Skip to content
Hacker News · AI·· 3 h agoAI score38

Riptide benchmarks its GPU attention kernels against AMD's own on MI350X, publishing every loss

Measured attention kernels vs. AMD's own on MI350X, with every loss published

AI summary

Riptide published measured benchmarks of its GPU attention kernels against AMD's own kernels on the AMD Instinct MI350X, reporting both wins and losses with 95% confidence intervals and float64 correctness checks. Its kernels reach 3.32× AMD's best at head dim 128 causal training with S=512, but fall to 0.82–0.84× at 16k tokens for head dim 128 bf16. The kernels are not open source and are offered to partners for evaluation and licensing.

Source:Hacker News · AI · github.com