PDF 原檔：報告_OCP2026_Google_NVIDIA_LVDC開放生態系_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_NVIDIA_LVDC開放生態系_20261015_p6.png |
p6 | 真資料圖 | 資料機房搬移電源轉換示意：上圖為機櫃內含電源轉換（藍／紅機櫃層）、下圖將轉換移出機房，改由 SST（藍）與 ESS（紅）兩個設備箱經匯流排供電給整排機櫃 |
報告_OCP2026_Google_NVIDIA_LVDC開放生態系_20261015_p7.png |
p7 | 真資料圖 | 48–54VDC 母排供電方案：415V AC→54V DC 與 800V DC→54V DC 兩種機櫃電源架選項；文字標 DC:DC 電源架 55kW→95kW per RU；右上為 1RU 電源架前後視圖（6 顆 PSU） |
報告_OCP2026_Google_NVIDIA_LVDC開放生態系_20261015_p10.png |
p10 | 真資料圖 | 通用 busway 結構：DC（+ − + − PE）與 AC（L1 L2 L3 N PE）導體截面對照，以及 busway→tapbox→cable/whip→connector 的連接示意 |
報告_OCP2026_Google_NVIDIA_LVDC開放生態系_20261015_p11.png |
p11 | 文字卡 | LVDC 四階段路線：技術孵化（已完成）→單排驗證（2026，進行中）→資料中心級驗證（2027–2028）→大規模部署 GW 級 LVDC native（2029+） |
原始內容
p1
Jason Adrian - Systems Architect, Google Jared Huntington - Power Architect, Nvidia LVDC: Building an Open and Evolving Ecosystem AI OPEN DATA CENTER: Power Distribution Jason Jared
p2
• LVDC revolution impacts everything from the grid to the chip • ML racks already transitioning to LVDC out of necessity • LVDC enablement via sidecar tax is growing - 1MW racks drive imperative to native LVDC data centers! • Entire ecosystem of interoperable products needed, and must scale to 10’s of GW rapidly • How do we handle non-ML racks and other AC equipment? LVDC Revolution
p3
• 80%+ of the data center power is xPUs, but there are many other IT devices: • Compute, Storage, Scale-up/out & Core Networking, CDUs, Facility Fans, etc The Big Picture ●How to handle a mix of DC-native and AC loads and sources? ●Racks often need to be distributed across rows for power utilization, network cable length requirements, etc ●Splitting the data center into AC and DC powered rows would create complex placement challenges and fixed configurations
p4
What drives LVDC adoption? As rack power reaches ≥100kW, in-rack power conversion & backup consume up to 35%+ of available space: ●AC:DC Power Conversion: Takes 10–15% of rack capacity ●N+N redundancy: Increases power conversion 2x ●In-Rack BBUs: Consumes an additional 10–15% of space. Moving to LVDC offloads power conversion, freeing space for maximum xPU / accelerator density. Above 600kW: In-rack AC/DC conversion becomes completely unviable. Breaking Points For In-Rack Power
p5
xPU adoption of LVDC is obvious and under way, but what about everything else? • High density networking and some dense compute evaluating LVDC - hitting the rack power wall • Storage and low-density racks could remain AC (forces AC rows in future data centers, or AC inverters) or simply replace the AC:DC power shelves with 800VDC->54VDC power shelves • To ensure the most durable and fungible data centers we need to support a rapidly evolving ecosystem of products, we should unify around LVDC distribution to the IT rows. Should all racks move to LVDC?
p6
Moving Power Conversion Out of the Data Hall
p7
Powering Compute, Storage & Networking Standard Distribution 48-54 VDC Busbars Use 48-54 VDC busbars for anything < 100 kW if possible to ensure optimal efficiency and compatibility. Adapter Flexibility Modular Power Shelves Allows power shelves to flexibly provide required adaptation across generations: ● Existing Shelves: Optimized for 415 / 480 VAC ● New DC/DC Shelves: Built for 800 VDC architectures High-Power Loads & Density 2x DENSITY New designs can consider adapters or 800V direct to trays for high power requirements. DC:DC Power Shelves: Doubling power density Increases capacity from 55 kW → 95 kW per RU Rack Power Architecture Options 415V AC to 54V DC PAYLOAD 54V Busbar 800V DC to 54V DC PAYLOAD 54V Busbar
p8
• Reuse the same standard power whips where possible • Move CDUs and fans to DC motors - use same input connector and distribution as DC:DC power shelf • Standard 800 VDC DC/DC bricks are available now • PDB reference designs are available for cost optimization Making everything else LVDC
p9
• Inverters as a last resort • 800 VDC -> 54 VDC -> 240 VAC • DC/DC shelf + 1RU converter - not elegant but ok for small components • 800 VDC -> 240 VAC - Might be some uses • Crash cart plugs, controls equipment, lighting, etc… Enabling the last puzzle pieces
p10
Universal Busway • Not having interoperability between tapbox and busway causes issues • If the busway and tapbox don’t land at the same time the datacenter can’t be powered on. • If common solution can’t be converged on then RPP will be used • Need to standardize to limited set of connectors to enable rack fungibility
p11
Journey from R&D to Scale While the LVDC ecosystem does not yet fully exist, Google and Nvidia are driving an open ecosystem from innovation to mass deployment. 01 Tech Incubation LVDC architecture, SST testing, distribution, protection Completed 02 Single Lineup Validation Full single lineup testing (2026) Active Phase 03 Datacenter Validation Data center-scale validation (2027–2028) Upcoming 04 Mass Deployment Mass deployment of GW’s of LVDC native (2029+) Future Scale Hyperscalers and Data Center Facility builders are designing and establishing the ecosystem.
p12
The LVDC transition impacts everything from chip to grid, and the pace required to make this change requires open collaboration, debate, and open standards. • Google, Meta, Microsoft, NVIDIA are introducing new OCP workstreams to standardize: • Rack interface connectors • DC rack whips • DC tapbox • DC:DC power shelves • DC:DC power conversion modules • Asking for all consumers and producers of the LVDC technology ecosystem to partner on open high level specifications for interoperability and velocity. Join us! Call to Action & Creating Open Standards
p13
Thank You!