I am able to run OpenCL programs, including running them as MPI + OpenCL, but OpenCL is hopelessly ugly as a programming language for end users. It takes a 150 lines of extra junk to run a simple OpenCL kernel, whereas it takes only 4 lines to do it in CUDA (not that CUDA is a great language either). See attached example from the OpenCL spec. The dot product kernel is a few lines of code, followed by 3 pages of additional stuff needed just to run it. OpenCL examples fron NVIDIA are in ~thakur/OpenCl-1.0/OpenCL/src on breadboard. oclVectorAdd is the one I added some MPI too. Rajeev