OptiML: An End-to-End Framework for Program Synthesis and CUDA Kernel Optimization

Researchers unveiled OptiML, an AI framework that automatically generates and optimizes high-performance CUDA kernels for GPUs. It uses a two-stage process: first generating code from natural language prompts, then refining it using Monte Carlo Tree Search guided by hardware feedback. The system consistently discovers verified performance improvements over strong LLM baselines, producing interpretable optimization trajectories grounded in profiler evidence. It navigates the combinatorial space of low-level transformations that typically challenges developers.