Can LLMs write fast multi-GPU kernels? We built a benchmark to find out.