https://pytorch.org/docs/stable/notes/mps.html

In my very limited experience it's not a ton faster than using a CPU backend though.