65fc9fb4897a89789352e211ca2d398f-AuthorFeedback.pdf

Author response: The speed-up from FP32 to FP8 strongly depends on the chip architecture and any additional10 compiler andsoftwareoptimizations. Detailed comments: there should be better baseline comparisons (although: this method seems to match normal16 training, so there'svery little margin for ittobe out-performed.