AMD/NVDIA or some other vendors provide GPU-accelerated BLAS and FFT. The program needs minimal changes to work on cluster with mixed CPU/GPU. This happens to the TOP500 benchmarking since HPL uses MPI/BLAS.

If the core computation is not based on BLAS or FFT, the user needs to mix MPI with CUDA, stream SDK or OpenCL to gain full power of GPU on such cluster. This is not easy job for large applications where there are lots of no-so-hot hot spots. For applications with fewer hot spots the GPU/CPU hybrid cluster is definitely a must buy.

I think some HPC vendors will fear very much about this trend, but it seems inevitable that the time of cheaper super computers is coming.

Therefore buy blah or blah, and sell blah, blah and blah!