Best Laid Schemes
The whole idea that GPUs can be used for more than just graphics processing has been given the label GPGPU, General Purpose Graphics Processing Unit. The most public application of GPGPU technology thus far is the Folding@Home campaign, launched in September of 2006. However, GPUs have been used in scientific and medical applications for quite a few years now. NVIDIA’s Tesla GPU Computing Solutions are the first mainstream hardware products devoted to GPGPU work. Tesla GPUs should be pumping out calculations at previously unheard of rates for far more companies and individuals than ever before. Now supercomputer level floating point performance is available to the masses, rather than the few who are fortunate enough to run their applications on current real supercomputers. Scientific calculations and highly-complex 3D rendering will probably be the forte of Tesla GPUs for the time being. Unfortunately for those holding out for more useful applications of the GPGPU concept, you might have a bit of a wait ahead of you, still.

It is true that GPU-based physics processing is right around the corner. However, GPU physics was announced originally by Havok with their Havok FX engine that had been in development since 2005. In March of 2006 at the Game Developers Conference in San Francisco, NVIDIA in a joint announcement with Havok announced that GPU-based physics would be possible on SLI GeForce hardware in the near future. Needless to say more than a year has past without much news on the topic beyond a couple of NVIDIA tech demonstrations. Despite Havok’s attempts at creating a middleware for physics computing on a GPU, the technology we were excited for more than a year ago still is not here.

After reviewing several white papers and technology presentations from the GPGPU organization, it becomes quite apparent that the main thing holding back the development of more applications for GPUs is the difficulty of programming such applications. Since standard code based on the x86 instruction set is not (yet) able to be run on GPUs, any developer making an application for the GPU needs to do so from the ground up. Unfortunately, nobody at FPS Labs has the faintest knowledge of programming beyond html and php, so we can’t really elaborate on the troubles that developers face when creating applications to run on GPUs. What we can say, however, is that there are steps being taken by graphics card manufacturers such as NVIDIA, who have developed something called CUDA, which includes a C-compiler development environment that should aid in coding applications to run on GPUs.

Beyond physics, the only other really promising application of GPGPU technology for gamers is AI calculations. With games becoming increasingly more complex and ‘smart’, AI is getting more and more complex. The calculations necessary to run the AI are becoming a burden for even the fastest CPUs. Since AI calculations are basically floating point algorithms, GPUs would potentially handle such a load much more easily than current hardware.