SOL-ExecBench：针对硬件极限的实用GPU内核基准测试
机器学习
2026-03-20 v1 人工智能
摘要
随着智能体AI系统在生成和优化GPU内核方面日益强大，进展受限于那些奖励软件基线速度提升的基准测试，而非接近硬件高效执行的程度。我们提出SOL-ExecBench，一个包含235个CUDA内核优化问题的基准测试套件，这些问题从124个生产级和新兴AI模型中提取，涵盖语言、扩散、视觉、音频、视频和混合架构，目标为NVIDIA Blackwell GPU。该基准测试覆盖BF16、FP8和NVFP4上的前向和反向工作负载，包括预期依赖Blackwell专属能力的内核。与先前基于软件实现评估内核的基准测试不同，SOL-ExecBench将性能衡量为对速度-光界限（Speed-of-Light, SOL）的接近程度，这些界限由我们用于推导硬件 grounding SOL界限的SOLAR管道计算得到，为硬件高效优化提供固定目标。我们报告SOL得分，用于量化候选内核在解决释放定义评分基线与硬件SOL界限之间剩余差距的程度。为支持智能体优化器的稳健评估，我们额外提供了一个包含GPU时钟锁定、L2缓存清除、隔离子进程执行以及针对常见奖励作弊策略的静态分析检查的沙箱环境。SOL-ExecBench重新定义了GPU内核基准测试，从击败可变软件基线转向缩小至硬件速度-光界限的剩余差距。
引用
@article{arxiv.2603.19173,
title = {SOL-ExecBench: Speed-of-Light Benchmarking for Real-World GPU Kernels Against Hardware Limits},
author = {Edward Lin and Sahil Modi and Siva Kumar Sastry Hari and Qijing Huang and Zhifan Ye and Nestor Qin and Fengzhe Zhou and Yuan Zhang and Jingquan Wang and Sana Damani and Dheeraj Peri and Ouye Xie and Aditya Kane and Moshe Maor and Michael Behar and Triston Cao and Rishabh Mehta and Vartika Singh and Vikram Sharma Mailthody and Terry Chen and Zihao Ye and Hanfeng Chen and Tianqi Chen and Vinod Grover and Wei Chen and Wei Liu and Eric Chung and Luis Ceze and Roger Bringmann and Cyril Zeller and Michael Lightstone and Christos Kozyrakis and Humphrey Shi},
journal= {arXiv preprint arXiv:2603.19173},
year = {2026}
}