F@H and Mathematical Processing Performance
Video cards do so much more than just graphical processing and 3D rendering. They have become an incredible source of mathematical processing power. Applications like Folding @ Home, media encoding, and even games can benefit from this untapped resource.
As most people active in the F@H scene know, it has been speculated that CUDA and a lack of ATI optimizations in the F@H client have put NVIDIA in the lead. Whatever the reason, NVIDIA is vastly ahead as the 9800GTX+ chews through work units while the HD4850 isn’t quite as productive. We hope that Stanford will further optimize its cores to better utilize ATI hardware.
Although F@H definitely prefers NVIDIA CUDA supporting hardware, Sisoft Sandra’s synthetic GPGPU benchmark paints a very different picture. The HD4850 outperforms the 9800GTX+ by a significant margin here, especially in “Native Double Shader” operations. Although these numbers are completely synthetic, this should be somewhat reassuring for those frustrated with their F@H performance on ATI hardware. Clearly the processing power is there; harnessing it is the problem.
When it comes to memory bandwidth in Sisoft Sandra’s GPGPU test, the 9800GTX+ is the clear leader in both internal memory bandwidth and data transfer bandwidth.
Math and physics processing on GPUs is still in its infancy and what we have seen today is likely just the tip of the iceberg. In the next year or two, we expect to see applications make much better use of this untapped well of potential.