Speed in FP16 under k80 is almost as FP32, the architecture doesn't work well with FP16. As you said, you need Volta to notice improvements with FP16. The othe

Speed in FP16 under k80 is almost as FP32, the architecture doesn't work well with FP16. As you said, you need Volta to notice improvements with FP16. The other two main features are: Eager execution and TensorFlow Lite