We should investigate how to enable fp16 e.g. in benchmarks. This [TensorFlow issue](https://github.com/tensorflow/benchmarks/issues/77) may provide some clues.