Skip to content

Comment on Scalable MatMul-Free Language Modeling

Comments

This is why the NPU built into your processor could quickly become a liability instead of a benefit.

Most NPUs advertise TOPS when most models still run on floating point. The approach they outlined will also run faster on NPUs.

It's ok, the AMD one still doesn't even work on Linux without a custom kernel

AboutSource Built by g1lg1l

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