TensorFlow has its framework XLA (https://www.tensorflow.org/versions/master/resources/xla_prerelease.html) to compile the pipeline. Perhaps you can have an accelerator.