Houmao agents auto-generate GPU kernels, beat hand-written CUDA by 1101x

A new arXiv paper from Mao Luo and colleagues demonstrates that general-purpose code agents can autonomously generate state-of-the-art GPU kernels without any manually written CUDA. The team built Houmao, a multi-agent orchestration framework that coordinates heterogeneous coding agents to generate, debug, profile, and optimize kernels. Humans only defined the workflow, enforced correctness constraints, and supplied references—never reviewing or editing the actual kernel code. The experimentation used FlashInfer-Bench workloads—Fused MoE, DSA TopK Indexer, and DSA Sparse Attention—evaluated on NVIDIA B200 GPUs under a correctness-gated protocol.

Across roughly 1.9 billion agent tokens, the generated kernels achieved remarkable speedups over PyTorch baselines: 92.68x on Fused MoE, 1101.02x on DSA TopK Indexer, and 181.35x on DSA Sparse Attention. They also significantly outperformed the hand-optimized FlashInfer baselines. In the official MLSys 2026 FlashInfer AI Kernel Generation Contest, Houmao's generated Fused MoE kernel scored a 1.71x speedup over FlashInfer—beating the top result from the agent-assisted track (1.68x). This signals that disciplined, correctness-first AI workflows can match or exceed human expertise in low-level GPU programming.