Benchmark report backs Helion’s H100 attention speedup over PyTorch SDPA

RisingSayak · x · 2026-07-29

A GitHub Gist provides the benchmark report behind the Helion vs. PyTorch SDPA claim on H100.

It documents the test setup and reiterates the headline result: Helion attention kernels were faster on all tuned shapes, faster on 9/10 unseen shapes, and numerically matched PyTorch on 29/29 tests.

Related event: Helion Outperforms PyTorch SDPA on H100(2 posts)→

Original post →

More from Infra

Infra channel →