報告_OCP2026_廣達QCT_AI工廠方案_20261015

PDF 原檔：報告_OCP2026_廣達QCT_AI工廠方案_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

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L Jack Luoh Head of Product Marketing, QCT/Quanta

p2

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L Why AI Is Processing More Tokens Per Task? TOKENS The fundamental unit of AI processing The units AI models process and generate — from understanding inputs to reasoning and producing outputs. 1 2 3 Reason More Act More Remember More Multi-Step Thinking Agentic AI Long Context AI thinks through multiple steps before producing the final answer, especially for complex tasks. AI agents repeatedly reason, call tools, receive results, and continue until a task is completed. AI needs to process longer conversations, documents, and historical information across interactions. More reasoning tokens per task More model calls & tokens per task More input tokens to process

p3

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L 1 2 3 Reason More Act More Remember More Reasoning Agentic AI Long Context AI thinks through multiple steps before producing the final answer, especially for complex tasks. AI agents repeatedly reason, call tools, receive results, and continue until a task is completed. AI needs to process longer conversations, documents, and historical information across interactions. More reasoning tokens per task More model calls & tokens per task More input tokens to process From More Tokens to Massive AI Compute Infrastructure Global AI chip sales translate to rapidly expanding compute capacity Source: EPOCH AI Dataset Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 2022 2023 2024 2025 2026

p4

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L AI Growth Is Outpacing the Power Infrastructure Behind It Frontier AI power requirements are rising rapidly, while grid expansion struggles to keep pace. Source: EPOCH AI Dataset ~31 GW Global AI data center power capacity by Q4 2025 Comparable to peak power usage in New York State, and outstripping many developed countries. Source: EPOCH AI Dataset ~40% of facility power goes directly to GPUs • Total server power is 1.53× GPU power alone. • IT equipment uses 1.14× as much power as the servers. Source: EPOCH AI Dataset Grid expansion struggles to keep pace ~268 GW Power supply gap by 2030 Total data center power demand Grid-deliverable capacity 490.7 GW 222.6 GW ~268 GW supply gap Source: EPOCH AI Dataset

p5

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L Drives demand for Requires power to operate The focus is no longer just More Compute but how to generate More Tokens Per Watt Resource Demand Expanding infrastructure requirements More Tokens → More Compute COMPUTE Processing Resource ~30M H100e Cumulative AI compute capacity by 2026 COMPUTE Source: EPOCH AI Dataset Physical Limit Constrained by grid and site infrastructure More Compute → More Power POWER Physical Infrastructure & Constraint ~31 GW Global AI data center power capacity by Q4 2025 Source: EPOCH AI Dataset AI Demand More sophisticated workloads More AI Workloads → More Tokens TOKENS AI Output 300T+ / DAY AI tokens already disclosed Source: Tokens Per Day As AI Demand Scales, Power Becomes the Constraint

p6

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L QCT collaborates with the DSX ecosystem to optimize tokens per watt from AI Factory perspective Tokens-per-Watt Optimization with NVIDIA DSX DSX AIDC Cooling system Power system MaxLPS DSX Flex DSX Exchange DSX Sim Manufacturing Healthcare Education & Research Logistics Entertainment Transportation Industrial Workloads Compute Storage Network QCT AI Cluster Infrastructure Software POD-Level Efficiency 1 Deliver software-defined systems tailored to workload requirements. Infrastructure- Level Optimization 2 Integrate cooling and power systems to optimize data center design and operations.

p7

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L POD-Level Efficiency: Optimized for Every AI Workload POD-Level Efficiency 1 Match each AI workload with the right compute architecture Right Workload Right Architecture X Optimize performance REASON Agentic AI & Reinforcement Learning Vera CPU Rack • High-performance CPUs • Optimized for agentic AI and RL • Scalable and flexible architecture Purpose-built architectures to deliver higher performance and efficiency across diverse AI workloads. ACCELERATE Large-scale AI Compute Vera Rubin NVL72 • Highest AI compute density • Multi-GPU at scale • Optimized for training and advanced reasoning REMEMBER Context & Memory Vera BlueField-4 STX Storage • High-capacity memory • Optimized for long context • Efficient data access and retrieval REAL-TIME RESPONCE High-throughput Inference Groq 3 LPX • Low-latency inference • High throughput • Optimized for real- time and large-scale serving CONNECT High-speed AI Fabric Spectrum-6 SPX Networking • Scalable rack-to-rack connectivity • High bandwidth, low latency • Optimized for AI clusters

