NVIDIA looks back at its agent loop that auto-generates GPU kernels with DeepSeek-R1
bingxu_ · x · 2026-08-21
NVIDIA engineer Bing Xu reflects on the technical blog he co-authored with Terry Chen, started in 2024 and published in 2025: a closed-loop agent workflow combining a verifier with DeepSeek-R1 plus inference-time scaling to automatically generate numerically correct, optimized GPU attention kernels for different attention variants—without explicit programming.
- Key results: on Stanford's KernelBench, 100% of Level-1 and 96% of Level-2 problems yielded numerically correct kernels.
- The blog introduced the idea of a "loop" in agents and was initially criticized as "just a loop"; looking back, the author notes it's striking that even a Llama-70b-class model could generate GPU kernels.
- An early representative application of test-time scaling to real engineering tasks.
More from coding & agent
- Devin Oncall tops SREGym with 100% end-to-end success, auto-resolving production incidents — tianyin_xu · 2026-10-16
- Running a multi-model UI design arena: Paseo orchestrates four design drafts per model — vista8 · 2026-10-16
- Cortex A/B-tests coding-agent skills and rules by replaying developer corrections before keeping them — Samuele Giampieri · 2026-10-16
- Tool calls are the hidden bottleneck killing AI coding speed — Crabbox and 5 fixes — steipete · 2026-10-16
- Agentic sandbox WorldBox doubles GPT-6 Astra's MazeBench score, 23 to 61 gems — ZhitingHu · 2026-10-16
- Two prompts: Gemini 4 Argon turns a single-player game into full multiplayer — chongdashu · 2026-10-16