SanDisk ULLtra DIMMS: Terabytes of low-latency flash storage directly off the RAM channel

Diablo Technologies developed this new type of NAND memory design that eschewed PCIe slots or SATA interfaces in favor of attaching flash directly to the DIMM channel. This promised incredibly low latencies and consistent performance. SanDisk has taken an interest Diablo Technologies and has partnered with the company to release a shipping product.

Dubbed ULLtra DIMM (Ultra Low Latency), This company has already signed IBM to ship the new hardware in System x3850 and x3950 X6 servers, with up to 12.8TB of installed flash capacity. The only reason why IBM is adopting the modules, because of latency. XFlash hitting a write latency of 5-10 microseconds — The listed performance, per DIMM, is 1GB/sec read and 750MB/sec write. When it comes to NAND industry its far lower than anything else. 

High-frequency stock trading is the market of these DIMMS — SanDisk mentions transaction processing, virtual desktop interfaces, virtualization and cloud computing, but most of these workloads aren’t so latency critical as to demand 5-10 microsecond response times. On the other hand, HFT is a market where microseconds in response time really can make the difference between making and losing money on a trade. if demand scales up then there no reason to wait to implement the tech in those fields to justify production for other kinds of servers.

Flash grows up

Over the years there are multiple attempts to reduce NAND latency by marrying the NAND to SATA controllers, PCI-Express, SAS and now main memory. Before you or anyone starts jonesing this product in the consumer space, is that NAND is still orders of magnitude slower than conventional DRAM. A 5 microseconds latency is amazing for non-volatile storage, but RAM write latency is measured in nanoseconds — thousands of times faster.

If your are keeping terabytes of database information sitting in local main memory is indeed very expensive —  Intel’s upcoming Ivy Bridge v2 Xeons is a four-socket server with 24 DIMMs per socket and 16GB DIMMs would “only” allow for about 1.536TB of local memory. If you instead had 12.8TB of local NAND flash, that might change things somewhat. What we’re seeing here, at the high end, is NAND reaching up to reduce the impact of the access-time pyramid gap (shown below).

We might see this technology showing up in consumer hardware at any point in the next few years, but it’s not impossible. PCIe-based NAND storage has been inching towards the consumer market as conventional SSDs drop well below the $1/GB mark.