Skip to main content

Supermicro shows off GPU supercomputing gear

While high-end graphics cards usually get their workouts playing a rousing game of Crysis, there's a growing market for supercomputers based on General Purpose GPU computing - and Supermicro is at Computex to show off its take on the concept.

The Supermicro GPU SuperBlade is an impressive name for a product, but the specifications are more impressive still: 120 CPUs and 120 GPUs, packed into a single 42U rack for high-performance number crunching at a surprisingly low power draw.

Based on Nvidia's high-end Tesla M2070 GPU - with the Tesla M2050 as a cheaper option for those trying to supercompute on a budget - the SuperBlade offers impressive compute power in an unexpectedly small footprint for those willing to investigate the possibilities of Nvidia's CUDA or Microsoft's DirectCompute.

As well as the GPUs, Supermicro's creation includes Intel's Xeon 5600 processors - or the 5500 series for those trimming costs - and up to 96GB of DDR3 memory in each node. Inter-node communication takes place over either a pair of QDR 40Gb/s InfiniBand connections, or 10GB Ethernet for the traditionalists - and a SATA DOM, similar to those created by InnoDisk, provides the boot partition.

GPU-based supercomputing is becoming increasingly popular, with many computationally-intensive tasks being well suited to running on the massively parallel architecture of a graphics processing unit - but competing products from companies such as Tilera promise to shake up the high-performance market, meaning Nvidia's current reign as the champion of low-power HPC could be short-lived.

As for the price? If you have to ask, you're probably not in Supermicro's target demographic.