabyte.
While either NVDIMM class would boost system performance dramatically, especially for in-memory databases, neither one can reach the Holy Grail standard.
However, 3D XPoint memory, a new technology co-announced by Intel and Micron, has substantially faster data access times than flash.
The SSD models fit into existing servers and storage easily, while the NVDIMM version has the same problem as today's NVDIMM-N: the compiler, link loader and operating system support aren't available yet. Intel has quite an edge with this, delivering one of the most popular NVDIMM compiler packages.
As an NVDIMM, 3D XPoint memory would have approximately 20% of the speed of standard volatile DRAM. This is fast enough that the two technologies could be intermixed in a server with the CPU memory cache masking the speed difference. But subtle software changes are needed to take advantage of this.
You would need a persistent memory class in the compiler, for example, and a way to easily guarantee atomicity in writing data. Also, applications would need to be very aware of the I/O differences. These changes are in the pipeline, and Intel and Micron are hopeful that they will arrive in conjunction with the delivery of the 3D XPoint memory.
An Optane upset in performance
Sadly, the controller for the memory used in Intel's 3D XPoint versions, named Optane, didn't meet its performance goals, delivering only the 4x to 10x speed boost seen by Micron in SSDs. The first-generation offering does not appear to be fast enough for NVDIMMs.
This speed problem appears to have shifted Intel's release of a second-generation, full-speed Optane product from 2016 into 2017 for SSDs and even into 2018 for NVDIMMs. The first-generation SSD will still be available in 2016, but it will not be dramatically faster than the NAND flash drives released at year end. Add in the expected production ramp up for the Optane product, and NVDIMM-N and NVDIMM-F look like they just received a reprieve.
Perhaps more importantly, Samsung and Western Digital's SanDisk business are not just standing still and waiting to be crushed. They have their own flash replacement products in the works, though they were quite a bit behind Intel's original schedule. This effective slip in the delivery schedule of Optane NVDIMMs means these competitors have some chance of catching up, which could open up a race for market share.
Competitor RRAM could horn its way in
In 2015, Western Digital's SanDisk unit announced that it had developed a storage-class resistive RAM (RRAM) product with Hewlett Packard Enterprise. This would provide stiff competition for 3D XPoint, with similar speeds and, perhaps, more durability, since first-generation 3D XPoint write wear numbers aren't much better than NAND. RRAM lends itself nicely to 3D structures and might achieve higher density than the "bulk cell" approach in 3D XPoint.
This race for the gold is a balance of hype, solving tough physics and engineering issues and creating production capability. 3D NAND foundry capacity is growing rapidly in China, with the new coarse/fine stacking approach of piling 48 layer structures on top of each other also increasing density by a big factor, while 3D XPoint is currently confined to the Intel/Micron joint venture foundry in Lehi, Utah. This suggests that the sweet spot for solid-state storage will continue to be 3D NAND for the next few years, and that the battle for the next product offering won't reach the market until late 2017 or 2018 -- and will then take several more years to settle. At that point, 3D XPoint memory will likely have serious competition, which is great for the user.
How big will 3D XPoint's impact be on flash storage?
Compare and contrast NVDIMM and memory channel storage
HPE launches persistent memory server products