# GPU Forecasters: Predicting GPU Kernel Performance with Large Language Models to Accelerate Optimization > This article introduces the GPU Forecasters method, which uses large language models as selective proxies to predict GPU kernel performance. It can evaluate several times more candidate solutions under the same GPU budget, significantly improving kernel search efficiency. - 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm) - 发布时间: 2026-05-29T15:56:08.000Z - 最近活动: 2026-06-01T02:56:17.669Z - 热度: 101.0 - 关键词: GPU内核优化, 大语言模型, 性能预测, 强化学习, 代码优化, 深度学习系统, 代理模型, 编译器优化 - 页面链接: https://www.zingnex.cn/en/forum/thread/gpu-forecasters-gpu - Canonical: https://www.zingnex.cn/forum/thread/gpu-forecasters-gpu - Markdown 来源: floors_fallback --- ## Introduction: GPU Forecasters—Accelerating GPU Kernel Optimization with LLM # Introduction: GPU Forecasters—Accelerating GPU Kernel Optimization with LLM This article introduces the GPU Forecasters method published on arXiv on May 29, 2026. Its core is to use large language models (LLM) as selective proxies to predict the relative performance of GPU kernels, reducing reliance on actual hardware measurements. Under the same GPU budget, this method can evaluate several times more candidate solutions, significantly improving kernel search efficiency and aiding deep learning system optimization. ## Background: Measurement Bottlenecks in GPU Kernel Optimization and Contradictions Brought by LLM # Background: Measurement Bottlenecks in GPU Kernel Optimization and Contradictions Brought by LLM GPU kernels are core workloads in deep learning, and their performance directly determines system efficiency. Existing optimization methods require repeated hardware measurements. Although a single measurement (compilation + execution + synchronization) is short, the total cost of large-scale searches is extremely high. In recent years, LLM's code generation capability has improved, but hardware evaluation has become a bottleneck—they can generate a large number of candidates but lack sufficient GPU time to evaluate them. The core question: Can we use LLM to predict performance and reduce hardware reliance? ## Method: Selective Proxy Mechanism of GPU Forecasters # Method: Selective Proxy Mechanism of GPU Forecasters The GPU Forecasters framework uses LLM as a selective proxy, which needs to have two characteristics: 1. Accuracy (predicting relative performance); 2. Selectivity (delegating hardware measurement when uncertain). The prediction target is relative performance (whether A is faster than B), which is easier than absolute values and meets search requirements. Proxy evaluation metrics include accuracy, calibration (confidence matches reliability), and utility (finding faster kernels under limited budget). ## Method: RL Enhancement and Kernel Search Process # Method: RL Enhancement and Kernel Search Process The study uses reinforcement learning (RL) to improve proxy performance: penalizing incorrect high-confidence predictions and rewarding correct selectivity. Experiments show that accuracy and calibration are significantly improved after RL training. Steps to integrate into the search process: 1. Generate candidates; 2. Proxy predicts pairwise performance; 3. Accept high-confidence predictions, perform hardware measurement for low-confidence ones; 4. Update search state and repeat until budget is exhausted. ## Experimental Evidence: Efficiency Improvement and Value of Selective Mechanism # Experimental Evidence: Efficiency Improvement and Value of Selective Mechanism Key findings: LLM's relative performance prediction accuracy exceeds random; RL training further improves it; under the same GPU budget, it can evaluate 3-5 times more candidates, and the best kernel found is 10-20% faster. The selective mechanism is key: no selectivity degrades search quality, and retaining 10-30% of candidates for measurement achieves the best balance. ## Technical Details: Input Representation and Prediction Head Design # Technical Details: Input Representation and Prediction Head Design The input needs to include kernel code, hardware specifications, and problem scale; the combination of code and hardware representation yields the best results. Prediction head design: pairwise comparison head (outputs the probability that A is faster than B), ranking head (relative performance score), confidence head (prediction + confidence). Training data comes from historical search logs, synthetic data, and public benchmarks. ## Practical Recommendations: Steps to Apply GPU Forecasters # Practical Recommendations: Steps to Apply GPU Forecasters Data collection: Record search history, establish benchmark suites, and continuously update data; Proxy deployment: Start with simple pairwise comparisons, then gradually introduce confidence and RL; Integration with search: Dynamically allocate budget, tolerate errors, and update the proxy via online learning. ## Conclusion: A New Direction for AI-Assisted AI Infrastructure Optimization # Conclusion: A New Direction for AI-Assisted AI Infrastructure Optimization GPU Forecasters embodies the recursive value of AI-assisted AI infrastructure optimization—LLMs not only generate code but also understand performance characteristics. In the future, it can be extended to hardware such as TPU/NPU and multi-objective optimization, becoming a key tool for complex system optimization and driving the development of AI from the application layer to the infrastructure layer as an optimization engine.