Helion lands in Hugging Face Kernels and beats PyTorch SDPA on H100 by 1.17×
RisingSayak · x · 2026-07-29
Helion is now supported in 🤗 Kernels, and an H100 benchmark says it beats PyTorch SDPA on attention.
- 19/19 tuned shapes were faster.
- On 10 unseen shapes, it was faster on 9/10.
- The reported geomean speedup was 1.17×.
- Numerical results matched on all 29/29 tested shapes.
The benchmark runs with one line of Python.
Related event: Helion Outperforms PyTorch SDPA on H100(2 posts)→
More from Infra
- Microsoft and Meta may lift CapEx, but the real question is whether AI bottlenecks remain physical — tengyanAI · 2026-07-29
- Goldman sees AI hardware stocks hit by forced selling as HBM valuations may reset higher — firstadopter · 2026-07-29
- MoonshotAI open-sources FlashKDA, a high-performance Kimi Delta Attention kernel — MoonshotAI · 2026-07-29
- SK Hynix’s $476,000 bonuses are fueling a Samsung AI chip talent war — eyishazyer · 2026-07-29
- AI companies are hiring thousands of electricians and carpenters — pstAsiatech · 2026-07-29
- Tencent Hunyuan open-sources AngelSpec, a full-stack speculative decoding framework — 腾讯混元 · 2026-07-29