Gemini 4 Carbon Performance on KernelBench-Mega

Gemini 4 Carbon high achieved notable performance on KernelBench-Mega by writing a specialized W4A16 Kimi-Linear Decode megakernel for RTX PRO 6000. According to the evaluation, it operated significantly faster than the optimized PyTorch baseline by utilizing persistent blocks, a single-word software barrier, and custom int4 GEMVs. While not exclusively a kernel writer, the model demonstrated strong practical performance and clarity when managing complex tasks. Open article →