You may not realize it, but that's a quite damning statement on the CUDA compiler.