How to make the right hardware choice for deep learning training and inference applications