News for hackers. Stuff that matters.

A comment on Run CUDA, unmodified, on AMD GPUs. Parent comment

Re:Run CUDA, unmodified, on AMD GPUs by ashvardanian on Monday July 15, 2024 @08:57PM (#40972826)
Have you seen anyone productively using TMA on Nvidia or async instructions on AMD? I’m currently looking at a 60% throughput degradation for 2D inputs on H100: https://github.com/ashvardanian/scaling-democracy/blob/a8092...
[ Parent | on HN ]

There are no comments on this comment yet.