Unlocking ML Performance on TPUs and GPUs