THE AUSSIE Commonwealth Scientific and Industrial Research Organisation (CSIRO) has just knocked together a supercomputer that's based around a cluster of graphics processing units (GPUs).
Using GPUs in this way gives the computer a processing processing capacity that competes with supercomputers over twice its size.
The as yet unnamed CSIRO system is one of the first supercomputers to combine traditional CPUs with powerful GPUs. The fifth fastest supercomputer on the latest Top 500 supercomputers list includes ATI Radeon HD4870-2 GPUs.
The CSIRO machine has 28 Dual Xeon E5462 Compute Nodes for a total of 1024 2.8GHz compute cores. It has 500GB of SATA hard disk storage and DDR InfiniBand interconnects.
The system also has 64 Tesla S1070 modules which means 256 GPUs with a total of 61,440 streaming processor cores.
It can manage 200+ TeraFLOPS (TF), more than the 140 TF capable supercomputer announced by the Australian National University last week, although that is a 64-bit machine while the CSIRO machine is a 32-bit machine.
The CSIRO machine's GPUs can deliver up to 4.14 TF of single-precision floating-point performance and 345 GigaFLOPS of double-precision floating-point performance for each Tesla on board.
They can be accessed with Nvidia's CUDA parallel computing architecture or by using compiler technology released by the Portland Group .
CSIRO science applications have already seen one to two orders of magnitude increases of speed by using Nvidia GPUs. µ