Scaling TensorFlow and Caffe to 256 GPUs - IBM Systems Blog: In the Making