Notes on LLM Infra.

LLM Infra 笔记。

• 2026.10 The Roofline Model, Starting from Measured Data

  Arithmetic intensity, how to measure the hardware peaks, how to read a roofline plot, and where LLM prefill and decode land, using measurements from eight GPUs.

• 2026.10 Roofline 模型：从实测数据出发

  用八种 GPU 上的实测数据讲 arithmetic intensity、硬件峰值怎么测、roofline 图怎么读，以及 LLM 推理的 prefill 与 decode 落在哪里。