PDF 原檔：報告_OCP2026_NVIDIA_主題演講_AI工廠規模化_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_NVIDIA_主題演講_AI工廠規模化_20261015_p6.png |
p6 | 真資料圖 | MGX 1.2（第三代）規格重點：無線纜無水管運算托盤、液冷匯流排 5,000 A、45C 液冷盲插歧管、latch 可承 6,000 lb 配合力、+100 lb 鋼補強、UQD08 流率、機櫃實拍照 |
報告_OCP2026_NVIDIA_主題演講_AI工廠規模化_20261015_p8.png |
p8 | 真資料圖 | Visionbay.ai 首座 800 VDC AI 工廠：16 MW 部署單元由 4 個 4 MW 可擴充單元組成，端排集中轉換，800 VDC 匯流排架空配電，機櫃端 800 VDC 125 A tap can |
報告_OCP2026_NVIDIA_主題演講_AI工廠規模化_20261015_p11.png |
p11 | 真資料圖 | Vera CPU MGX 機櫃「Shipping now」：無線纜、模組化、100% 液冷，四種可客製托盤，列名 ODM／OEM logo（Foxconn、Compal、Pegatron、Wistron、Wiwynn、QCT、Gigabyte、MSI、ASUS、Inventec、MiTAC 等） |
報告_OCP2026_NVIDIA_主題演講_AI工廠規模化_20261015_p14.png |
p14 | 真資料圖 | Spectrum-X Photonics Multiplane：Spectrum-6 交換機剖面圖、ConnectX-9、BlueField-4；2.6 ms 故障偵測、40% 更少交換器、5x 能效、10x MTBI；Meta／Microsoft／OCI logo |
報告_OCP2026_NVIDIA_主題演講_AI工廠規模化_20261015_p15.png |
p15 | 真資料圖 | Spectrum-6 乙太網交換器供 Meta：Meta 102.4T FBOSS switch（N-IceCube）實物圖，基於 NVIDIA Spectrum-6；故障偵測比軟體快 30x |
報告_OCP2026_NVIDIA_主題演講_AI工廠規模化_20261015_p18.png |
p18 | 真資料圖 | Vera Rubin Ultra NVL576 CPO：Vera Rubin Ultra 運算托盤、NVLink 7 CPO 交換托盤，以及 8 座機櫃以黃色光纖在頂部串接的 NVL576 CPO 整體圖 |
原始內容
p1
AI Factories at Scale: Advancing Open Infrastructure for Agentic AI Ian Buck Vice President of Hyperscale and HPC Computing, NVIDIA
p2
OCP Global Summit | October 12, 2026 Ian Buck, VP of Hyperscale and HPC AI Factories at Scale: Advancing Open Infrastructure for Agentic AI
p3
History of OCP Contributions MGX 1.1 Blackwell Ultra 2025 ConnectX-7 adapter 2023 SwitchX-2 2013 ConnectX-3 2014 MGX 1.0 Blackwell rack 2024 HGX-H100 baseboard 2022 A100 OAM and baseboard v1.0 2020 V100 OAM form factor 2019 Tesla P100 HGX-1 baseboard Tesla P100 2017 ConnectX-4 Tesla M40 2015 Participating Projects Power Distribution IT Infrastructure Digital Twin Server MHS NIC Security Open Accelerator Data Center Facility Open Platform Firmware Open Data Centers / Systems for AI Time Appliance Project Hardware Managem ent Open Network Install Environment S/W for Open Network in the Cloud Switch Abstraction Interface Cooling Environme nt
p4
Built on OCP open standards, every model, every workload, every partner with appreciating value NVIDIA AI Factories Are Productive, Durable, and Fungible Year 1 Year 10 Every Model, Every Workload | Open, Closed, and Frontier 10M Developers | CSP, NCP, OEM, ODM, 300+ Supply Chain CUDA | Faster Every Release Rich Ecosystem Fungible Durable Accelerated Computing Pre-Training Post-Training Inference Agentic V100/DGX- 1 10 yrs in service A100 6 yrs old 9 yrs sold NVL72 1 yr old 7 yrs sold Vera Rubin Now in full production
p5
DeepSeek-V4-Pro | SemiAnalysis AgentX Full-Stack Co-design Delivers up to 30x Better AI Factory Throughput on Agentic Workloads 45x Lower $0.00 $0.10 $0.20 $0.30 $0.40 50 250 Cost per … Interactivity … Cost Per Million Tokens Disaggregated serving Rate matching Large-scale expert parallel Distributed KV-caching KV-cache offloading KV-aware routing Fused CUDA kernels Vera Rubin NVL72 MGX Rack GB300 NVL72 Vera Rubin NVL72 30x 0M 20M 40M 60M 80M 50100150200250300 Throughput … Interactivity … Total Token Throughput Per MW 10x 2x
