As graphics processors continue to advance, their potential for use as general-purpose processors continues to increase. General-purpose GPU (GPGPU) programming is still a fairly new field, but with tools like nVidia's CUDA, it is finally becoming accessible and practical for real-world usage.
Multithreading is a great way to increase performance, given enough CPU cores; however, passing massively data-parallel computations to a GPU is even faster, when done properly. Significant performance bonuses are available when taking advantage of the full power of a GPU.
However, there are some caveats. If data needs to be transferred between the CPU and GPU often (e.g. every frame or three), the slow speed of the system bus can introduce a serious bottleneck, and eliminate the gains from using GPGPU techniques entirely. In such situations, GPGPU programming is not currently particularly beneficial.
Fortunately, there are still plenty of things to do on the GPU that don't require such frequent communication. For instance, it is now practical to perform complex tessellation or even procedural geometry generation directly on the GPU, avoiding the expense of transferring vertex buffers across the bus. (For a great demonstration of this, check out Jack Hoxley's journal here on GameDev.Net.)
Another area of interest to watch is the introduction of compute shaders in DirectX 11. Compute shaders promise the ability to perform complex computations on data that has passed through the graphics pipeline, opening up a vast new world of possibilities for performing heavy calculation on the GPU.
Since GPGPU development is still relatively nascent, the user APIs for tools like CUDA are still slightly slower than what the driver-level APIs can truly offer. In some cases this additional overhead may be prohibitive; but while it does limit the potential of GPGPU computation in some contexts, the gap should prove easy enough to close. In fact, given the pace of hardware development, it may be completely closed by the time current GPGPU code ships.
Overall, GPGPU technology promises to feature heavily in upcoming hardware, and upcoming software would do well to exploit it as much as possible. Watch for a host of tools for GPGPU programming to appear (or mature to the point of market-readiness) over the next couple of years.
Southwest Gaming Expo
