Can AI Agents Write CUDA Kernels That Beat PyTorch? A DGX Spark Benchmark
lmoroney · x · 2026-10-10
Chien Vu tested whether Claude Code agents can write CUDA kernels that beat PyTorch on an NVIDIA DGX Spark, covering softmax, layernorm, fused GELU with bias and residual, and matmul with ReLU. Every kernel passed strict correctness checks against an fp64 reference, and the writeup's most useful contribution is its discussion of proper measurement methodology for such benchmarks.
Related event: Claude Code Writes CUDA Kernels That Beat PyTorch in Tests(2 posts)→
More from coding & agent
- LangChain's Model Router Cut Median Coding-Task Cost 64% With No Quality Loss — hwchase17 · 2026-10-10
- Keenan Crane shows AI-generated shader with coarse-to-fine decomposition diagram — keenanisalive · 2026-10-10
- One-line agent skill taps Gaia DR3 to build 3D star maps from real astronomy data — TinfoilTricorn · 2026-10-10
- Dev uses Claude to build native Swift RealityKit 3D simulations, demos a saw-safety training scene — Scobleizer · 2026-10-10
- Weekly agent transcript reviews: mine your sessions into AGENTS.md to level up agents — JeremyNguyenPhD · 2026-10-10
- DuckDB v2.0 CLI agent mode cuts agent-read tokens by 59% on TPC-H benchmarks — josh_wills · 2026-10-10