p6
MGX 1.2 is a factory-scale, high-density, liquid cooled architecture designed for easy assembly and maintenance Announcing NVIDIA Contributes 3rd Generation MGX Specification to OCP + - Rack Reinforcements +100 lb. of steel support Compute Tray Bay System Highly configurable Rear Rack Extenders Protects manifold fittings Compute Tray DC-SCM Fits 1RU form factor Floating Blind Mate Reduces breakage Larger Rack Cable Tray Fits higher cable density Copper Spine 4 Cable Cartridges for NVLink or Spectrum-X Ethernet Blind Mate Manifold 45C Liquid Cooling Latch Ejector Bar Support 6,000 lb. mating force Switch Tray Form Factor 1U LC: NVLink or Spectrum-X Blind Mate Slide Rails NVLink and LC connections Liquid-Cooling Flow Rates UQD08 Cableless, Hoseless Compute Tray 100% liquid cooled, faster assembly Liquid-Cooled Busbar 5,000 amp capacity In-tray MQD/MQDB Easier tray servicing
p7
Vera Rubin: No cables, fans, or hoses 3rd Gen MGX Improves Manufacturability and Maintenance Switch Tray Vera Rubin NVL72 (3rd Gen OCP MGX 1.2) Compute Tray GB200 NVL72 (1st Gen OCP MGX 1.0)
p8
16 MW Deployment Unit Centralize conversion in redundant end-of-row power centers Upstream AC infrastructure 800 VDC 125 A rack-level tap cans 800 VDC busbar distributes power overhead to 4 MW scalable units SU 2 SU 3 SU 4 SU 1 Scalable power design for high-density, multi-gen AI infrastructure, aligned with OCP 800 VDC architecture Visionbay.ai Introduces the First 800 VDC AI Factory Engineered for the Next Decade of Accelerated Computing
p9
The max single-threaded CPU at scale NVIDIA Vera CPU for Agents Single-Threaded Core Performance Throughput Low High Low High Pre-cloud CPU Cloud CPU Max Single- Threaded CPU at Scale NEW NVIDIA Vera CPU 2x Performance NVIDIA Olympus Core 3.4 TB/s Core-to-core bandwidth Zero Chiplet Tax Scalable Coherent Fabric 1.2 TB/s Memory bandwidth per core LPDDR5X subsystem
p10
Built for sustained AI execution at scale NVIDIA Vera Delivers 2x Performance for Agentic AI 1.3x 1.8x1.8x1.8x1.9x2.0x2.0x2.1x2.2x2.2x NVIDIA Grace NVIDIA Vera 2x HPC Golden Suite TACC 2x Hyperscale Cloud Meta DCPerf 3.5x Apache Kafka Redpanda
p11
Partner ecosystem building customizable Vera OCP MGX racks Announcing NVIDIA Vera CPU MGX Rack Now Available Cableless, Modular, 100% Liquid Cooled Faster assembly and easier tray servicing Faster assembly and easier tray servicing Customizable Vera MGX Rack Shipping now Customizable across inference, agentic, storage, and more
p12
Grid resilience, more compute per megawatt, and simpler AI factory operations NVIDIA DSX Advances OCP-Compatible AI Factories NVIDIA Config Manager NVIDIA Cluster Readiness Engine NVIDIA Rack Management Service Power GPU nodes New DSX-compatible Spec Contributions Open data centers & grid resilience MaxLPS – 40% More GPUs in Same Power Turns stranded power to GPU capacity New DSX AI Factory Open Software Easier factory configuration & operations Stranded power Unlocked capacity
p13
37% more GB300 GPUs deployed within the same provisioned power budget (Kimi K2.5, 8K/1K, Iceland) Nscale Increases Throughput Nearly 50% with MaxLPS Nscale GB300 Cluster Nearly 50% higher throughput in same power 2… 2… Power Same Pow… 1 4… 1 9… GPUs 37% Mor… 1… 1… oughput (to 49% Hig… Baseline…
