Claude Code Writes CUDA Kernels That Beat PyTorch in Tests

Chien Vu tested Claude Code on an NVIDIA DGX Spark, having the agent write CUDA kernels across four operators. The results outperformed PyTorch in some cases, including a fused GELU kernel about 2.5x faster, with the agent inventing its own fp32 splitting trick.

2026-10-10 ~ 2026-10-10 · 2 related posts

• Claude Code agents beat torch.compile on CUDA kernels: 2.5x fused GELU, 1.57x matmul via fp32-splitting trick — lmoroney · 2026-10-10
• Can AI Agents Write CUDA Kernels That Beat PyTorch? A DGX Spark Benchmark — lmoroney · 2026-10-10