跳转至

文章背景与核心概要

矩阵乘法(如半精度通用矩阵乘法 HGEMM)是深度学习和现代人工智能工作负载的核心计算瓶颈。尽管 NVIDIA 的 cuBLAScuBLASLt 等闭源库经过了数十年的极致人工优化,但人类工程师在探索庞大的参数与代码配置空间时依然存在极限。

本文介绍了一种名为 CUDA-L2 的创新系统,它将大语言模型(LLM)与强化学习(RL)有机结合,用于自动优化 HGEMM CUDA 内核。CUDA-L2 以实际 CUDA 执行速度作为强化学习的奖励信号,系统化地探索高达 1,000 种配置方案。实验结果表明,无论是在离线模式还是在模拟实时推理的服务模式下,CUDA-L2 的性能均全面超越了包括 torch.matmulcuBLAS 以及 cuBLASLt 在内的行业顶尖基准,展示了 LLM 引导的自动化技术在突破极端性能内核方面的巨大潜力。

CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learning

arXiv:2512.02551 [cs.LG]
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
Authors: Songqiao Su, Xiaoya Li, Albert Wang, Guoyin Wang, Jiwei Li, Chris Shum
Submitted: 2 Dec 2025 (v1); Last revised: 5 Aug 2026 (v4)
DOI: 10.48550/arXiv.2512.02551
Project Code & Resources: GitHub - ornith-ai/CUDA-L2


📋 Summary

CUDA-L2 is an innovative system that integrates Large Language Models (LLMs) with Reinforcement Learning (RL) to automatically optimize Half-precision General Matrix Multiply (HGEMM) CUDA kernels. By utilizing CUDA execution speed as the RL reward signal, CUDA-L2 systematically explores vast configuration spaces—scaling to 1,000 configurations in ways that are impractical for human engineers.

Across rigorous evaluations, CUDA-L2 outperforms leading matrix-multiplication baselines, including torch.matmul and NVIDIA’s closed-source, highly-optimized libraries (cuBLAS and cuBLASLt), in both offline and server inference modes.

📑 摘要

本文提出了 CUDA-L2,这是一个将大语言模型(LLMs)与强化学习(RL)相结合的系统,用于自动优化半精度通用矩阵乘法(HGEMM)CUDA 内核。利用 CUDA 执行速度作为 RL 奖励,CUDA-L2 在 1,000 种配置中自动优化 HGEMM 内核。

在各项严苛的评估中,CUDA-L2 系统性地超越了迄今为止主要的矩阵乘法基准,从广泛使用的 torch.matmul 到 NVIDIA 顶尖的闭源库(即 cuBLAScuBLASLt)。

  • 离线模式(内核连续执行且无时间间隔):
  • 平均比 torch.matmul 提升 +22.0%
  • 在采用最优布局配置(标准-标准 NN 和转置-标准 TN)时,比 cuBLAS 提升 +19.2%
  • cuBLASLt-heuristic(查询 cuBLASLt 库并根据启发式建议选择算法)提升 +16.8%
  • 比最具竞争力的 cuBLASLt-AutoTuning 模型(从 cuBLASLt 的建议中最多 100 个候选里选择最快算法)提升 +11.4%

  • 服务模式(以随机时间间隔执行内核以模拟实时推理):

  • 性能加速进一步提升至:torch.matmul+28.7%cuBLAS+26.0%cuBLASLt-heuristic+22.4% 以及 cuBLASLt-AutoTuning+15.9%
  • Offline Mode (kernels executed consecutively without time intervals):
  • +22.0% over torch.matmul on average.
  • +19.2% over cuBLAS using the optimal layout configuration (normal-normal NN and transposed-normal TN).
  • +16.8% over cuBLASLt-heuristic (which queries the cuBLASLt library and selects algorithms based on heuristic suggestions).
  • +11.4% over the most competitive cuBLASLt-AutoTuning model (which selects the fastest algorithm from up to 100 candidates from cuBLASLt's suggestions).

  • Server Mode (kernels executed at random intervals simulating real-time inference):

  • Speedups further increase to +28.7% (torch.matmul), +26.0% (cuBLAS), +22.4% (cuBLASLt-heuristic), and +15.9% (cuBLASLt-AutoTuning).

CUDA-L2 demonstrates that even the most performance-critical, heavily-optimized kernels like HGEMM can be significantly improved through LLM-guided RL automation by exploring configuration scales previously impractical for human developers.


🔗 访问与链接


📦 提交历史

  • [v1] Tue, 2 Dec 2025 09:20:15 UTC (545 KB)
  • [v2] Fri, 12 Dec 2025 00:47:40 UTC (545 KB)
  • [v3] Mon, 13 Jul 2026 00:51:09 UTC (1,494 KB)
  • [v4] Wed, 5 Aug 2026 18:31:31 UTC (1,494 KB) (this version)