KernelBench (arXiv 2502.10517) scores a model on replacing PyTorch operators with GPU kernels that are both correct and faster than the reference, with a public leaderboard maintained by Stanford’s Scaling Intelligence Lab. In this corpus it is the leaderboard that GPT-5.5 xhigh topped at 6.57%, the item the newsletter reads as a model optimizing the hardware it runs on; a megakernel variant, KernelBench-Mega, appears in the seventeen-leaders item of July.
Benchmark
KernelBench
A benchmark from Stanford's Scaling Intelligence Lab that asks models to write correct, faster-than-baseline GPU kernels for PyTorch reference workloads.
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to April 29, 2026 (theinnermostloop.substack.com)
Development timeline
Newest reporting dates first. Reporting dates may differ from event dates.
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.