PDF 原檔：報告_OCP2026_AMD_主題演講_AI基礎設施新世代_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_AMD_主題演講_AI基礎設施新世代_20261015_p7.png |
p7 | 真資料圖 | AMD 內部預估 TAM：資料中心 CPU TAM 由 2025 約 $26B 增至 2030 約 $220B（>50% CAGR，分一般用途／AI host／Agentic）；AI 加速器 TAM 由約 $200B 增至約 $1.4T（>45% CAGR，分訓練／推論） |
報告_OCP2026_AMD_主題演講_AI基礎設施新世代_20261015_p13.png |
p13 | 真資料圖 | Helios 網路分層：Front-end（安全／儲存／SDN）、Scale-up（UALoE、ESUN）、Scale-out（UEC、MRC、MetaRoCE）；列名 Broadcom、Cisco、Arista、HPE、DriveNets |
報告_OCP2026_AMD_主題演講_AI基礎設施新世代_20261015_p15.png |
p15 | 真資料圖 | Helios 機櫃規格：2.9 EF FP4、1.48 EF FP8、31 TB HBM4、1.7 PB/s HBM4 頻寬、260 TB/s scale-up、43 TB/s scale-out；up to 30% more tokens/$；DeepSeek R1 MLPerf 17,000 tokens/s/GPU |
報告_OCP2026_AMD_主題演講_AI基礎設施新世代_20261015_p17.png |
p17 | 真資料圖 | Helios 部署：OpenAI／Meta／Anthropic／Microsoft／Oracle logo，三張機房實拍（黑色機櫃、橘黃走線與紫色線纜） |
報告_OCP2026_AMD_主題演講_AI基礎設施新世代_20261015_p24.png |
p24 | 真資料圖 | Helion Agentic AI Rack 規格：ORv3 44OU、up to 32,768 Zen 6 核、up to 16 MI350P（74 PF FP4）、up to 102 Tbps 網路、標準模組化運算板 |
報告_OCP2026_AMD_主題演講_AI基礎設施新世代_20261015_p27.png |
p27 | 裝飾·logo·banner | Helion Agentic Rack「Available November '26」並貢獻 OCP；列名 Dell／HPE／Lenovo／Supermicro／Cisco／Wiwynn／QCT／Foxconn／MiTAC／Inventec／ASUS／ASRock Rack／Gigabyte logo |
原始內容
p1
AMD Keynote Dr. Lisa Su
p2
Solving The World’s Most Important Challenges Accelerating AI from Silicon to Solutions
p3
15 Years of Collaboration 2012 2017 2023 2024 2025 2026 Open Server Platforms Open AI Systems Open AI Infrastructure AMD OPEN 3.0 MODULAR SERVER SPECIFICATION “NAPLES” + MICROSOFT PROJECT OLYMPUS ULTRA ETHERNET CONSORTIUM FORMS MULTIPATH RELIABLE CONNECTION YOSEMITE V4 ULTRA ACCELERATOR LINK CONSORTIUM FORMS AMD HELIOS DEBUTS AT OCP ESUN COLLABORATION FORMS OPEN ADAPTER CARD ESUN
p4
HEALTHCARE MANUFACTURING FINANCE EDUCATION SCIENCE SOFTWARE AGRICULTURE MEDIA & ENTERTAINMENT AI Is Transforming Every Industry
p5
Demand for AI Continues to Accelerate Increase in tokens per month consumed over past 2 years 95x
p6
DeepSeek V4 Pro But Inference is Now the #1 AI Workload TRAINING FLOPS INCREASING 5X EVERY YEAR SINCE 2020 Training Continues to Scale DRIVEN BY ACCELERATED AGENTIC ADOPTION Source: Epoch AI, Trends in Artificial Intelligence, October 5, 2026 1027 1026 1025 1024 1028 Claude 3.7 Sonnet Claude 3.5 Sonnet GPT-4 Mistral Large 2 Llama 3.1 405B Grok 3 Gemini 1.0 Ultra GPT-4.5 Grok 4 Llama 4 Behemoth GPT-3.5 Mistral Large Claude 2 GPT-5 Grok 2 GPT-6 Astra DeepSeek V 4.1 Flash Kimi K3 Source: Openrouter.ai/ratings 7.5 T 6.5 T 5.5 T 4.5 T 3.5 T 2.5 T 1.5 T 0.5 T Agentic Human Mixed
14x Growth YTD
p7
Demand Accelerating Across CPUs and GPUs ~$200B ~$1.4T
45% CAGR DATA CENTER AI ACCELERATOR TAM ($B) Inference Training ~$220B 50% CAGR AI Host General Purpose Agentic DATA CENTER CPU TAM ($B) ~$26B
p8
Strategy Powering AI Everywhere Open Platforms Compute Leadership
p9
Rack-Scale Innovation Agentic Infrastructure Collaboration at Scale
p10
Frontier AI Requires Integrated Rack Architectures Rack Scale Efficiency & Serviceability Open Fabric to Scale AI Open Rack Architecture Leadership Compute Capability ESUN Turnkey Solutions
p11
The Highest Performance Rack-Scale AI Solution
p12
Helios: High-Performance Rack Co-Engineered with AI Leaders Open Rack Wide OCP Power Shelves OCP GPU Management
p13
Helios: High-Performance Networking Delivered Through An Open Ecosystem Open Performant Programmable FRONT-END Security, Storage, SDN Acceleration SCALE-UP UALoE, ESUN SCALE-OUT UEC, MRC, MetaRoCE
p14
ROCm: High-Performance Software Powered By An Open Ecosystem VERL OPEN SOURCE COMMUNITY ENABLING AMD
