Databricks tops NVIDIA's SOL-ExecBench by looping GPT-6 and Opus to write CUDA kernels
Yuchenj_UW · x · 2026-09-29
Databricks announced it now ranks #1 on NVIDIA's SOL-ExecBench kernel leaderboard across all 4 tracks. They put GPT-6 Astra and Opus 5 in a self-hillclimbing loop where the agents researched, implemented, verified and iterated on performance-sensitive CUDA kernels — and beat top GPU kernel engineers.
Key facts:
- Total token spend: $70K, far less than what top kernel engineers get paid.
- One learning: GPT-6 Astra and Opus 5 are still much better than OSS models at writing GPU kernels.
- Built on collaborations including NVIDIA's open-source KDA (Kernel Design Agents) (1.1k stars on GitHub), an agent-centric workflow for high-performance CUDA kernel tasks with prompt templates, skills and docs, plus the Humanize project.
- The effort was driven by @leshenj15 at Databricks, who built the kernel harness; more details to come.
Related event: Databricks Tops NVIDIA Kernel Leaderboard with AI Agents(2 posts)→
More from coding & agent
- Google Research open-sources PaperOrchestra, a multi-agent framework for writing AI papers — google-research · 2026-11-16
- ffmpeg-skill gives coding agents local video editing via 42 FFmpeg tools and MCP — kajisho5 · 2026-11-16
- Every's Frontier team on Claude Fable 5.2: overnight agent jobs, cost tradeoffs, real builds — every · 2026-11-16
- Agents Now Reply to Slack Messages on Their Own, Resolving a Prod Incident Ghost-to-Ghost — iamrobotbear · 2026-11-16
- Engineer's Interview Prep Method: Use Claude to Find Which Answers Undersell Your AI Work — brandon_galang · 2026-11-16
- DeepSeek open-sources its app frontend "chat-demo", already at 3.9k stars — NielsRogge · 2026-11-16