← All resources

CudaForge: An Agent Framework with Hardware Feedback for CUDA Kernel Optimization

Auto-discovered

Paper Advanced

Zijian Zhang, Rong Wang, Shiyang Li, Yuebo Luo, Mingyi Hong, Caiwen Ding

Visit resource ↗

Description

CudaForge: An Agent Framework with Hardware Feedback for CUDA Kernel Optimization presents a training-free multi-agent workflow for CUDA kernel generation and optimization by Zijian Zhang, Rong Wang, Shiyang Li, Yuebo Luo, Mingyi Hong, and Caiwen Ding (University of Minnesota Twin Cities), indexed on hgpu.org and arXiv cs.LG.

It covers an iterative Coder and Judge agent loop: draft kernels, test correctness, incorporate hardware feedback such as Nsight Compute metrics, and refine. Reported results include high correctness, speedups over PyTorch baselines, cross-GPU and cross-model tests, and notes on generation time and API cost versus heavier agentic baselines.

It is for advanced CUDA, LLM-for-code, and GPU performance researchers who already write or evaluate kernels. Readers should know CUDA and basic LLM tooling.

The page is a free abstract-style paper entry with citation metadata and abstract text, with the full preprint linked from the listing.

It is useful as a multi-agent recipe that ties kernel search to profiler feedback instead of one-shot code generation. Link last verified accessible on 2026-09-29. Link last verified accessible on 2026-09-28.