Wanted to share something I’ve made recently - a visualisation for ML number formats: thecharlieblake.co.uk/visual……
Hopefully this gives some intuition for the trade-offs involved (e.g. FP16 vs BF16). It also introduces the two newly proposed FP8 formats. Feedback much appreciated🙏
Oct 17, 2022 · 7:42 AM UTC
Just wanted to tag a couple of people on my feed who might be interested: @charles_irl - did some similar stuff a few months back which got me into all this! and @Tim_Dettmers - whose LLM.int8() paper shows some v interesting aspects of using low-precision formats