FP8 Is All You Need (Part 1): Debunking Hardware FP64 as the HPC Holy Grail

I feel like there's gotta be a catch in there somewhere, like "fails catastrophically when computing logarithms" or something. skeptical

Assuming this is correct that is a very intriguing result.

FP64 emulated with FP8 running faster than the native FP64 implementation.