On Tue, Jan 24, 2012 at 1:38 AM, Adam Ralph <adam.ra...@nuim.ie> wrote: > CUDA is a set of extensions for C which will allow you to access hardware > accelerators (certain NVidia cards in this case). CUDA has been around for > a > while and there are CUDA libraries for FFT and BLAS. > I have not used cuFFT myself, I know that its APIs are based on those of > FFTW. The capabilities and ease of use of these cards are improving with > each generation. If you are in the game of speeding up your FFTs then I > recommend you take a look.
Unfortunately this isn't going to make refinement programs much faster either. I found that cuFFT was about 20x faster on a state-of-the-art NVidia accelerator versus a single Intel Xeon core - but the memory transfer knocks it down to 4x. OpenMP parallelization can give a similar speedup without spending $2500 extra on the GPU, and with much less butchering of the code. (And even that doesn't help much, because FFTs still take up less than half the runtime during refinement, at least in Phenix - I would be surprised if other programs were significantly different in this respect.) -Nat
