GPUBench is a benchmark suite designed to analyze the performance of programmable graphics processors in areas of particular importance to general purpose computation. Included in the suite are tests that rigorously exercise a system's GPU, determining statistics such as memory input/output bandwidth to floating point buffers, texture cache bandwidth, data download and readback rates, instruction throughput, and instruction precision. More so than being a measuring stick for GPU performance, GPUBench is intended to be a tool for developers of GPU-accelerated applications, since a more complete understanding of machine capabilities will dictate better software design decisions. Information included in GPUBench reports is useful in accurately estimating the performance of shader programs as well as determining if a particular GPU is a viable target platform for a particular computation.
A poster on GPUBench was presented at the 2004 GP^2 Workshop in Los Angeles.
The full GPUBench source code can be obtained at: http://sourceforge.net/projects/gpubench.
GPUBench was developed at the Stanford University Graphics Lab.