PDF 原檔：報告_OCP2026_Microsoft_訓練功率振盪與機櫃內濾波_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_Microsoft_訓練功率振盪與機櫃內濾波_20261015_p3.png |
p3 | 真資料圖 | AI 訓練 GPU 功率時序（正規化，約 200 秒）：運算段約 0.8–1.0、通訊重同步段降至約 0.1–0.2 的週期方波；右側列電網共振頻帶（0.1–0.5、0.5–0.8、0.8–2、5–55 Hz） |
報告_OCP2026_Microsoft_訓練功率振盪與機櫃內濾波_20261015_p6.png |
p6 | 真資料圖 | 各電網大型負載要求表：MISO、ATC、ERCOT、SOCO、AESO、LIPA 的發電容量、有功功率分鐘內變動上限與爬升率上限（例 MISO 10MW／5s 滑動窗、30MW/min；AESO 10MW/min） |
報告_OCP2026_Microsoft_訓練功率振盪與機櫃內濾波_20261015_p8.png |
p8 | 真資料圖 | 產業緩解策略圖：電網→HV 變電站→MV 開關→降壓變壓器→LV 開關→雙轉換 UPS→IT 機櫃，標示 MV UPS＋BESS 貨櫃、機櫃內電容托盤、PSU 與軟體功率控制迴路 |
報告_OCP2026_Microsoft_訓練功率振盪與機櫃內濾波_20261015_p10.png |
p10 | 真資料圖 | 電容式功率濾波器放置位置：軟體／GPU 功率平滑→近 GPU 主動濾波（48V IBC 後 DC-DC 加電容）→機櫃級電源架 DC 匯流排→機房列 busway＋PDU；左下為放電／充電削峰曲線 |
報告_OCP2026_Microsoft_訓練功率振盪與機櫃內濾波_20261015_p11.png |
p11 | 真資料圖 | 機櫃內主動濾波選項：PSU 內部 VBlk 軟匯流排與外掛電容托盤（OCP 2024 Power Capacitor Shelf）電路對照，DoD＝1−(Vmin/Vmax)² 公式與電容托盤優點 |
原始內容
p1
Bo Wen Microsoft Xianyong Feng Microsoft Luz Belmont Microsoft Mitigating AI Training Power Oscillations: Utility Requirements and In-Rack Power Filtering ] RACK & POWER
p2
• Workload characteristics of large AI training clusters • Key power grid stability concerns • Utility requirements for instantaneous active power at large datacenters • System-wide synchronized monitoring framework • Mitigation strategies used across the industry • In-rack power filtering approach • Call to Action Outline
p3
• Resonant frequencies in the bulk power system 1. Inter-area oscillations: 0.1–0.5 Hz 2. Regional oscillations: 0.5–0.8 Hz 3. Local oscillation modes: 0.8–2 Hz 4. Sub-synchronous and torsional oscillations: 5–55 Hz • Traditional PSU bulk capacitance can attenuate power components above line frequency seen by downstream IT load Large AI Cluster Training Workload Computing stage Idling – data comm. & resync Typical AI Training Load Profile Ref: Power Stabilization for AI Training Datacenters
p4
North America Bulk Power System (BPS) Ref: How it Works: The Role of a Balancing Authority
p5
- Forced oscillations a) Inter-area oscillations – large-grid stability and blackout risk b) Regional oscillations – regional grid stability c) Torsional oscillations – turbine-generator shaft damage
- Flicker
- Grid–datacenter interface impedance stability (13–27 Hz)
- Large grid disturbances a) AI training cluster (100s of MW to GW) start-up and shutdown b) Large load transfer during utility grid short-circuit faults Major Grid Concerns
p6
Region Generation Capacity (GW) Load Variability Limit (or proposed) Ramp Rate Limit (or proposed) MISO ~138 GW 10 MW per 5s sliding window 30 MW/min ATC (in MISO) - 25 MW per 5s sliding window No individual freq. exceeds 10 MW 30 MW/min ERCOT ~105 GW 10 MW per 5s sliding window - SOCO (SBA) ~50 GW 10 MW for 0.1-5 Hz for single freq. 2.5 MW for 5-55 Hz for single freq. 10 MW per 4-6s for P75 1 MW for median change 20 MW/min AESO ~23.4 GW Originally 0.16 MW (study-based now) 10 MW/min LIPA - 20 MW for 0.1-5Hz range 10 MW for 0.1-5 Hz for single freq. 3.5 MW for 5-55 Hz for single freq. 20 MW/min Utility Requirements for Large Loads Active Power Intra-Minute Limit and Ramp Rate Limit
p7
System-Wide Synchronized Monitoring Framework • Synchronized, high-resolution measurements at strategic points to characterize AI workload dynamics 1. HV substation: 100s of MW 2. MV switchgear: 10s of MW 3. LV switchgear: a few MW 4. Row: 100s of kW to MW 5. Rack input: 10s-100s of kW 6. IT chassis: kWs ... Database / dashboard NW
