Where does these magic optimization flags come from? https://github.com/NVIDIA/DeepLearningExamples/issues/57