KernelAgent: Hardware-Guided GPU Kernel Optimization via... Kaiming Cheng, Jack Khuu & Laura Wang

Why watch

This presentation covers KernelAgent, Meta's hardware-guided GPU kernel optimization system that extends the original PyTorch KernelAgent with a multi-agent workflow integrating GPU hardware-performance signals. The system achieves 2.02x speedup over previous versions and 1.56x speedup vs default torch.compile on H100 GPUs, targeting ML engineers and systems programmers working on GPU performance optimization.