PDF 原檔：報告_OCP2026_Meta_主題演講_GW級AI基礎設施_20261015_original.pdf
講者公開簡報（OCP 官方 Recorded Content 頁提供下載）。文字以 PyMuPDF 逐頁抽取，
### p<n>為頁碼；圖表內容見原檔。
圖片清單（已驗證 2026-10-19）
| 檔名 | 頁 | 分類 | 親眼所見內容 |
|---|---|---|---|
報告_OCP2026_Meta_主題演講_GW級AI基礎設施_20261015_p3.png |
p3 | 真資料圖 | Prometheus 多棟園區俯視示意：各棟間以藍色光纖線連接，標註距離 321m、800m、885m、1500m、1500m、3400m、3800m，左下與右上為帳篷式廠房 |
報告_OCP2026_Meta_主題演講_GW級AI基礎設施_20261015_p7.png |
p7 | 真資料圖 | 兩組 Region 各含 4 個 GPU 區塊（每個區塊為 2×2 共 4 個「20k GPUs」方格，即每區塊 4×20k），經 Backend Aggregation (BAG) 彙整，兩個 BAG 之間以雙向箭頭互連 |
報告_OCP2026_Meta_主題演講_GW級AI基礎設施_20261015_p8.png |
p8 | 真資料圖 | MetaRoCE 對照圖：今日傳輸把網路當單一路徑、壅塞即停（粗訊號、全部重送）；MetaRoCE 多路徑、逐路徑訊號、只重送遺失封包 |
報告_OCP2026_Meta_主題演講_GW級AI基礎設施_20261015_p11.png |
p11 | 真資料圖 | Prometheus 現場空拍：多座長條形帳篷式（tent）廠房並列施工中 |
報告_OCP2026_Meta_主題演講_GW級AI基礎設施_20261015_p15.png |
p15 | 真資料圖 | GPU-1 與 GPU-2 每 8K GPU 每日中斷次數對數曲線，目標線 2.0；兩條趨勢線 5.5%/wk 與 5.9%/wk，實測值約降至 4–5 區間，趨勢線與目標線在右側交會 |
報告_OCP2026_Meta_主題演講_GW級AI基礎設施_20261015_p20.png |
p20 | 真資料圖 | Titan AI 訓練叢集：Prometheus 標示「1 GW+ cluster」、Hyperion 標示「5 GW over the next few years」，各附空拍施工照 |
原始內容
p1
Concrete to Code: Engineering AI Infrastructure at Gigawatt Scale Santosh Janardhan Head of Global Infrastructure, Meta
p2
Assumption 1 A cluster lives under one roof.
p3
PROMETHEUS Our multi-region AI training cluster 800m 3400m 3800m 1500m 1500m 885m 321m
p4
10km 5km
p5
100km 10km 5km
p6
5km Backend Aggregation (BAG) Region A 20k 20k 20k 20k GPUs Region B 20k 20k 20k 20k GPUs Region C 20k 20k 20k 20k GPUs Region D 20k 20k 20k 20k GPUs
p7
10km Region W 20k 20k 20k 20k GPUs Region X 20k 20k 20k 20k GPUs Region Y 20k 20k 20k 20k GPUs Region Z 20k 20k 20k 20k GPUs 5km Backend Aggregation (BAG) 5km Backend Aggregation (BAG) Region A 20k 20k 20k 20k GPUs Region B 20k 20k 20k 20k GPUs Region C 20k 20k 20k 20k GPUs Region D 20k 20k 20k 20k GPUs
p8
MetaRoCE A transport that can see every path. TODAY Treats the Network as one single path. Congestion stalls everything. MetaRoCE Treats the Network as several paths. Routes around congestion. Coarse signals · Resend everything Per-path signals · Resend only what's lost
p9
（本頁無可抽取文字，內容在圖中）
p10
Assumption 2 Data centers are built from concrete and steel.
p11
（本頁無可抽取文字，內容在圖中）
p12
（本頁無可抽取文字，內容在圖中）
p13
Assumption 3 Once it’s built, it just works.
p14
Nvidia GB200 Compute System
p15
300 200 100 50 30 20 10 7 5 3 2 1 Target: 2.0 /8K GPUs/day GPU-1 vs GPU-2 Interruption Trends Log Scale | Interruptions per 8K GPUs/day | Dotted = projection Interruptions per 8K GPUs/day (log) GPU-1 Actual GPU-2 Actual GPU-1 Trend (5.5%/wk) GPU-2 Trend (5.9%/wk)
p16
NO ONE BUILDS THIS ALONE DATA CENTERS NETWORK CHIPS & RACKS Co-designed at every layer, at every stage. Option A - 1 slide
p17
Assumption 4 Software is the easy part.
p18
（本頁無可抽取文字，內容在圖中）
p19
（本頁無可抽取文字，內容在圖中）
p20
1 GW+ cluster 5 GW over the next few years PROMETHEUS HYPERION Our Titan AI Training Clusters
p21
INFRASTRUCTURE