When you write for PyTorch and make it go fast it's easy to to optimize the high level model and the architecture towards CUDA and Nvidia chips without even kn

When you write for PyTorch and make it go fast it's easy to to optimize the high level model and the architecture towards CUDA and Nvidia chips without even knowing it. Nvidia has made the choice where everything works with CUDA and is optimized to run CUDA fast. It also restricts their design choices. Unless AMD and Intel make similar commitments, they are always little behind in performance and reliability.