GPU Computation and GPU TeraSort

There has been a lot of noise lately about the potential of using either GPUs or GPU-like streaming architectures for HPC. There’s lots of excellent coverage at the gpgpu.org news site – in particular see the Scientific Computing category.

We’ve been discussing the use of NVidia GPUs for general-purpose computation in the largescale seminar here at UCSD, and I thought I’d highlight a paper we read recently that goes into more than the usual detail about how an existing algorithm (bitonic sort) was adapted to get high performance on a GPU for sorting database keys. GPUTeraSort shows very high price-performance, using a GPU as a co-processor along with an inexpensive CPU.