afaik Xeon phi doesn't come close to the performance per buck of gpus in deep learning