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
- Agent Loop is the engine, outcome ownership is the product — dfinke · 2026-08-21
- Remote Co-founder Builds Agent Icarus to Automate Tacit Company Knowledge — dscape · 2026-08-21
- Agent Design: Capabilities Should Be Tools, Domains Should Be Agents — andreisavu · 2026-08-21
- Long-Horizon Agents: A Technical Primer on METR Doubling Trends — agihouse_org · 2026-08-21
- Harrison Chase shares strategies to harden Agent evals for better signals — hwchase17 · 2026-08-21
- mcp-anything: Index 75k MCP Servers and Let Models Discover and Call Any Tool — Outcome_Hour · 2026-08-21