https://openreview.net/forum?id=HJe88xBKPr¬eId=Hyx3iPknoH
Mixed Precision Training With 8-bit Floating Point | OpenReview
We demonstrated state-of-the-art training results using 8-bit floating point representation, across Resnet, GNMT, Transformer.
8 bit floating pointmixed precisiontrainingopenreview