Q. Do the PGI Accelerator compilers support using two or more GPUs in the same program?
A. Multiple GPUs are supported when each GPU is connected to a distinct host thread. Multiple GPUs could be used if the accelerator region were contained in an OpenMP parallel region, for instance. Implicit use of multiple GPUs adds another level of complexity, because not only are the GPU memories distinct from the host memory, they are separate from each other. Managing multiple GPUs with the PGI Accelerator model requires data distribution similar to that used in High Performance Fortran. We think this is important, but we haven’t come up with an elegant solution yet.