KernelSkill: A Multi-Agent Framework for GPU Kernel Optimization

Researchers developed KernelSkill, a multi-agent framework that optimizes GPU kernel performance using expert knowledge rather than trial-and-error approaches. The system achieved 100% success rates and significant speedups (1.92x to 5.44x) over existing methods, addressing a critical bottleneck in AI system efficiency.