Has your server's memory reached its 192GB limit? Last week a company called Netlist was demonstrating a new RDIMM technology that it claims can double per-CPU memory capacity without doubling the cost and without replacing any other hardware. At Supercomputing 09, the high performance computing industry's annual conference, Netlist showed off its so-called HyperCloud 16GB DRR3 RDIMMs elevating a machine to 384GB of memory per CPU.
Set to begin shipping in the first quarter of 2010, HyperCloud RDIMMs also will be available in 4GB and 8GB versions, and according to Paul Duran, director of business development for Netlist, work interchangeably with today's standard parts. "They are essentially plug-and-play with the JEDEC standard with no issues," even when they are installed on the same channel, and require no bios changes, he added.
The technology employs proprietary logic that Duran said appears to the CPU as two physical ranks of memory (the maximum), but actually contains four, allowing dual socket servers to be populated with as many as 24 16GB RDIMMs. Duran said that vRank also helps performance. "The CPU is tricked into thinking it has a single load, therefore it will run at 1,333 [million transfers per second] when you have four DIMMs populated per channel." In today's systems, the CPU supports 1333 MT/s only when populated with one DIMM per channel, he said. With two, it decreases to 1067 MT/s and with three it drops to 800. Doing the job are two ASICs and a register device with the rank multiplication logic. Isolation circuits between each DRAM and the bus connector provide the load reduction.
While exact pricing was not disclosed, Duran said HyperCloud's cost-per-gigabyte will be close to that of other products--which vary depending on the DRAM spot price--plus "a 20 percent premium for this technology."