Скорость

NVIDIA GPUs can run operations in float16 faster than in float32
TPUs can run operations bfloat16 faster than in float32

The Keras mixed precision API allows you to use a mix of either float16 or bfloat16 with float32, to get the performance benefits from float16/bfloat16 and the numeric stability benefits from float32.

XLA: Optimizing Compiler for Machine Learning