Intel’s new Xeon E5 series will be powering plenty of application servers in the near future. But at CeBIT, high-end hardware company SuperMicro has demonstrated another use for the platform: as the basis of a GPU-based supercomputer.
The catchily named SuperServer SYS-7047GR-TRF is a 4U system based on Intel’s Xeon E5-2600, mounted on SuperMicro’s C602-based X9DRG-QF motherboard. This supports four PCI-E 3.0 x16 slots, each supporting the chipset’s full QPI bandwidth of 8GT/sec – effectively, twice the speed of a PCI-E 2.0 x16 slot.
SuperMicro has filled each of these slots with one of Nvidia’s 512-core Tesla M2090 cards. For single precision calculations, each card is rated at 1.3 teraflops, delivering an aggregate of 5.2 teraflops of computing power (2.6 teraflops for double precision).
On paper this may not sound exceptional: just one of AMD’s domestic Radeon HD 7970 cards is good for 3.8 single-precision teraflops. But while AMD’s card comes with 3GB of onboard memory, Nvidia’s Tesla cards provide 6GB each, providing far more breathing space for computations.
What’s more, these cards are based on Nvidia’s Fermi architecture, meaning they can be programmed in more or less native C++ (with CUDA extensions), helping them slot neatly into a programming workflow.
The cards are powered by twin, redundant, hot-swappable 1.6kW power supplies with a remarkable efficiency rating of 94%. The GPUs use passive heatsinks, kept cool by four chunky case fans, of the sort we’re used to seeing in rack servers. There’s eight drive bays, and support for up to 512GB of RAM across 16 DIMM sockets, scattered around the board so as not to impede airflow.
Clearly, this type of specialist, enterprise-grade kit is out of most hobbyists’ price range: BroadBerry currently sells Tesla-based systems built on last year’s Xeon 5600 series for £5,400 and up, and SuperMicro’s new design will doubtless cost at least as much. But it points to the growing commoditisation of GPU computing – surely a sign of the times.
