I realize the list is large and I don't want to clutter inboxes, but anyone who may be lightly considering GPU implementations should understand that this technology, while useful for some applications, is probably not economical for crystallography refinements. I do not have personal programming experience with them, but the AMBER community has found that it takes a very knowledgable programmer and a highly regular problem to reap the most significant benefits from this technology. It is a significant commitment of time to program a GPU code, both in terms of the language learning curve and the extensive reoptimization of the CPU implementation.
Multi-dimensional FFTs really are not the forte of GPUs, and in general conditional statements must be avoided as much as possible. Random memory lookup is also problem for GPUs, so these cryo-EM docking problems, while they may seem like CPU-intensive and regular/uniform problems, may not be so much suited for GPUs as for grid computing. Intel and other companies are coming out with massively multicore chips that have their own drawbacks but at least run commodity code. The benefits of GPUs to crystallography have been, and will probably continue to be, smoother visualization. Dave P.S. I agree with the consensus on this thread, lossy approximations to FFTs are not of use to crystallographers but I they may be of use to someone like me if I go back into docking and in-silico scanning applications :-)
