Cake from Scratch: Agents Write GPU Kernels (gangliao.me)

NVIDIA and Carnegie Mellon University have introduced CAKE (Compiler-Agent co-design for frontier Kernel Evolution), a new framework that enhances the ability of AI coding agents to autonomously write GPU kernels. Previously, while agents could propose and test kernels, they struggled with complex tasks typically reserved for expert programmers. CAKE addresses this by creating a new language tailored for GPU scheduling, allowing agents to receive in-depth feedback from the compiler that highlights specific performance issues. This iterative process not only improves the kernels produced but also adapts the compiler itself based on the challenges faced by the agents. This advancement is significant for the AI/ML community as it bridges the gap between automated code generation and the specialized requirements of modern GPU architecture. By enabling agents to learn from their mistakes and improve over time, CAKE could reduce the reliance on human experts for kernel writing, potentially accelerating the development of advanced AI applications. The paper indicates improved performance metrics, showing that the enhanced agent-compiler interaction leads to faster and more efficient GPU code execution, ultimately benefiting large language models that depend heavily on optimized GPU operations.