Stanford's ParallelKernelBench Shows Top LLMs Get Only 32% on Multi-GPU Kernels

Stanford and UCSD released ParallelKernelBench (PKB), the first benchmark evaluating LLMs' ability to write multi-GPU high-performance kernels, accepted as a NeurIPS 2026 poster. The best model achieved only 32% zero-shot correctness, exposing a major gap in multi-GPU code generation.

2026-11-04 ~ 2026-11-05 · 2 related posts

• Stanford's ParallelKernelBench lands NeurIPS 2026: LLMs nail only 32% of multi-GPU kernels — simonguozirui · 2026-11-04
• Stanford's ParallelKernelBench: best LLM solves only 32% of multi-GPU kernel tasks — burny_tech · 2026-11-05