PDF 原檔：報告_OCP2026_Google_晶片到電網協同設計_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）

原始內容

p1

Tom Garvens VP Data Center Technology & Systems From Chip to Grid: Scaling Holistic AI Data Center Co- Design Through Open Standards OCP Global Summit · October 2026 Amber Huffman Principal Engineer OCP Board of Directors

p2

AI Exponential Demands Data source: “Computation used to train notable intelligence systems”, Our World in Data AI is not only a powerful enabler, but a platform shift. Innovating on compute to power the AI-driven future ELMo (93.6M) BERT (340M) GPT-2 (1.5B) Megatron-LM (8.3B) GPT-3 (175B) LaMDA (137B) Gopher (280B) PaLM(540B) Megatron-Turing NLG(530B) T5 (11B) ~10x/yr trajectory 2018 2019 2020 2021 2022 0.01 0.1 1 10 100 1000 10000

of dense parameters

Parameters

p3

Exaflops 2021 2022 2023 2024 2025 AI Model Momentum 2026 Milestones Shipped Gemini 3.5 | 3.6 | 3.7 | 3.8

22 billion tokens/min 10B 16B 22B via direct API use by our customers Feb Apr Jul (100 tokens roughly equals 75 words) Gemini 4 Argon Our new frontier model · introduced Sep 30

p4

AI Stack Services Models AI Infrastructure Software AI Infrastructure Hardware Data Centers Energy

p5

TPU Supercomputing TPU 8t & TPU 8i 2026 Ironwood 2025 Over 9k chips & designed to meet the growing demands of the most complex thinking models Trillium 2024 Exceptional performance & efficiency. Enabling the next frontier of AI models TPU v5p 2023 A powerful, scalable, and flexible AI accelerator TPU v5e 2023 Purpose-built cost-efficiency and performance for medium/large- scale training and inference TPU v4 2022 Optically reconfigurable 3D Torus 4k chips with distributed shared memory TPU v3 2020 Liquid cooled 1k chips distributed shared memory TPU v2 2018 Domain-specific AI supercomputing 256 chips distributed shared memory TPU v1 2015 Internal inference accelerator

p6

A Small Slice Becoming a Grid Resource 2023 ~4,000 TWh total Source: LBNL, 2024 United States Data Center Energy Usage Report (Dec 2024) for usage Source: EIA, Electricity Explained and Annual Energy Outlook for totals Data Centers as a Share of Total U.S. Electricity Use 4.4 % 176 TWh 6.7 - 12 % 325 - 580 TWh All other U.S. demand 2028 scenario ~4,300 TWh total

p7

Impacts due to ML workloads Ramp Rate Oscillations Intra-Minute Fluctuations Example 24 hour Load Profile for Training and Inference Workloads

p8

New Utility Requirements Region / Grid Operator Required Timeframe FinGrid (Finland) 90% within 1s ERCOT (Texas, USA) 90% within 2s AESO (Alberta, Canada) 100% within 1.0 to 10.0 seconds (configurable) Energinet (Denmark) 80% within 5 seconds, then 90% within 20 seconds, 95% within 30s Fingrid POI Voltage (pu) TenneT SPP RTE Energinet ERCOT ATC AERSO 1 2 3 4 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Source: Energy Systems Integration Gro up. Time (s) Low Voltage Ride-Through Large load disconnect ride-through and restore requirements

p9

Mitigation Across Every Timescale No single layer covers the full frequency range of an AI load. Coverage has to be built up by timescale. µs ms 100 ms 1 s 10 s Silicon design features Targeted design features Layered software control Workload and resiliency CBU / near-rack supercaps Near load, high di/dt Facility-level BESS Longer duration, multi-purpose Holistic Solutions Required

p10

Mitigation in Action Mitigation on: the swing goes flat. See it at Booth C1.

