https://lilinji.github.io/tags/roofline%E6%A8%A1%E5%9E%8B/Recent content in Roofline模型 on Ringi's LogHugo -- 0.152.2zh-cnThu, 03 Sep 2026 00:00:00 +0800https://lilinji.github.io/2026/09/%E7%AC%AC10%E8%AE%B2-ai-infra-%E6%80%A7%E8%83%BD%E5%B7%A5%E7%A8%8B%E6%96%B9%E6%B3%95%E8%AE%BA%E4%B8%8E%E4%B8%89%E8%B4%A6%E6%9C%ACflops-%E6%98%BE%E5%AD%98%E4%B8%8E%E9%80%9A%E4%BF%A1/Thu, 03 Sep 2026 00:00:00 +0800https://lilinji.github.io/2026/09/%E7%AC%AC10%E8%AE%B2-ai-infra-%E6%80%A7%E8%83%BD%E5%B7%A5%E7%A8%8B%E6%96%B9%E6%B3%95%E8%AE%BA%E4%B8%8E%E4%B8%89%E8%B4%A6%E6%9C%ACflops-%E6%98%BE%E5%AD%98%E4%B8%8E%E9%80%9A%E4%BF%A1/建立大厂 AI Infra 性能分析第一性原理：以生活直觉与极简小算盘拆解 MFU/HFU/TFLOPS 真实算力指标、2P/6P/8P 浮点计算量账本、静态与动态显存四账本、跨卡通信量账本与 Roofline 算术强度模型，手把手推导 LLaMA-2/3 7B/70B 训练与推理集群资源规划实战。https://lilinji.github.io/2026/09/%E7%AC%AC12%E8%AE%B2gpu-util-100-%E7%AE%97%E5%8A%9B%E5%8D%B4%E5%8F%AA%E6%9C%89-15sm-%E6%89%A7%E8%A1%8Cwarp-%E8%B0%83%E5%BA%A6tensor-core-%E4%B8%8E-roofline-%E6%A8%A1%E5%9E%8B%E7%9A%84%E7%AC%AC%E4%B8%80%E6%80%A7%E5%8E%9F%E7%90%86/Thu, 03 Sep 2026 00:00:00 +0800https://lilinji.github.io/2026/09/%E7%AC%AC12%E8%AE%B2gpu-util-100-%E7%AE%97%E5%8A%9B%E5%8D%B4%E5%8F%AA%E6%9C%89-15sm-%E6%89%A7%E8%A1%8Cwarp-%E8%B0%83%E5%BA%A6tensor-core-%E4%B8%8E-roofline-%E6%A8%A1%E5%9E%8B%E7%9A%84%E7%AC%AC%E4%B8%80%E6%80%A7%E5%8E%9F%E7%90%86/深入剖析 GPU 硬件执行与存储层次：从 SM 微架构、Warp 调度、Tensor Core MMA 原语到 Register/Shared Memory/L2/HBM 存储金字塔，结合 Arithmetic Intensity 与 Roofline 模型彻底破解 GPU 算力利用率指标陷阱。