- Test with MPI parallelizing across GPUs rather than OpenMP (but use the OpenMP target regions) - Try with mixture of MPI and OpenMP