p8

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L POD-Level Efficiency: Optimized for Every AI Workload POD-Level Efficiency 1 Match each AI workload with the right compute architecture Right Workload Right Architecture X Optimize performance Purpose-built architectures to deliver higher performance and efficiency across diverse AI workloads. ACCELERATE Large-scale AI Compute Vera Rubin NVL72 • Highest AI compute density • Multi-GPU at scale • Optimized for training and advanced reasoning REAL-TIME RESPONCE High-throughput Inference Groq 3 LPX • Low-latency inference • High throughput • Optimized for real- time and large-scale serving CONNECT High-speed AI Fabric Spectrum-6 SPX Networking • Scalable rack-to-rack connectivity • High bandwidth, low latency • Optimized for AI clusters Beyond accelerating AI, new architectures are emerging to optimize how AI reasons and remembers. REASON Agentic AI & Reinforcement Learning Vera CPU Rack • High-performance CPUs • Optimized for agentic AI and RL • Scalable and flexible architecture REMEMBER Context & Memory Vera BlueField-4 STX Storage • High-capacity memory • Optimized for long context • Efficient data access and retrieval

p9

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L Accelerate Agentic AI Workflows QuantaPlex T46X-1U Independent CPU Scaling Scale CPU capacity independently from GPU compute, adding execution resources where agentic workloads need them. Massive Parallel Execution Support 22,500+ concurrent CPU environments for sandbox execution, tool use, and RL evaluation. High-Density CPU Deployment Pack massive CPU capacity into a dense rack-scale architecture, maximizing compute within the data center footprint. Vera CPU Rack Accelerate Execution & Generate More Tokens High-performance, energy-efficient CPU platform for agentic AI.

p10

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L Vera Bluefield-4 STX Storage Reuse Context & Generate More Tokens. AI-native context memory infrastructure for efficient inference. Efficient Context Memory for AI Efficient Context Reuse Shared context memory enables KV cache reuse across AI workloads. Reuse KV Cache Avoid redundant GPU computation to improve compute efficiency. Reduce Recomputation Efficient Context Access Prestage KV cache before decoding to reduce GPU stalls and idle time. Faster Context Access Offload context from GPU memory to free valuable HBM for AI compute. Free GPU Resources QCT collaborates with leading storage solution providers to deliver AI-native storage solutions QuantaPlex S26F-2U

p11

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L Beyond System-Level Efficiency: Addressing the Remaining 29% 41% 22% 8% 29% AI Data Center Power Distribution GPU Other Server Components Networking & Other IT Cooling & Facility Overhead Estimates are for a data center using GB200 NVL72 chips, at peak operation Source: EPOCH AI POD-Level Efficiency 1

p12

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L Beyond System-Level Efficiency: Addressing the Remaining 29% 41% 22% 8% 29% AI Data Center Power Distribution GPU Other Server Components Networking & Other IT Cooling & Facility Overhead Estimates are for a data center using GB200 NVL72 chips, at peak operation Source: EPOCH AI POD-Level Efficiency 1 Infrastructure- Level Optimization 2

p13

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L QCT Supports Infrastructure-level optimization QCT collaborates with ecosystem partners to support the end-to-end AI Factory journey and optimize tokens per watt. Infrastructure- Level Optimization 2 • Utilizes simulation result to streamline construction processes. • Ensure system operation to meet with digital twins for customer acceptance. • Enables continuous operational feedback and traceability for AI Factory management. • Collect AI factory capacity, power & density requirements and assemble required SimReady assets. • Build AI factory digital twins and validate design with simulation tools. DSX SIM, Omniverse, MaxLPS, DSX Exchange, DSX Flex Design Construction Operation Delivery Deployment DC Bring-up Optimization & Maintenance Collection Simulation Report

p14

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L lT, Facility Data Continuous Improvements Virtual Environment Simulate, analyze and validate in the virtual environment Thermal & Airflow Simulation Analyze temperature, detect hotspots and optimize cooling Power & Energy Analysis Evaluate power capacity, PUE and energy efficiency ◇ Configuration & Layout Adjust rack and equipment configurations Real World Green Field / Brown Field Data Center Physical Systems & Equipment Racks, servers, cooling and power systems Ecosystem Partners Provide equipment models, parameters and real-world data Streamline AI Factory Innovations with Digital Twins By leveraging digital twins for simulation and validation, QCT optimize AI Factory performance and accelerate data center deployment

p15

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L QCT AI Factory Based on NVIDIA DSX Together with ecosystem partners, QCT enables a Neocloud customer to reduce execution risks, accelerate AI Factory deployment, scale out faster, and optimize performance Power Digital Twins with IT & Facility Data Provide data center inputs for digital twin simulations, reducing AI Factory design iterations and accelerating deployment. Reduce Deployment Risk through Simulation Minimize deployment risk through pre-deployment simulation and validation. Optimize AI Factory Performance Continuously synchronize data with the digital twin to optimize AI factory performance. . Scale Faster with Modular Architecture Validate modular designs with ecosystem partners to ensure standardized infrastructure and future scalability.

