voir sur phys.org : Distributing control of deep learning training delivers 10x performance improvement