HANDS ON GUIDE

2 - Graphics cards are getting there gradually

Older hardware can run GPGPU, too

GPGPU may have only come to popular attention with new graphics cards such as the GeForce 8800 and Radeon X1000 and HD 2000s, but much older graphics cards could theoretically be used. Nvidia GPUs from NV20 onwards (GeForce3 and GeForce4 non-MX) have the potential to be used for GPGPU, because their vertex calculation capabilities can be accessed for purposes other than Direct3D acceleration. However, this generation of cards only offers FP16 precision, limiting its usefulness. The NV30 generation (GeForce FX 5200 or later) and ATi’s R300 (Radeon 9700) offer FP32 and FP24 precision respectively, plus Pixel and Vertex shader 2 capabilities, yielding much more complex programmability and greater accuracy.

However, what has really allowed GPGPU to take off is the arrival of a unified shader architecture. Instead of dividing processing rigidly between vertex and pixel units, a unified shader architecture uses more general-purpose stream processors, incorporating ALUs, which can process either type of calculation (and with DirectX 10, the new geometry shader instructions, too). Our DirectX 10 feature explains this in more detail.

What makes GPUs so much faster than CPUs is their parallelism. Whereas your average Core 2 can execute a few SIMD instructions per clock cycle, the latest GPUs handle hundreds. ATi’s Radeon X1900 took parallelism to a new level with its 48 pixel shading units. It is this hardware which forms the basis of ATi’s Stream Processor, partnered with 1GB of GDDR3 DRAM – the eight vertex shading units are still present, but basically remain idle. But Nvidia’s G80 (GeForce 8800) has 128 stream processors and ATi’s R600 (Radeon HD 2900XT) 320 stream processors, so the parallel processing potential for both is huge.

Unfortunately, the current generation of graphics lack one important feature compared to dedicated floating point accelerators used in supercomputers. Where the latter offer FP64 (double-precision) processing, ATi and Nvidia still use FP32 (single-precision), although Nvidia claims it will have FP64 hardware by the end of 2007.

Latest Hands on guides

These are the most popular features articles over the past three months.

Mobile Broadband

Compare prices

Fastest, cheapest 3G mobile broadband dongles from 3, Vodafone, T-Mobile and Orange
from just £10/month