MPS implementations generally lag behind CUDA kernels, especially for new and cutting edge stuff. Sure, if you're only running CPU inference or only want to use

MPS implementations generally lag behind CUDA kernels, especially for new and cutting edge stuff. Sure, if you're only running CPU inference or only want to use the GPU for simple or well established models, then things have gotten to the point where you can almost get the plug and play experience on Apple silicon. But if you're doing research level stuff and training your own models, the hassle is just not worth it once you see how convenient ML has become in the cloud. Especially since you don't really want to store large training datasets locally anyways.