In principle most if not all inference hardware should be usable for training.Efficiency is the question.