News for hackers. Stuff that matters.

A comment on Scalable MatMul-Free Language Modeling.

Re:Scalable MatMul-Free Language Modeling by PaulHoule on Sunday June 09, 2024 @07:50AM (#40623809)
This is why the NPU built into your processor could quickly become a liability instead of a benefit.
[ Parent | on HN ]
Re:Scalable MatMul-Free Language Modeling by sva_ on Sunday June 09, 2024 @08:15AM (#40623927)
It's ok, the AMD one still doesn't even work on Linux without a custom kernel
[ Parent | on HN ]
Re:Scalable MatMul-Free Language Modeling by imtringued on Monday June 10, 2024 @10:10AM (#40633873)
Most NPUs advertise TOPS when most models still run on floating point. The approach they outlined will also run faster on NPUs.
[ Parent | on HN ]