For some reason, frequency of M1 Max gpu is low - 400HZ instead of maximum possible ~1300HZ. In this post details are presented.
Turns out, PyTorch could be 3x times faster if it would just use a boosted GPU frequency.
What could be the reason of this? Maybe they just don’t use Metal API properly? In comparison, MLX runs at boosted frequency and 3x faster
This is a repo with code to reproduce:
You can also just use asitop to see frequency of GPU while using PyTorch on MPS