Comment by juancn

In principle most if not all inference hardware should be usable for training.

Efficiency is the question.