结果是显著的：StitchCUDA 将 Hacking 率从 Kevin-32B 的 52% 降至 16%， Hacking 从 4 次降至 0 次。而去除 Rubric 的 StitchCUDA-A 变体，Hacking 率回升至 32%，进一步验证了 Rubric Reward 的因果效应。

本文旨在梳理作者学习路径，带领读者共同探索 GPU Kernel 性能分析从宏观到微观的技术演进。 引言 作为一名使用eBPF进行CPU性能分析的工程师，在转向学习GPU性能优化分析时，一直在思考GPU上是否有技术也可以实现用户自定义探针式性能分析？学习NVIDIA Nsight ...

这项由香港科技大学、字节跳动、香港中文大学（深圳）以及南洋理工大学联合开展的研究发表于2026年，研究团队开发出了一套完整的训练系统，让大语言模型学会编写高性能的GPU内核代码。这项突破性工作首次系统性地解决了用强化学习训练AI模型编写内核 ...

AI 流行的当下，你有没有想过： 大模型推理服务到底怎么跑起来的？大模型推理服务的运行过程中，CPU和GPU分别负责哪些工作？ 用GPU一定比CPU跑的快么？哪些场景需要用GPU? GPU最初的使命是加速图形渲染。而渲染一帧图像，本质上就是对数百万个像素点进行相似 ...

The optimisation of GPU kernels through performance tuning and auto-tuning approaches has become essential in maximising computational efficiency on modern heterogeneous architectures. Researchers ...

Back in 2000, Ian Buck and a small computer graphics team at Stanford University were watching the steady evolution of computer graphics processors for gaming and thinking about how such devices could ...

Support for unified memory across CPUs and GPUs in accelerated computing systems is the final piece of a programming puzzle that we have been assembling for about ten years now. Unified memory has a ...

Project Babylon would extend the reach of Java to foreign programming models such as machine learning models, GPUs, SQL, and differential programming. Java would be extended to foreign programming ...

Graphics processing units (GPUs) were originally designed to perform the highly parallel computations required for graphics rendering. But over the last couple of years, they’ve proven to be powerful ...