p8
• HV / MV • MV UPS + BESS • LV • LV UPS • Rack • Capacitor tray • Software solution Major Mitigation Strategies in Industry
p9
Power Filters with Capacitor: intuitive solutions from other industries Grid Non-Linear Load DC-AC Active power filter for electric arc furnace (EAF) with rapid and unpredictable load. Energy buffer for double- line frequency power: Google little-box challenge 2014. 1-Phase AC-DC or DC-AC Buffer Converter Energy buffer for active suspension system: Controller module for highly dynamic loads. DC-DC Automative 48V Active Suspension
p10
Power Filters with Capacitor: where to put them Software / GPU Power Smoothing GPU / server node accelerator DC rail Rack level power shelves DC bus Data hall / row busway + PDU Near-GPU active filter In rack active filter or active filter rack Active filter rack 48V Intermediate Bus Converter (IBC) VBus Load (VR, GPU) DC-DC Figures from: Lei Gu, APEC 2026
p11
Power Filters with Capacitor: where to put them Software / GPU Power Smoothing GPU / server node accelerator DC rail Rack level power shelves DC bus Data hall / row busway + PDU Near-GPU active filter In rack active filter or active filter rack Active filter rack Power capacitor shelf (PCS), OCP 2024 AC PFC + DC-DC VBus Load (IBC, VR, GPU) DC-DC AC PFC DC-DC DC VBlk A natural soft bus inside the AC-DC PSU: VBlk 1. Bulk capacitor’s Depth of Discharge (DoD) is limited a) AC voltage and component voltage rating 2. Capacitor shelf/tray is a more attractive solution a) Higher DoD b) Supercapacitors offer higher energy density than E-cap c) Careful coordination with the PSU is required DoD = 1 - (Vmin/Vmax)2 AC-DC PSU AC-DC PSU
p12
Power Filters with Capacitor: where to put them Software / GPU Power Smoothing GPU / server node accelerator DC rail Rack level power shelves DC bus Data hall / row busway + PDU Near-GPU active filter In rack active filter or active filter rack Active filter rack Busway level capacitor rack solution is not an ideal solution 1. Downstream components like PSU, IBC still suffers 2. Future DC bus way solution prefers capacitor shelf 3. Can be a patch solution for existing data centers DC Bus AI Racks DC-DC AC Bus AI Racks DC-AC
p13
Smarter Way to Control for Power Filtering ‘AI data center load forecasting: a liquid neural network approach,’ Mariam Mughees, 2026, IEEE TRANSACTIONS ON SMART GRID AC PFC + DC-DC VBus Load (IBC, VR, GPU) DC-DC Forecasting and coordination by power filter and GPU load 1. Help to prepare the charge and discharge positing 2. Upper stream UPS and BESS system can do the same for low frequency component filtering 3. GPU telemetry to power system deliver system is needed Forecasting Forecasting
p14
Call to Action 1. There is no technical blocker from power electronics point of view 2. Better capacitor technologies are needed to further push energy density 3. Smarter real-time forecasting algorithm is strongly desired 4. Standardized AI datacenter system-level synchronized monitoring framework 5. Collaborate through OCP to standardize AI workload characterization, synchronized monitoring, power-filtering performance metrics, and validation methods across rack, facility, and utility layers
p15
Thank You!