> but it’s light years behind on compute.

Is that the only factor though? I wonder if pytorch is lacking optimization for the MPS backend.