Second edition第二版
How the networks behind AI training workAI 訓練背後的網路如何運作
For readers with no background in networking or machine learning. Read the thirteen chapters in order: each term is defined where it first appears.寫給沒有網路或機器學習背景的讀者。請依序讀這十三章:每個名詞都在它第一次出現的地方定義。
13 chapters · ≈11 h13 章 · 約 11 小時Start with Chapter 1 →從第 1 章開始 →What one training step computes一個訓練步驟算什麼
53 papers, 2024–202653 篇,2024–2026Research digests →研究導讀 →Recent papers in 11 topics, about 10 minutes each近期論文,分 11 個主題,每篇約 10 分鐘
162 terms162 個名詞Glossary →名詞表 →Every term, linked to where it is defined每個名詞都連到它的定義處
The thirteen chapters十三章
Part I · Foundations第一部 · 基礎
- 01What training computes, and why it needs thousands of GPUs訓練在算什麼,又為何需要上千顆 GPU30′
- 02The hardware: GPU, server, network card, switch and cable硬體:GPU、伺服器、網路卡、交換器與線纜45′
- 03Splitting the model: five parallelisms and the traffic they make切分模型:五種平行化與它們產生的流量60′
- 04Collective communication: algorithms, cost and the library that runs them集體通訊:演算法、成本與執行它們的函式庫55′
- 05The step-time budget: why the network decides how far training scales步驟時間帳:為何網路決定訓練能擴展到多大45′
Part II · The network第二部 · 網路
- 06Topology: how switches are wired into a network拓樸:交換器如何連成一個網路70′
- 07Transport: RDMA over Ethernet, lossless networks and congestion control傳輸層:Ethernet 上的 RDMA、無損網路與壅塞控制60′
- 08Routing and load balancing: why hashing fails and what replaces it路由與負載平衡:為何雜湊會失效,以及取代它的做法50′
- 09The front end: data, checkpoints and control前端網路:資料、檢查點與控制45′
- 10Failures and stragglers: the arithmetic of staying up故障與落後者:保持運轉的算術50′
- 11Sharing the cluster: isolation, placement and spare capacity共用叢集:隔離、擺放與備用容量45′
- 12Models and simulation: predicting step time before the hardware exists模型與模擬:在硬體存在之前預測步驟時間30′
Part III · Putting it together第三部 · 綜合
Beyond the chapters章節之外
Research digests研究導讀53
One recent paper per page, grouped by topic. Each topic says which chapter to finish first.每頁一篇近期論文,依主題分組。每個主題都標明要先讀完哪一章。
- Inside the server伺服器內部3 · after Ch. 43 篇 · 第 4 章後
- Workloads and traces工作負載與追蹤2 · after Ch. 52 篇 · 第 5 章後
- Parallelism and MoE平行化與 MoE6 · after Ch. 66 篇 · 第 6 章後
- Collective communication集體通訊16 · after Ch. 616 篇 · 第 6 章後
- Topology and optics拓樸與光學5 · after Ch. 65 篇 · 第 6 章後
- Load balancing and path monitoring負載平衡與路徑監測5 · after Ch. 85 篇 · 第 8 章後
- Transport and congestion control傳輸層與壅塞控制3 · after Ch. 83 篇 · 第 8 章後
- Checkpoints and the front end檢查點與前端網路3 · after Ch. 93 篇 · 第 9 章後
- Failures and diagnosis故障與診斷3 · after Ch. 103 篇 · 第 10 章後
- Sharing and simulation共用與模擬5 · after Ch. 125 篇 · 第 12 章後
- Whole-network designs整體網路設計2 · after Ch. 132 篇 · 第 13 章後
Guided readings論文導讀9
The production papers this primer is built on, read one at a time.這份入門所依據的生產環境論文,一次讀一篇。