GPUs surpass computers at repetitive calculations

Abstract

Major performance gains can be obtained by implementing computational intelligence algorithms on graphics processing units. However, a certain amount of skill is needed for the implementation; in some cases the performance gains can be as high as 200 times, but as low as two times or actually less than CPU operation. It is necessary to understand the limitations of the graphics processing hardware and to take these limitations into account in developing algorithms targeted at the GPU. It is important to note that all the examples in this document were operating on last-generation hardware. The next generation of graphics hardware is now available, and includes an order of magnitude more shader units per processor, as well as improved branching capabilities. Consider the possible capability of 512 programmable pipelines working in parallel at 1.5 GHz each, providing the amount of computing power previously seen in large supercomputers on a single desktop machine, at a fraction of the cost. These extremely powerful computing tools are now at the disposal of software designers for entertainment, scientific- computing, and computational intelligence applications.