p16

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L AI Factory Full-stack NVIDIA AI infrastructure platform for large-scale AI training and accelerated workloads. End-to-end NVIDIA platform Large-scale AI Heterogeneous Systems Flexible system integration with different platforms to support diverse AI workloads. Flexible Workload-Driven Different Partners. to AI Infrastructure. Choose the path that best accelerates your AI journey.

p17

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L QCT AI POD: Cluster-Level Infrastructure Optimized for Diverse AI Workloads LLM Inference Generative AI Inference with RAG LLM Fine-tune Training Computer Vision Deep Learning Training Agentic AI AI Reasoning Trillion-parameter LLM Training Choose The Right Solution Package Tailored to Your Workloads Needs NVAIE For Enterprise QCT SW Stack DSX OS For Enterprise For NCP Platform Software Orchestration and Scheduling Infrastructure Software Infrastructure Provisioning & System Management Rapid Deployment System Management Realtime Monitoring Development Environment System Provisioning Workload Orchestration AI Application QCT AI Cluster Infrastructure Storage S55J-2U General server D55Q-2U 8 GPU server D75T-7U Intel General server S44N-1U NVIDIA 8 GPU server D75H-10U 4/8 GPU AI server D75E-4U AMD 8 GPU server D76T-2U

p18

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L Workload LLM Inference Generative AI Inference with RAG Computer Vision Lighter AI Workloads Compute Storage Management Management node Capacity Node 4-GPU / 8-GPU D75E-4U - Unlock Infinite AI Possibilities Intel GNR-2P D55Q-2U Intel SRF/GNR-2P -High Flexibility for Expansion GPU-accelerated compute for inference, fine-tuning, perception & visual computing Manage, monitor and orchestrate AI resources for reliable and efficient operations Keep datasets, models and application data continuously available to AI compute S55J-2U Intel SRF/GNR-1P - Flexible, Ultra-Dense Architecture Built on Platform Build the Right AI POD for Your Workloads Intel - Lighter AI Workloads: Right-sized infrastructure for efficient inference and everyday AI workloads.

p19

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L Workload Compute Storage Management Management node Capacity Node 4-GPU / 8-GPU D75E-4U - Unlock Infinite AI Possibilities Intel GNR-2P D55Q-2U Intel SRF/GNR-2P -High Flexibility for Expansion GPU-accelerated compute for inference, fine-tuning, perception & visual computing Manage, monitor and orchestrate AI resources for reliable and efficient operations Keep datasets, models and application data continuously available to AI compute S55J-2U Intel SRF/GNR-1P - Flexible, Ultra-Dense Architecture Deep Learning Training Agentic AI LLM Fine-tune Training Advanced AI Workloads AI Reasoning HGX Platform D75H-10U - Enabling Diverse AI Workloads Intel GNR-2P Built on Platform Build the Right AI POD for Your Workloads Intel - Advanced AI Workloads: High-performance infrastructure for compute-intensive AI training and reasoning.

p20

Q C T C O N F I D E N T I A L Q C T C O N F I D E N T I A L Workload Compute Storage Management GPU-accelerated compute for inference, fine-tuning, perception & visual computing Manage, monitor and orchestrate AI resources for reliable and efficient operations Keep datasets, models and application data continuously available to AI compute D75H-10U - Enabling Diverse AI Workloads Intel GNR-2P HGX Platform Management node D55Q-2U Intel SRF/GNR-2P -High Flexibility for Expansion Capacity Node S55J-2U Intel SRF/GNR-1P - Flexible, Ultra-Dense Architecture Deep Learning Training Agentic AI LLM Fine-tune Training Advanced AI Workloads AI Reasoning Built on Platform D75T-7U -Modern AI-HPC AMD Turin-2P 8-GPU Capacity Node S44NL-1U AMD Genoa/Turin-1P - Light Weight and Cost-Effective S44N-1U AMD Genoa/Turin-2P -Compute intensive Management node Build the Right AI POD for Your Workloads AMD - Advanced AI Workloads: Flexible GPU infrastructure for scalable AI training and HPC workloads.

p21

Q C T C O N F I D E N T I A L 21 OPTIMIZE EVERY WATT POD-level efficiency More AI output from every watt OPTIMIZE THE AI FACTORY Infrastructure-level optimization Power, cooling, networking & operations BUILD AROUND REAL WORKLOAD Flexible AI POD solutions Right platform. Right scale. Right workload. from optimized systems and AI Factory infrastructure to workload-ready AI solutions.

p22

Q C T C O N F I D E N T I A L 22 B1

p23

AI 5G w w w . q c t . i o Thank You