PDF 原檔：報告_OCP2026_神達_先進散熱_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_神達_先進散熱_20261015_p5.png |
p5 | 真資料圖 | 神達×AMD 機櫃級產品線：AI 液冷櫃（MI355X）、AI 氣冷櫃（MI350X）、儲存氣冷櫃、HPC 液冷／氣冷櫃（EPYC 9005）五種機櫃外觀 |
報告_OCP2026_神達_先進散熱_20261015_p14.png |
p14 | 裝飾·logo·banner | 熱管理生態系三項選擇：冷板 JWS／Cooler Master、冷卻液 Castrol／PFX、AI 風格宣傳圖、無數據；架構 單相（CoolIT、Nidec）／兩相（ZutaCore）／鑽石冷卻（Akash） |
報告_OCP2026_神達_先進散熱_20261015_p16.png |
p16 | 真資料圖 | JWS 微通道冷板 GPU PQ 曲線（40°C 實線 vs 模擬 HSSVC@45°C 虛線，流量 0～3.5 LPM，縱軸標 Power (kW)，圖為示意風格；左為冷板渲染圖），標示 1.5 LPM 附近壓降降 44% |
報告_OCP2026_神達_先進散熱_20261015_p17.png |
p17 | 真資料圖 | Castrol PG25 冷卻液在 45°C 供液溫度下，五種 R:C 混合比例的流量（3.5～5.5 LPM）對 ΔT（約 8.5 降至 5.5～6°C）曲線，五條幾乎重疊 |
報告_OCP2026_神達_先進散熱_20261015_p18.png |
p18 | 文字卡 | Akash 鑽石冷卻 G8825Z5 AMD 8 GPU 伺服器：標示免降頻、GPU 溫度最多降 10°C、冷卻耗電最多省 100%、FLOPs/Watt 最多 +22%（約 75°F 環境）、token 吞吐最多 +15%（約 120°F 環境） |
報告_OCP2026_神達_先進散熱_20261015_p19.png |
p19 | 真資料圖 | ZutaCore 兩相 DLC 三步驟（蒸發、輸送、冷凝）與實測成果：>4,000W 晶片、>250 W/cm²、年化 pPUE 1.03、冷卻能耗較單相降 74%、蒸氣輸送約 0.3 L/min per kW |
原始內容
p1
MiTAC Confidential All third-party brands and names are the property of their respective owners. MiTAC Computing Technology Corp. 1 Raymond Huang General Manager, MiTAC Computing Technology USA Sachin Dhingra Sr. Director, Enterprise AI Solutions, AMD Powering Future Compute | www.mitaccomputing.com Scaling the Future of Open AI Infrastructure with Advanced Cooling Oct.14, 2026
p2
2
p3
3 Enabling Open Standard Hardware Architecture with MiTAC OCP Multi-node Servers Type Air-cooled Liquid-cooled Air-cooled Picture Model Name C2810Z6/ C2610Z6 (AMD EPYCTM 9006) C2611Z6 (AMD EPYCTM 9006) Capri 3 Ultra Capri 3 Standard (AMD EPYCTM9005/9004) Form Factor 2OU2N 2OU4N 2OU3N
p4
4 MiTAC Diversified AI Infrastructure Solutions with AMD 2U 2U 4U 8U MiTAC B8050/ MiTAC TN85-B8261 MiTAC TN85-B8261 MiTAC G8825Z5 (Air-cooled) MiTAC G4826Z5 (Liquid-cooled) MiTAC Rack-scale Solutions • Power-efficient, air-cooled for distributed AI inference • Lower TCO for GenAI at the edge • Flexible enterprise AI inference • High performance with efficient agent logic & multi- tasking • GPU intensive AI compute • Air-cooled • Massive memory & high-speed NVMe storage • Extreme-scale AI compute • Liquid-cooled • High CPU core density, high memory & ultra-fast storage • AI liquid-cooled rack • AI air-cooled rack • HPC liquid-cooled rack • HPC air-cooled rack Workstation/Server Server AI-Optimized Server AI-Optimized Server Rack AMD Radeon AI PRO R9700S AMD Instinct MI350P (PCIe) AMD Instinc MI350X x8 (UBB) AMD Instinct MI355X x8 (UBB) AMD Instinct MI355X/MI350X x8 (UBB) • Edge AI, retail • Image generation • Traditional AI (CNN, DLRM) • SLM inference • Mainstream LLM inference • Agentic AI • Content generation • RAG • LLM inference • Model fine tuning • SLM training • Code generation • LLM inference • Model fine tuning • SLM training • Code generation • Hyperscaler/CSP inference • Frontier model training • Low latency AMD EPYC 9005 Series and Next-gen CPU
p5
5 MiTAC Rack-Scale Solutions with AMD
p6
6 Open Ecosystem for AMD AI Infrastructure From Open-Source FW to Full AI SW Stacks for AMD EPYCTM and InstinctTM MI355X Platforms
p7
7 | AMD ENTERPRISE AI High-Performance Computing Leadership Leadership Foundational IP Broad Product Portfolio Open, Proven Software Ecosystem Deep, Collaborative Partnerships
p8
8 | AMD ENTERPRISE AI AMD Enterprise AI Solutions Seamless Enablement of High Value Enterprise Applications High Performance Platforms Delivering Performance and TCO value; On-prem & Cloud Enterprise AI Software Complete solutions for deploying & customizing industry-standard AI models Strategic Partnerships Enable high-impact use cases optimized for AMD platforms LLAMA Stack Complete Solutions = Hardware + Software + ISV partners
