Slide 1: Cautionary Statement
CAUTIONARY STATEMENT This presentation contains forward-looking statements concerning Advanced Micro Devices, Inc. (AMD) such as the features, functionality, performance, availability, timing and expected benefits of AMD products, product roadmaps, partnerships and collaborations; AI demand; AMD’s t
Slide 2: AI Is Transforming Every Industry
AI Is Transforming Every Industry MEDIA & HEALTHCARE MANUFACTURING FINANCE EDUCATION SCIENCE SOFTWARE AGRICULTURE ENTERTAINMENT
Slide 3: Demand for AI Continues to Accelerate
Demand for AI Continues to Accelerate 158x Increase in tokens per month consumed over past 2 years
Slide 4: Training Continues to Scale
Training Continues to Scale TRAINING FLOPS INCREASING 5X EVERY YEAR SINCE 2020 1028 1027 Grok 4 GPT-4.5 Grok 3 1026 Llama 3.1405 B GPT-5 Gemini 1.0 Ultra Llama 4 Behemoth Claude 3.5 Sonnet Claude 3.7 Sonnet Grok 2 GPT-4 Mistral Large 2 1025 Mistral Large GPT-3.5 Claude 2 1024
Slide 5: Training Continues to Scale But Inference is Now the AI Workload
Training Continues to Scale But Inference is Now the TRAINING FLOPS INCREASING AI Workload 5X EVERY YEAR SINCE 2020 TRAINING VS INFERENCE SPLIT 1028 40% 50% 60% 1027 Grok 4 GPT-4.5 Grok 3 1026 Llama 3.1405 B Gemini 1.0 Ultra GPT-5 Llama 4 Behemoth 60% Claude 3.5 Sonnet GPT-4 Grok 2 Claude 3.7 Sonnet
Slide 6: Agentic AI Further Accelerates Demand for Compute
Agentic AI Further Accelerates Demand for Compute MoreInference AGENTIC LOOP (Orchestration) MoreAgents REQUEST RESULT MoreTools INFERENCE/ SUB TOOL REASONING AGENTS USE MoreCompute
Slide 7: AI Demand Driving Acceleration in TAM
AI Demand Driving Acceleration in TAM DATA CENTER AI ACCELERATOR TAM ($B) ~$1.4T > 45% CAGR ~$200B Training Inference
Slide 8: AI Demand Driving Acceleration in TAM
AI Demand Driving Acceleration in TAM DATA CENTER CPU TAM ($B) ~$220B > 50% CAGR ~$26B General Purpose AI Host Agentic
Slide 9: AI Will Run Wherever the Work Happens
AI Will Run Wherever the Work Happens CLOUD DATA CENTER CLIENT EDGE
Slide 10: AMD Total Compute TAM
AMD Total Compute TAM ~ 40% CAGR $2T ~ 2030 $365B 2025 Data Center Client & Embedded
Slide 11: Strategy
Strategy Compute Leadership Open Platforms Powering AI Everywhere
Slide 12: Advancing AI 2026
Advancing AI 2026 AI Compute AI Powering for the Agentic Era Software AI Everywhere
Slide 13: Advancing AI 2026: AI Compute for the Agentic Era
Advancing AI 2026 AI Compute for the Agentic Era
Slide 14: EPYC Momentum Continues to Accelerate
EPYC Momentum Continues to Accelerate The Industry Standard For Server CPUs 46% 1st Gen 3rd Gen 5th Gen “Naples” “Milan” “Turin” AI Cloud Digital Enterprise OEM 2017 2026 SERVER CPU REVENUE SHARE
Slide 15: AI Leaders Trust AMD Instinct At Scale
AI Leaders Trust AMD Instinct At Scale
Slide 16: Frontier AI Requires Integrated Rack Architectures
Frontier AI Requires Integrated Rack Architectures ESUN Leadership Compute Open Rack Open Fabric Rack Scale Efficiency Turnkey Capability Architecture to Scale AI & Serviceability Solutions
Slide 17: Rackscale Performance For At-Scale AI
LAUNCHING TODAY AT ADVANCING AI 2026 RackscalePerformance For At-Scale AI
Slide 18: AMD Helios Delivers Leadership Across All Dimensions
AMD Helios Delivers Leadership Across All Dimensions COMPUTE MEMORY NETWORKING CPU Core Performance HBM4 Memory Capacity Scale Out Bandwidth GPU Compute HBM4 Memory Bandwidth Scale Up Bandwidth MI455X GPU “Venice” CPU Salina DPU Vulcano AI NIC 5th Generation Instinct 6th Generation EPYC AMD Pensando
Slide 19: AMD Helios Compute Architecture and Instinct MI455X
AMD Helios Compute Architecture Delivers The Highest Performing GPU With Instinct MI455X 40 PF|20 PF 432 GB 23.3 TB/s FP4 & FP8 Flops HBM4 Capacity Memory Bandwidth CDNA 5 2nm|3nm 320B GPU Architecture Advanced Process Transistors
