Kimi K3 生成 H100 CUDA 内核速度比优化 PyTorch 快 14.82 倍

Kimi K3 now writes a H100 CUDA kernel 14.82x faster than optimized PyTorch, nearly matching Claude Opus 4.8.

KernelBench tests hardware work, not just whether generated code successfully compiles.

KernelBench gives an AI agent a readable PyTorch implementation and asks it to produce a custom GPU implementation that is both correct and faster.

The code must compile, return the right numerical results across test inputs and beat a reference implementation on actual hardware. A kernel that is fast but wrong fails. A kernel that is correct but slow is not particularly useful.