Finishing our benchmark for GPU accelerated compute. The state of sparse matrix libraries on GPU seems to be a complete mess. Running cuSOLVER currently, but it's incredibly primitive.