On Thu, Jul 25, 2013 at 8:05 AM, rjguan
We are wondering if the GPU computing will help significantly or not on phenix autosol, model building and refinement.
Not at all, especially since we don't have any code for this in place. GPUs really aren't very useful for macromolecular crystallography, unfortunately.
Has anyone used such machines and how is the experience? Would you like to share with us?
Several of us at LBL (Nick Sauter, Peter Zwart, Billy Poon, and myself) have done some experiments with the high-end NVidia chips. Basically they work very well for certain specialized tasks, which don't really apply to MX. I tried the simplest route, which was to call the CuFFT library, and it's about 20x faster than a single Intel CPU core, but once you add in the memory transfer it's only about 4x faster, and that's not really a fair comparison since the system had 1 GPU and 8 CPU cores. To really take advantage of the speedup we'd need to port other calculations to the GPU as well. I'm pretty certain we could make classical density modification (i.e. charge flipping) and maximum entropy calculations at least an order of magnitude faster on a GPU, but these aren't really rate-limiting steps in the overall process. Speeding up phenix.refine ultimately requires algorithmic improvements, not expensive proprietary hardware. -Nat