CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning

The exponential growth in demand for GPU computing resources has created an urgent need for automated CUDA optimization strategies. While recent advances in LLMs show promise for code generation, current SOTA models achieve low success rates in improving CUDA speed. In this paper, we introduce CUDA-L1, an automated reinforcement learning framework for CUDA optimization that employs a novel contrastive RL algorithm. CUDA-L1 achieves significant performance improvements on the CUDA optimization task: trained on NVIDIA A100, it delivers an average speedup of x3.12 with a median speedup of x1.42 across all 250 CUDA kernels of KernelBench, with peak speedups reaching x120. Furthermore, the model also demonstrates portability across GPU architectures, achieving average speedups of x3.12 on L40, x2.50 on RTX 3090, x2.39 on H100, and x2.37 on H20 despite being optimized specifically for A100. The capabilities of CUDA-L1 demonstrate that, RL can transform an initially poor-performing LLM into an effective CUDA optimizer through speedup-based reward signals alone, without human expertise or domain knowledge. This paradigm opens possibilities for automated optimization of CUDA operations, and holds promise to substantially promote GPU efficiency and alleviate the rising pressure on GPU computing resources. We also identify important challenges posed by training RL models for tasks like CUDA development, where RL often learns to exploit loopholes in reward functions rather than solve the intended optimization problems. By identifying these failure modes and analyzing their root causes, we develop practical methods for creating more robust training procedures that prevent reward hacking.