p14
Spectrum-X Multiplane Photonics enables AI factories to scale up to 512K GPUs with a flat, two-tier network NVIDIA Spectrum-X Photonics Multiplane Ethernet Uninterrupted AI Connectivity NVIDIA Spectrum-6 In deployment now ConnectX-9 SuperNICs BlueField-4 DPUs 40% Fewer switches 2.6 ms Fault detection 10x Higher MTBI 5x Better power efficiency
p15
Spectrum-X Multiplane collaboration for higher resiliency and smooth deployment at scale World’s fastest fault detection, 30x faster than traditional software-based fault detection NVIDIA open Ethernet silicon with OCP stack Open Networking for AI with FBOSS and OCP SAI Announcing Spectrum-6 Ethernet Switch for Meta’s AI Factory Based on NVIDIA Spectrum-6 Meta 102.4T FBOSS Switch (N-IceCube)
p16
Extends NVIDIA networking to Google data center transport Supported across NVIDIA Spectrum-X SuperNICs and NVIDIA BlueField DPUs Predictable performance at any fabric scale Silicon-to-software partnership for consistent performance Announcing Google Falcon Transport on NVIDIA Networking Multipath Load Balancing Programmable Congestion Control Spreads traffic, reduces bottlenecks Adapts traffic to reduce delays Hardware Loss Recovery Per-Connection Security Keeps data moving Protects data in transit
p17
One accelerated Ethernet platform supporting multiple AI transports Spectrum-X Ethernet Extends Open Networking Leadership Falcon Meta RoCE MRC Spectrum-X Adaptive RDMA Interoperability with Google’s data center transport RoCE networking for Meta’s AI infrastructure Multipath delivery for resilient AI communication Adaptive routing for efficient AI data movement
p18
Vera Rubin Ultra Compute Tray NVLink 7 CPO Switch Tray Designed for performance per watt for 10T+ parameter models NVL576 Is Built for Breakthrough Scale and Efficiency Vera Rubin Ultra NVL576 CPO
p19
OCP Talks—33 Total Sessions Join Us at OCP Session Date, Time, and Location AI Tokenomics Explained Shruti Koparkar and Eduardo Alvarez, NVIDIA Tuesday, Oct 13, 2:30 p.m. | Concourse Level, 220B 800 VDC AI Factories: From Arch. Alignment to Indust. Deployment Mike Tu and Jared Huntington, NVIDIA Tuesday, Oct 13, 4:10 p.m. | Concourse Level, 210CG Flexibility the Grid Can Count On: Turn Curtailable Load to Capacity Sean James, NVIDIA Tuesday, Oct 13, 4:35 p.m. | Concourse Level, 210CG Revolutionizing Networking for Gigawatt AI Factories Gilad Shainer, NVIDIA Wednesday, Oct 14, 9:15 a.m. | Concourse Level, 210CG Next-Gen ML Infrastructure: The Open DC Architecture Ayan Majumdar and Heather Huang, Google | Mike Sabotta, NVIDIA | Paul Artman, AMD | Mike Edie, Meta Thursday, Oct 15, 12:30 p.m. | Concourse Level, 210CG The AI Factory Digital Twin: Full Stack Tokens/Watt Optimization Kourosh Nemati, Tony Marfil and Usha Kodali, NVIDIA | Sherman Ikemoto and Matthew Käufeler, Cadence Thursday, Oct 15, 3:35 p.m. | Concourse Level, 210CG
p20
Nearly 100 partners on display 800 VDC power technology Next generation innovations MGX Ecosystem
p21
（本頁無可抽取文字，內容在圖中）