Diablo’s Memory1 technology leverage its unique memory-channel architecture, to get DRAM performance out of NAND flash and reduce server CAPEX costs by up to 75 per cent.
Ottawa-based memory vendor Diablo Technologies has announced its Memory1 technology, all-flash DDR4 system memory modules designed to provide the same performance as high end DRAM [dynamic random-access memory] for a small fraction of the price.
“DRAM demand is growing driven by virtualization, and it is big in in-memory data bases, as new applications like HANA and Hadoop drive demand for memory and drive power demands as well,” said Kevin Wagner, VP Marketing at Diablo. “The challenge is that it is very expensive – 30 times as expensive as any other memory technology.”
With Memory1, what Diablo has done is leverage its unique memory-channel architecture, which lets flash interface directly onto the high performance, highly parallel DDR bus. It uses NAND flash to get standard technologies to behave like DRAM.
“Memory1 fits in the thermals of current servers,” Wagner said. “There is no DRAM on Memory1 itself. Using our MCS [Memory Channel Storage] architecture and our software layer, it becomes one pool of system memory that behaves like DRAM – but with the economics and capacity of flash.”
Wagner said that it is necessary to have some true DRAM in the system along with Memory1, with the amount depending on the specific workload. He said an 8-1 or 10-1 ratio of flash to DRAM is typical.
Even with a little DRAM, Memory1’s economics of flash are much better than DRAM,” Wagner said. As well, DRAM is more expensive with size. A 32 GB stick is more than twice as much as a 16 GB stick. That’s not the case with flash, where there is no premium for density.”
Memory1 also allows more memory per server, which is particularly important in caching applications.
“Memcached performance is dictated by performance of the network, but in practice it leads to bottlenecks,” said Jerome McFarland, Principal Product Marketer at Diablo. “Memory1 allows more memory in each server so you can cache more data per server and minimize network access.”
McFarland further illustrated the economics of Memory1 with Internet search as practiced by hyperscale vendors and large enterprises.
“As indexes grow, because of the huge amount of data, they overgrow their DRAM, so they buy more servers to get more memory slots,” he said. “Having more memory per server with Memory1 lets them scale much more efficiently. It reduces servers by 90 per cent. Now the cost of Memory1, which is more expensive than flash, does make each server more expensive. But with a hyperscale user using 100,000 servers, reducing it to 10,000 still cuts their CAPEX by 75 per cent.”
While the greatest savings from Memory1 would come from users with the largest deployments, McFarland said this is not a technology which becomes uneconomic unless a certain volume threshold is reached.
“Even at the smaller scale it makes sense, with an in-memory database,” he said. “You don’t need 100,000 servers in a HANA deployment. It can be one server. But you will still save money on that one server for this kind of application by replacing expensive DRAM. This addresses the entire high-value market, as well as enterprise memory with HANA and enterprise customers looking for increased performance.”
Still, the initial go to market will focus on delivering Memory1 to the cloud and hyperscale data centers, which will mean that Diablo will be selling primarily to the ODMs and OEMS who serve those markets. Some SIs and ISVs will also be touched by Memory1. Diablo does say that it will be broadly available next quarter however. Pricing has not yet been announced.