p9
9 | AMD ENTERPRISE AI LOCAL/ AGENTIC AMD Ryzen AI Halo Build local AI agents 1x users USER SCALE 192 GB MEMORY 300B MODEL SIZE Desktop FORM FACTOR AIR | 120W SYSTEM POWER EDGE/ PROTOTYPE Radeon AI PRO R9700 Prototype. Deploy. Repeat 100x users USER SCALE 32 GB MEMORY 35B MODEL SIZE Wkstn/Server FORM FACTOR AIR | 1-15kW SYSTEM POWER SCALE/ ENTERPRISE Instinct MI350P Scale Enterprise inference 1,000x users USER SCALE 144 GB MEMORY 175B MODEL SIZE Server FORM FACTOR AIR | 15-30kW RACK POWER INDUSTRIALIZE/ AI FACTORY Instinct MI355X x8 Build the AI factory 10,000x users USER SCALE 2.3 TB MEMORY 1T+ MODEL SIZE AI Server FORM FACTOR AIR / LC | 30- 60kW RACK POWER Start with the use case. Match the platform. Scale on one AMD software foundation. Edge AI Retail AI Image Gen SLM Inference RAG Chat Copilots Content Gen Production LLM Large LLM Fine-Tuning Code Gen Training Agent Apps Sensitive Docs Personal Assistant Code Edits HYPERSCALE/ RACKSCALE AMD Helios 72× MI455X GPUs 100,000+ users USER SCALE 31 TB MEMORY
1T MODEL SIZE Rack FORM FACTOR LC | 100kW+ RACK POWER Hyperscale AI Sovereign AI Frontier Train Low-Latency Deploy the Right AMD Platform for Every Enterprise Workload
p10
10 | AMD ENTERPRISE AI AMD Enterprise AI Software Stack Removes underlying complexity, making AI deployments simple AMD Resource Manager AMD AI Workbench Ready-to-deploy templates that combine AI models and reference code to accelerate AI application development Fast, scalable and secure deployment of optimized AI models with industry- standard APIs Complete toolchain to develop AI models and deploy for real-time inference at scale Managing compute resources and quotas for efficient compute utilization across AI initiatives AMD Solution Blueprints AMD Enterprise AI Reference Stack AMD Inference Microservices
p11
11 MiTAC Diversified AI Infrastructure Solutions with AMD 2U 2U 4U 8U MiTAC B8050/ MiTAC TN85-B8261 MiTAC TN85-B8261 MiTAC G8825Z5 (Air-cooled) MiTAC G4826Z5 (Liquid-cooled) MiTAC Rack-scale Solutions • Power-efficient, air-cooled for distributed AI inference • Lower TCO for GenAI at the edge • Flexible enterprise AI inference • High performance with efficient agent logic & multi- tasking • GPU intensive AI compute • Air-cooled • Massive memory & high-speed NVMe storage • Extreme-scale AI compute • Liquid-cooled • High CPU core density, high memory & ultra-fast storage • AI liquid-cooled rack • AI air-cooled rack • HPC liquid-cooled rack • HPC air-cooled rack Workstation/Server Server AI-Optimized Server AI-Optimized Server Rack AMD Radeon AI PRO R9700S AMD Instinct MI350P (PCIe) AMD Instinc MI350X x8 (UBB) AMD Instinct MI355X x8 (UBB) AMD Instinct MI355X/MI350X x8 (UBB) • Edge AI, retail • Image generation • Traditional AI (CNN, DLRM) • SLM inference • Mainstream LLM inference • Agentic AI • Content generation • RAG • LLM inference • Model fine tuning • SLM training • Code generation • LLM inference • Model fine tuning • SLM training • Code generation • Hyperscaler/CSP inference • Frontier model training • Low latency AMD EPYC 9005 Series and Next-gen CPU
p12
12 | AMD ENTERPRISE AI AMD Instinct MI350P The Easiest Way to Add AI to Your Data Center STANDARD SERVER COMPATIBLE PERF/$ LEADERSHIP LARGE MODEL CAPABLE Up to 260 Billion Air Cooled Up to 4.2x Tokens/second/$ vs RTX Pro 6000 Model parameters @ FP4 450-600W TBP MI350P sweet spot: private, memory-intensive production inference in familiar enterprise servers.
p13
13
p14
15
p15
16 AI Server GPU Cold Plate Loop
p16
17 Minimizing pressure drop by 44% optimizes CDU power requirements and enhances overall heat dissipation efficiency.
p17
18
p18
19
p19
20
p20
21 Scaling Open AI Infrastructure Open by Design. Cooled to Scale.
p21
22