p15
DELIVERS On leading AI Models Up to 30% More Tokens/$ DeepSeek R1 Inference (MLPerf) 17,000 tokens/s/GPU AMD Helios Delivers Leadership Rack-Scale Solution 2.9 Exa Flops FP4 1.48 Exa Flops FP8 260 TB/s Scale Up Bandwidth 43 TB/s Scale Out Bandwidth COMPUTE 31 TB HBM4 Capacity 1.7 PB/s HBM4 Bandwidth MEMORY NETWORKING
p16
Helios Deployments Underway at Leading AI Customers VIDEO
p17
Helios Deployments Underway at Leading AI Customers
p18
AMD Helios Growing Adoption by AI Leaders
p19
Rack-Scale Innovation Agentic Infrastructure Collaboration at Scale
p20
Entering an Era of High-Volume Agentic Usage 10,000 Agents to solve the Navier–Stokes Millennium Prize Problem 30,000 R&D Agents building the next Frontier Model 5 Million Downloads of Meta Muse in first few weeks
p21
Agents Need the Right Compute TECHNICAL COMPUTE AGENTS Performance Servers (Max Agent Performance) Balanced Servers (Max # of Performant Agents) Dense Servers (Max Agents per Watt) PERSONAL AGENTS ENTERPRISE AGENTS
p22
And Agents Also Need the Right Rack ENTERPRISE All-in-One Agentic Rack HYPERSCALE Disaggregated Agentic Rack
p23
AMD Helion Agentic AI Rack Open & Flexible Rack-Scale Compute for Agentic AI Match your CPU compute to your agentic workloads Customize your mix of CPU core density, performance and power Choose any combination of CPU & GPU nodes CPU-only for hyperscale deployments & all-in-one CPU/GPU Agentic racks for enterprises Integrated token router Intelligently route your agents model calls across frontier and open source models Open hardware & open software OCP compliant hardware components ANNOUNCING TODAY
p24
ORv3 44OU Rack Native OCP Spec Compliant Up to 16 MI350P 74 PetaFlops - FP4 Up to 102 Tbps Networking Fully Programmable Standard Modular Compute Board Native OCP Spec Compliant Up to 32,768 Zen 6 x86 CPU Cores Open-Source Ecosystem Stack AMD Helion Agentic AI Rack
p25
AMD Helion Agentic Rack – Hyperscale Config Rack-Scale Agentic Infrastructure (CPU Only) Mix and Match CPU Nodes to Support Diverse Agentic Workloads PERSONAL AGENTS Dense Servers “Venice” 256c @ 400W ENTERPRISE AGENTS Balanced Servers “Venice” 256c @ 600W TECHNICAL AGENTS Performance Servers “Venice” 96c High Frequency
p26
AGENTIC NODE – CPU “SANDBOXES” Dense – “Venice” 256c @ 400W Performance – “Venice” 96c @ 5.0GHz INFERENCE NODE – GPUs Performance - Instinct MI350P GENERAL PURPOSE NODE Database, Storage – ”Venice” SP7, SP8 Technical Compute – ”Venice-X” AMD Helion Agentic Rack – Enterprise Config All-in-One Enterprise Agentic AI Rack (CPU+GPU)
p27
AMD Helion Agentic Rack Open & Flexible Rack-Scale Compute for Agentic AI AVAILABLE NOVEMBER ‘26 CONTRIBUTING TO
p28
EPYC “Venice” The Best Server CPU Portfolio for Agentic Racks High Performance High Density Technical Compute & HPC High Core Count Low Power Balanced Performance & Power High Core Count Low Power GENERAL PURPOSE CPU SERVERS AGENT CPU SERVERS “SANDBOXES” “Venice” 256c “Venice” SP7 “Venice” SP8 “Venice-X” GPU SERVERS “HOST NODE” “Venice” HF High Frequency CPU to GPU Bandwidth “Verano” RDIMM / MRDIMM LPDDR
p29
New EPYC “Venice” with LPDDR Lowest System Power Fastest CPU to GPU Bandwidth Up to 192c AVAILABLE IN 1H 2027 “Verano” Highest Density Highest Perf/Core Balance Perf/Power Venice 256c Venice HF Venice SP8
p30
Open Hardware Open Software Open Ecosystem Open Ecosystem Drives Innovation ESUN VERL
p31
Delivering the Next Era of AI Infrastructure Collaboration at Scale Leadership Components Flexible, Open Platforms Integrated Rack Solutions
p32
（本頁無可抽取文字，內容在圖中）
p33
（本頁無可抽取文字，內容在圖中）