Skip to content

Comment on Quantization from the Ground Upparent

Comments

Your understanding is correct. The key detail is that the author used an M1 Max and H100 for their testing.

M1 Max: FP16 hardware support, FP8 and Bfloat16 emulated in software (via dequantization)

H100: FP16 and FP8 hardware support

which I ran both on a MacBook Pro M1 Max and a rented H100 SXM GPU
AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.