I wonder why there are no official numbers on RV770/870 GPGPU perf in various realworld test (such as SGEMM, DGEMM). After all, AMD positioning GPU as application accelerator. As I understand AMD ACML supports matrix multiplication on GPU but the perf is pretty low (one says 300 GFLOPS on RV770). There are some highly specialized algorithms to extract more flops, but it is not clear if those algorithms are universal enough to use it in the real world.




Reply With Quote

Bookmarks