GPU Forecasters: Language Models as Selective Surrogates for Kernel Optimization
The paper discusses the use of language models (LLMs) as surrogates for optimizing GPU kernel performance. It highlights the challenges of evaluating kernel performance on GPUs due to high costs and proposes a method for LLMs to forecast kernel performance. The findings suggest that LLMs can enhance kernel search efficiency and lead to the discovery of faster kernels.
- ▪GPU kernels are essential for deep learning but require costly evaluations on hardware for optimization.
- ▪The study explores how LLMs can act as selective surrogates to forecast kernel performance, reducing the need for extensive GPU evaluations.
- ▪Experiments show that LLMs can accurately predict kernel performance and improve search efficiency through reinforcement learning.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.31464 (cs) [Submitted on 29 May 2026] Title:GPU Forecasters: Language Models as Selective Surrogates for Kernel Runtime Optimization Authors:Zaid Khan, Justin Chih-Yao Chen, Jaemin Cho, Elias Stengel-Eskin, Mohit Bansal View a PDF of the paper titled GPU Forecasters: Language Models as Selective Surrogates for Kernel Runtime Optimization, by Zaid Khan and Justin Chih-Yao Chen and Jaemin Cho and Elias Stengel-Eskin and Mohit Bansal View PDF Abstract:GPU kernels are the workhorse of modern deep learning, and optimizing them (via evolutionary search or coding agents) usually requires repeated measurement on target hardware.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
