In an effort to speed up computations for research proposes academics and other scientific elites have turned to the processing power of graphics cards. GPUs may not hold the frequency speed records but they certainly have more math processing power than any single CPU today.
Organizations such as www.GPGPU.org have devoted their efforts and resources to speed things up in the fields of physics simulations, medical imaging, speech recognition, and matrix/vector operations like Fast Fourier Transform (FFT) and Ray Tracing. It has been shown that the Intel Pentium 4 3.0 GHz processors can produce 12 GFLOPs / 5.96 GByte per second. Interestingly, by using graphics cards they can produce much higher throughput. The ATI X1800XT can produce almost 7 times as much throughput with 83 GFLOPs / 42 GByte per second.
In a paper by Daniel Reiter Horn, Mike Houston, and Pat Hanrahan at Stanford University, they show results of CPUs as well as video cards using hidden Markov models (HMMs) to solve fuzzy protein sequence matching calculations. The paper states "...unlike the ATI hardware, the Nvdia hardware cannot directly use previous outputs as inputs for future iterations without a copy."
ATI added features like write to memory that allow it to skip a copy step that was needed to move the calculations forward to the next step. This one addition is one of many in the GPU's arsenal. In the diagram below you can see where the efficiencies of the ATI X1800XT have even pushed its performance to a new level.
It will be interesting to see where this science and academic community goes with ATI's new hardware (once it actually hits the market).
You can read the paper here.