I’d argue you don’t need CUDA if PyTorch runs well on your platform, which is what everyone is trying to show with these MLPerf results. I think Intel’s strateg

I’d argue you don’t need CUDA if PyTorch runs well on your platform, which is what everyone is trying to show with these MLPerf results. I think Intel’s strategy with oneAPI is not bad, it’s just late.