alphaXiv(@askalphaxiv):"Hawkeye: Hardware-Aware GPU Kernel Optimization with Minimal Supervision" Hawkeye teaches AI agents how to optimize for a new GPU using just 10 expert-written unit tests. Instead of hand-porting kernels to every new architecture, experts encode the key hardware tricks once and the agent composes them into fast CUDA or HIP kernels. It works across NVIDIA and AMD GPUs (Ampere, Hopper, Blackwell, and MI350), and often matching or beating expert-tuned baselines.

2026.08.21 02:21

"Hawkeye: Hardware-Aware GPU Kernel Optimization with Minimal Supervision" Hawkeye teaches AI agents how to optimize for a new GPU using just 10 expert-written unit tests. Instead of hand-porting kernels to every new architecture, experts encode the key hardware tricks once and the agent composes them into fast CUDA or HIP kernels. It works across NVIDIA and AMD GPUs (Ampere, Hopper, Blackwell, and MI350), and often matching or beating expert-tuned baselines.