Orders of magnitude slower.

Isn't this GPU-to-CPU? And really slow. And only CUDA. And over IP. And implemented in software. I think it's really very different.