PDF 原檔：報告_OCP2026_Microsoft_電源架聚合與串接架構_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；2026-10-23 校正 p10／p11 說明）
| 檔名 | 頁 | 分類 | 親眼所見內容 |
|---|---|---|---|
報告_OCP2026_Microsoft_電源架聚合與串接架構_20261015_p4.png |
p4 | 真資料圖 | 機櫃電源架構演進圖：獨立（各電源架自供）、聚合（共用機櫃電力池）、串接／分區（Zone 1–3）三種模式，標示故障域與共用程度 |
報告_OCP2026_Microsoft_電源架聚合與串接架構_20261015_p5.png |
p5 | 真資料圖 | 聚合與串接架構比較表：聚合為單一 ±400VDC 母排貫穿全櫃，串接為 Zone 1／Zone 2 逐級經 ±400VDC 傳遞；逐項比較容量利用、故障隔離、控制複雜度、效率、擴充性與優缺點 |
報告_OCP2026_Microsoft_電源架聚合與串接架構_20261015_p6.png |
p6 | 真資料圖 | 階層式機櫃電源協調架構：機櫃電源協調器（政策、配置、備用容量、故障恢復）與四個電源架控制器（快速迴路保護），電源架含 PSU→±400VDC 匯流排→DC-DC→50VDC 區→GPU；遙測上行、預算下行 |
報告_OCP2026_Microsoft_電源架聚合與串接架構_20261015_p10.png |
p10 | 文字卡 | Call to Action：把機櫃級電源協調變成可重複的工程實務——1 Design as a Rack（以機櫃為單一協調電源系統）、2 Coordinate & Protect（機櫃級共享容量、快速保護留在本地）、3 Measure & Prove（量測可用容量、閒置容量、備用、故障影響、暫態響應與復原時間）、4 Standardize Together（透過 OCP 統一術語與指標、遙測與控制介面、運行狀態、故障與恢復語意） |
報告_OCP2026_Microsoft_電源架聚合與串接架構_20261015_p11.png |
p11 | 文字卡 | Open Discussion：向 OCP 社群提三個問題——介面（跨電源架的最小遙測、能力與控制原語，如 Redfish）、安全邊界（本地保護、機櫃政策與退守行為如何分工與驗證）、共同指標（可用容量、閒置容量、損耗、故障影響、暫態與復原、可靠度） |
原始內容
p1
Praveen Kumar Velpuri Principal Firmware Engineer, Microsoft Yang Hsiung Principal Firmware Engineer, Microsoft Scaling Power Delivery for AI Infrastructure – Cascade vs Aggregated RACK & POWER ]
p2
OCP GLOBAL SUMMIT 2026 | RACK & POWER What this session addresses A system-level view of multi-shelf power delivery WHY NOW GPU demand is pushing beyond isolated shelf assumptions. 01 THE PROBLEM Sharing capacity introduces coordination and fault challenges. 02 THE PATTERN Hierarchical orchestration combines local safety with rack- level policy. 03 THE OUTCOME Higher usable capacity, lower losses and bounded failures domains. 04 2
p3
OCP GLOBAL SUMMIT 2026 | RACK & POWER Problem statement AI racks share capacity, while limits, faults, and transients remain local. 1 USABLE CAPACITY ≠ INSTALLED CAPACITY Nameplate power is not fully usable. Redundancy, derating, limits, and reserve define usable capacity. Model usable capacity 2 SHARED BUS ≠ UNLIMITED SHARING Aggregation pools power, but shelves and paths retain limits. A shelf can reach its limit while rack capacity remains available. Coordinate rack-wide sharing 3 LOCAL FAULTS CAN BECOME RACK EVENTS Shared paths can expand fault impact. Isolate the smallest safe domain and preserve healthy capacity. Contain the fault domain 4 GPU TRANSIENTS OUTRUN SOFTWARE Correlated load steps can consume reserve in milliseconds. Local protection acts first; rack software re-models, reallocates, and recovers. Protect fast, recover safely The challenge: coordinate shared capacity while preserving local protection and predictable recovery.
p4
Rack Power Architecture Evolution Three patterns, three different fault and sharing boundaries INDEPENDENT Shelf-local power SHELF SHELF SHELF Simple isolation Limited sharing AGGREGATED Rack-wide power pool SHELF SHELF SHELF Maximum pooling Larger common domain CASCADED / ZONED Controlled sharing by domain SHELF SHELF SHELF ZONE 1 ZONE 2 ZONE 3 Smaller fault domains More coordination points
p5
（本頁無可抽取文字，內容在圖中）
p6
（本頁無可抽取文字，內容在圖中）
p7
OCP GLOBAL SUMMIT 2026 | RACK & POWER Manage the Rack Within a Safe Power Envelope One coordinated loop turns telemetry into a safe power budget; the operating condition determines the response. 7 1 UNDERSTAND CAPACITY Observe + Model Measure power, temperature and health. Determine current usable capacity and reserve. 2 COORDINATE CAPACITY Allocate + Verify Distribute rack power budgets within the safe envelope. Verify response and rebalance as conditions change. 3 APPLY THE RIGHT RESPONSE Condition → Action Normal: monitor / optimize Constrained: cap Fast excursion: local HW throttle Link degraded: conservative fallback Faulted: contain Design principle: rack software coordinates sustained capacity; local hardware protects fast electrical limits.
p8
OCP GLOBAL SUMMIT 2026 | RACK & POWER Contain the Fault, Recover Safely, Prove the Result Limit fault impact, restore only verified capacity, then measure the outcome. 8 FAULT → RECOVERY PROVE IT WITH 4 RACK KPIs 1 CONTAIN Detect locally and isolate the smallest safe domain 2 STABILIZE Confirm voltage, current, and thermal state 3 REALLOCATE Re-model surviving usable capacity and reset budgets 4 RECOVER Reconnect gradually only after path and capacity verification USABLE CAPACITY Safely deliverable power ÷ installed capacity % / kW TRANSIENT RESPONSE Time for protection to limit a fast event ms FAULT IMPACT Rack resources affected by the fault trays / shelves / zones RECOVERY Time to restore stable usable capacity sec / min VALIDATE ACROSS Steady state → GPU load step → Cap change → Link loss / fault → Controlled recovery Success = preserve safety, minimize blast radius, retain usable capacity, and return to a stable policy.
p9
OCP GLOBAL SUMMIT 2026 | RACK & POWER
p10
（本頁無可抽取文字，內容在圖中）
p11
（本頁無可抽取文字，內容在圖中）