[GitHub] [incubator-mxnet] ZHAIXINGZHAIYUE commented on issue #16531: Just use fp16 is slower than mixed_precision! Why?

ZHAIXINGZHAIYUE commented on issue #16531: Just use fp16 is slower than mixed_precision! Why? URL: https://github.com/apache/incubator-mxnet/issues/16531#issuecomment-547326899 it is influenced by kvstore(local or device). T

[GitHub] [incubator-mxnet] ZHAIXINGZHAIYUE commented on issue #16531: Just use fp16 is slower than mixed_precision! Why?

ZHAIXINGZHAIYUE commented on issue #16531: Just use fp16 is slower than mixed_precision! Why? URL: https://github.com/apache/incubator-mxnet/issues/16531#issuecomment-544313684 @anirudh2290 I mean the train speed. mixed_precision: 1000 samples/s, just single fp16: 900 samples/s -