Float16 mixed precision being slower than regular float32, keras, tensorflow 2.0...

tf2.4 mixed_precision with float16 return 0 gradient...

Can I speed up inference in PyTorch using autocast (automatic mixed precision)?...