PDF 原檔：報告_OCP2026_Google_AI負載功率波動硬體緩解_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_Google_AI負載功率波動硬體緩解_20261015_p3.png |
p3 | 真資料圖 | 每櫃 kW 隨年份曲線：AI/ML 2018–2028 急遽上升、Compute & Storage 持平；右側 AI 負載功率時序大幅鋸齒 vs 非 AI 負載平穩 |
報告_OCP2026_Google_AI負載功率波動硬體緩解_20261015_p6.png |
p6 | 真資料圖 | 三層儲能示意：晶片與機櫃級（整合電容＋機櫃內 CBU，暫態快速響應）→列級／sidecar（橋接與削峰）→設施級（大型 BESS／微電網，長時備援與電網支撐） |
報告_OCP2026_Google_AI負載功率波動硬體緩解_20261015_p8.png |
p8 | 真資料圖 | 儲能時間尺度示意：秒（快速波動與 ride-through）、分（橋接與削峰）、小時以上（長時備援與負載移轉），並建議機櫃端用雙功能儲能模組、AC 側長時備援用 BESS |
報告_OCP2026_Google_AI負載功率波動硬體緩解_20261015_p9.png |
p9 | 文字卡 | 儲能技術比較卡：LFP 電池（約 3k–10k 循環）、進階化學（Na、Zn 等）、電解電容（>5M 循環但僅短脈衝）、鋰離子超級電容（能量密度為 EDLC 的 4–5 倍） |
原始內容
p1
Vijay Boovaragavan Energy Storage Technical Lead, Google Xiong Li Rack Power Technical Lead, Google Hardware Mitigations for AI/ML Workloads Power Fluctuation ] RACK & POWER
p2
• The rise of Cloud computing led to the development of “hyperscale data centers” • MW scale facilities housing tens of thousands of machines • Typical IT rack power 10-20kW • Growth of AI/ML is transforming hyperscale data centers into AI factories • GW scale facilities optimized for the full AI application lifecycle from training to inference • ML rack power reaching 1MW and above! From Data Center to AI Factory Photo source Image source 2
p3
Challenges of AI/ML Workloads Power Demand Accelerated power increase Synchronized AI workloads Data center and in-rack di/dt events Non-AI Workloads kW per Rack 2018 2020 2022 2024 2026 2028 Compute & Storage AI/ML AI Workloads C. Nge, M. Iyengar, “How AI/ML Workloads are Transforming Data Center Power and Cooling”, APEC 2026 3
p4
4 Grid Instability & Utility Penalties Exceeding utility ramp-rate limits (e.g., 20 MW/min) triggers Supervisory Control and Data Acquisition (SCADA) alarms, violates power quality standards (IEEE 519/1453 flicker), and risks facility disconnection. Onsite Generators Rapidly dropping a load can cause "load shedding" issues, risking generator tripping. Datacenter Infrastructure Thermal and mechanical stress on electrical busducts, transformers, and mechanical cooling systems. Silicon Lifetime Impact Severe temperature cycling on accelerators reduces their long-term operational reliability. Infrastructure Impact: Datacenter & Utility Vulnerabilities Large cluster-level swings (measured in MW) have critical operational consequences
p5
Software vs Hardware Mitigations: The Performance Trade-off Mitigating Power Fluctuations: Software vs. Hardware Software Controls (Complementary Tool) Software controls serve as a key tool in the mitigation toolbox, enabling flexible workload-level management alongside hardware solutions. ● Intra-Job Smoothing (e.g., XLA ML WL): Dampens power variance by injecting synthetic compute, providing targeted workload stability. ● Power Capping & Sigkill: Provides direct control bounds over peak power execution rates when grid demands require immediate throttling. 5 Hardware-Based Smoothing (End-to- End) Eliminates performance penalties by letting workloads run flat-out at peak efficiency. ● Zero Performance Penalty: Maximizes hardware capabilities, enabling continuous execution at peak efficiency. ● Local Energy Buffering: Dampens the physical load locally to present a "smooth" power profile back to the utility grid and generators.
p6
Energy Storage Solutions for AI/ML Data Centers Managing the dynamic power demands of AI/ML workloads requires highly responsive, full-stack energy storage. Rather than a 'one size fits all' approach, a 3-tier solution tailored to different response timescales guarantees power resilience. 6
p7
Challenge: Advanced power systems require parallel, non-conflicting control of multiple power sources in a single power domain. Coexistence of power supply and energy storage modules 01 The Coexistence Challenge PSUs, Battery Backup Units (BBU), and Capacitor Backup Units (CBU) must operate in parallel on a common high-voltage/low-voltage DC bus: supply management, deployment speed, regulations, etc. 7 02 Transient Coordination A power plane configuration where the CBU actively handles dynamic load transients while the BBU remains passive until AC failure presents integration and stability challenges. Energy Storage Challenges 03 Current-Sharing (I-Share) Combining current-sharing signals (Ishare_PSU, Ishare_CBU, and Ishare_BBU) into a single peak- shaving controller that prevents control loop oscillations.
p8
Energy Storage Challenges 8 An optimization strategy: Minimize energy storage requirements and EDLCs at the payload rack by utilizing dual-function energy storage modules to manage both (1) critical power bridging/backup and (2) workload power smoothing. Hours+ backup on the AC-side can be addressed using a Battery Energy Storage System (BESS) or alternative backup solutions.
p9
Energy Storage Technologies Evaluating Energy Storage for High-Power, High-Temperature and High-Cycle Applications supporting both BBU and CBU use cases. Key Takeaway: Traditional batteries are unsuitable for high-frequency ML cycles (~1 million/year); hybrid supercapacitors, advanced Li-ion cell chemistries, emerging non-Li Tech could be promising candidates. Lithium-Ion (LFP) Batteries Safer but limited cycle life (~3k to 10k cycles). Requires massive oversizing to survive ML cycling. 9 Advanced Chemistries LFP/NMC pairing with advanced metal oxides anode. Sustainable materials: Na, Zn, HSB/C, etc Emerging Tech, Time-to- market. Electrolytic Capacitors Exceptional power density and low Equivalent Series Resistance (ESR). Virtually infinite physical cycle life (>5M+) but for only for short burst. Lithium-Ion Supercaps Hybrid approach. 4–5x higher energy density than EDLC with high power. Slightly lower cycle life, higher cost, not sufficient energy for backup use- case. Key Challenges: Supply Chain Management, Product Safety, Time-to-Market
p10
Industry Product Concepts Energy storage module supporting both BBU and CBU use cases, paired with rack-level EDLCs for ultra-high-frequency ride-through. 10 01 Sidecar/Row Level Combined B/C BU Must operate in parallel on a common DC bus. 02 Row Integration I-Share Integration Combining current- sharing signals into a single controller prevents loop oscillations. 03 10-Year Lifespan 100M+ Cycles Achieving hundreds of millions of cycles over a 10-year span to engage and guide. 04 Decentral Control DC Bus Signaling Eliminating high-speed communication lines simplifies system integration. 05 Supply Chain Souring Multiple technologies and sources to achieve the given set of specifications at the best Perf/TCO.
p11
[Call to Action] Building a Standardized OCP Ecosystem • Where to find additional information (URL links) • Mailing list: dcf-energy@OCP-All.groups.io • Problem to solve • Cooperatively develop and open-source high-voltage hybrid Battery/Supercapacitor module standards specifications within the OCP ecosystem at sidecar or row levels. • Get involved https://www.opencompute.org/community/energy • Timeline for contribution availability: next 2-3 quarters 11
p12
Thank You! 12