bfloat16 training - AshokBhat/ml GitHub Wiki

About

Improvement in time-to-train with bfloat16

Model % Speedup
ResNet 32%
Inception v4 12%
Inception v3 21%

Source: Internal experiments at Google (https://cloud.google.com/tpu/docs/bfloat16).

See also