high performance computing on graphics processing units: hgpu.org

We present a novel parallel implementation of N-body gravitational simulation. Our algorithm uses graphics hardware to accelerate local computation, and is optimized to account for low bandwidth between the CPU and the graphics card, as well as low bandwidth across the network. The number of bodies that can be simulated with our implementation is limited only by the memory of the graphics card, and results for small clusters indicate that it will scale well across larger numbers of nodes. Finally, we show that our algorithm significantly outperforms a comparable CPU implementation. Heretofore, commodity graphics hardware has been used mainly for graphics and visualization applications. This work shows that it can also be used effectively for scientific computation.