How to scale distributed deep learning?