Scaling out with distributed TensorFlow