Mr. Kevin Collville
Tel: 021 658 3962
E-mail:kcolville@csir.co.za

Novel specialised computing hardware used in commodity entertainment provides greater potential for the acceleration of parallel algorithms than commodity general-purpose computer hardware. The graphics processor units (GPUs) used for computer and console games contain more transistors than the central processing unit (CPU) of the average computer. These GPU transistors are, unlike in CPUs, used to provide many arithmetic logic units (ALUs) which allow deep parallelism to be achieved at the heart of a simulation code. A GPU will have 32 to 128 ALUs, while a dual-core CPU has 4 to 8. Using streaming multimedia instructions, a CPU can operate on 4 single-precision pairs simultaneously, while a GPU can work with 32. This greater parallelism can accomplish speed-up factors for 20 or more once codes are rewritten to take advantage of this novel hardware.

A video game is a simulation, albeit an inaccurate or fantastic one, and is based on the same computer arithmetic (addition, multiplication) as serious scientific and engineering simulations. Driven by the need to paint ever more photo-realistic imagery and handle more complex scenarios, the hardware used by the gaming industry has accelerated past the conventional.

Recent developments in GPUs make them more attractive for high performance computing (HPC): the improvement in floating point arithmetic (more accuracy and stability by implementing true rounding, and other features of the IEEE/ISO standard) and the availability of a general-purpose programming environment and language (Cuda from Nvidia).

The HECTec research group at the CHPC aims to utilise the power of GPUs, for example the Nvidia's 9000 series, or IBM's Cell Broadband engine, to accelerate appropriate scientific and engineering simulation codes, as well as general purpose numerical routine libraries.