p11

Battery Energy Storage Systems (BESS) Grid Service Participation Demand Response Peak Shaving Supplying Power Back A B C A B C Peak Shaving (covering local demand spikes) Demand Response (ex. grid stability during extreme weather) Supplying Power Back (if regulations allow) Buffer Smooths the load the grid sees Backup Reduces diesel runtime

p12

On March 19, 2026, we announced a milestone in demand- side flexibility in the United States. Along with other new resources we’re bringing to the system, demand response enables valuable capacity while helping our utility partners support reliability. Data Center Load Grid Demand Demand Response Window 12am 6am 12pm 6pm 12am ML workloads Shifted or reduced Google has integrated 1 GW of demand response capacity into our contracts with multiple U.S. utilities

p13

Data Centers For The AI Era

p14

ML Rack Density Challenge TPU rack

p15

LVDC Journey - Existing LVDC Sidecar Standard AC Distribution

p16

LVDC Journey - Today LVDC Sidecar Hybrid Transition Today Standard AC Distribution

p17

LVDC Journey - Native LVDC LVDC Native Optimized SST/TRU Today Standard AC Distribution LVDC Sidecar Hybrid Transition

p18

LVDC Native Architecture Benefits 20–30% Floor space densification Sidecar footprint returns to the data hall 9,000 Install hours eliminated Fewer field-assembled boxes and connections

1% Chip-to-grid efficiency One conversion stage instead of three Consolidating conversion drives both density and velocity. Figures are internal estimates compared to current AC baseline. Hundreds of rectifiers and batteries, distributed Centralized SST and BESS

p19

Open Standards For The AI Era

p20

Why OCP: Align in Months, Not Years 2024 ~1,500 MW dropped offline at once Feb 2026 First draft Jun 2026 Critical items aligned Sep 2026 v1.0 in effect Speed Specs at product speed Scale One spec, many vendors Trust A neutral venue Ride through dips 90% load back in 2 s

p21

800 VDC, From Grid to Chip Mt Diablo ±400 VDC power rack Native LVDC 800 VDC to the row SST MV AC to 800 VDC · v0.3 On-board power New workgroup, Q4 2026 1 2 3 4

p22

Liquid Cooling, Open Data Center Blueprints the whole supply chain can build to Project Deschutes 1.0 Open 5th-gen CDU spec Project Brazos Liquid Cooling for Air-Cooled DCs Open Data Center 1.0 7 standard spaces, 12–24 MW blocks

p23

PUE — Power Usage Effectiveness Google strives to build the world’s most energy-efficient computing infrastructure, outfitting each Google data center with high-performance servers designed to use as little energy as possible compared to the amount of data they process. 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 1.04 1.08 1.12 1.16 1.20 1.24 Quarterly PUE Trailing twelve-month (TTM) PUE

p24

Reducing Embodied Carbon Green concrete Goal: >50% lower concrete carbon Carbon disclosure 0–10 data quality score Product category rules PCR v0 for ICT hardware Shared methods let every builder reduce embodied carbon. Clean backup power requirements <10% of diesel CO₂ <10 s to full power

p25

Responsible by Design Sources: datacenters.google/energy · Google 2026 Environmental Report 12 GW new clean energy, 2025 35 GW since 2010 1 GW demand response 120% replenishment goal, 2030 7B+ gallons replenished $500M+ for community water Energy Water

p26

Go Deeper at OCP This Week GRID & POWER Grid Requirements & Mitigation Wed 4:05 PM · 210CG Hardware Mitigations: Power Fluctuation Thu 1:00 PM · 210CG 800 VDC On-Board Power Wed 8:00 AM · 210CG LVDC Ecosystem Wed 9:10 AM · 210CG 800V Power Safety Wed 12:55 PM · 210CG SST Integration Wed 3:15 PM · 210CG COOLING & OPEN DC CDU Sub-Project Tue 2:25 PM · 220C Single-Phase Cold Plate Thu 10:45 AM · 220C Open DC Architecture Thu 12:30 PM · 210CG SUSTAINABILITY Sustainability Project Updates Thu 8:00 AM · LL21DE High-Temperature Fuel Cells Thu 2:45 PM · 210CG Open Enablers Thu 12:50 PM · LL21DE PCR for LCA Thu 2:45 PM · LL21DE Carbon Disclosure Thu 3:35 PM · LL21DE AT BOOTH C1 · ML power-fluctuation demo · Mt Diablo 800 VDC power sidecar · Ironwood rack

p27

Collaborate With Us at Product Speed Visit us at Booth C1 Next to the Central Park Lounge Tue 11 AM–7 PM · Wed 9 AM–5 PM · Thu 8 AM–2 PM IN THE BOOTH ● ML power-fluctuation mitigation demo ● Mt Diablo 800 VDC power sidecar ● TPU Ironwood rack ● Ask a Google engineer: office hours Scan: Google @ OCP sessions & office hours