Graphics Processing Units

Assuming GPU can positively affect prince/performance curve of HPC machines, what the best way to use it?



How do we offload compute kernels onto GPU?
How to abstract GPU from the user?
Same questions for FPGA's too.

To start with run NAS benchmarks on GPU.
Or run HPL on GPU.

Q. How to use thread contention in multi-threading MPI apps to offload data to GPU?