Slide 20: AMD Helios Networking Powers Rackscale Performance
AMD Helios Networking Powers Rackscale Performance The Only Fully Programmable Networking Solutions Built for AI Front-End Scale-Out Scale-Up Salina DPU 800G AI NIC UALoE Secures Network Traffic AMD Pensando Vulcano End-to-end fabric reliability Accelerates SDN and Storage Broadcom Thor Ultra Indust
Slide 21: AMD Helios Delivers The Most Powerful Rackscale AI Infrastructure
AMD Helios Delivers The Most Powerful Rackscale AI Infrastructure + 15% + 50% + 50% 2.9 31 43 ExaFlops TB TB/s AI Compute FP4 HBM Capacity Scale Out Bandwidth + 15% + 6% Same 1.4 1.7 260 ExaFlops PB/s TB/s AI Compute FP8 HBM Bandwidth Scale Up Bandwidth Vera Rubin AMD Helios
Slide 22: The Most Powerful Rackscale AI Infrastructure
The Most Powerful Rackscale AI Infrastructure COMPUTE PERFORMANCE MEMORY NETWORKING 4,600 CPU Cores 2.9ExaFlops 31 TB 260 TB/s 18,000 FP4 Compute HBM Capacity Scale Up Bandwidth GPU Compute Units
Slide 23: AMD Helios In Production Today
AMD HELIOS | AMDZ1| AMDZI | AMDZI ME ereYC INSTINCT | P=ENSANDO F IN PRODUCTION TODAY
Slide 24: AMD and Anthropic
AMD 1 « ANTHROP\C
Slide 25: Instinct MI455X Delivers Generational Increase In Inference Throughput
Instinct MI455X Delivers Generational Increase In Inference Throughput Up to 4X 34x Higher Token Throughput vs MI355X 15X 34X Instinct MI355X Instinct MI455X
Slide 26: Instinct MI455X Delivers Generational Improvement In Inference Cost
Instinct MI455X Delivers Generational Improvement In Inference Cost Up to 18x Lower MI355X MI455X Token Cost vs MI355X Up to 18x
Slide 27: AMD Helios Delivers Industry Leading Rack Scale Performance
+15% +12% +10% THROUGHPUT (TOKENS / S / GPU) AMD Helios Delivers Industry Leading Rack Scale Performance Vera Rubin AMD HELIOS
Slide 28: Up To 30% More Tokens per Dollar
Up To 30%More Tokens /$ Using AMD Helios vs Vera Rubin NVL72
Slide 29: AMD Helios Adopted by Leading AI Companies
AMD Helios Adopted by Leading AI Companies AT SCALE AI PARTNERS INFRASTRUCTURE PARTNERS
Slide 30: AMD and OpenAI
AMDgt ~< G OpenAl
Slide 31: EPYC: 5 Generations of The World's Most Powerful Server CPUs
EPYC: 5 Generations of The World’s Most Powerful Server CPUs 5th Generation “Turin” AI OPTIMIZATION 4th Generation “Genoa” PURPOSE - BUILT DESIGN POINTS EPYC Delivers Core Performance 3rd Generation Socket Performance “Milan” CORE PERFORMANCE LEADERSHIP Thread Count Performance/Watt 2nd Generation “
Slide 32: EPYC Turin: The Best Server CPU in Production Today
EPYC “Turin” The Best Server CPU in Production Today PERFORMANCE DENSITY I/O BANDWIDTH LEADERSHIP LEADERSHIP LEADERSHIP Up to PCIe® GEN 5 @ #1 Performance 384 Threads 32Gbps 192 Cores
Slide 33: Three Tiers of Agentic AI Infrastructure Driving CPU Demand
Three Tiers of Agentic AI Infrastructure Driving CPU Demand GPU SERVERS AGENT CPU SERVERS GENERAL PURPOSE “HOST NODE” “SANDBOXES” CPU SERVERS Web Gateway Caching Apps DB & Storage High frequency cores Maximum # of performant Diverse balance of performance, core with leadership I/O cores per watt cou
Slide 34: Introducing 6th Gen AMD EPYC Venice
INTRODUCING 6th Gen AMD EPYC “Venice” World’s Best Server CPUs for the Agentic Era
Slide 35: 6th Gen AMD EPYC Venice: World's Best Server CPUs for the Agentic Era
6th Gen AMD EPYC “Venice” World’s Best Server CPUs for the Agentic Era PERFORMANCE DENSITY MEMORY BANDWIDTH TECHNOLOGY LEADERSHIP LEADERSHIP LEADERSHIP LEADERSHIP #1 512 Threads 1.6TB/s “Zen 6” CPU Architecture 1.8x 1.3x 2.6x 2nm|6nm Est. SPECint ® 2026 vs Turin vs Turin vs Turin Advanced Process
Slide 36: EPYC Venice: The Best Server CPU Portfolio
EPYC “Venice” The Best Server CPU Portfolio GPU SERVERS AGENT CPU SERVERS GENERAL PURPOSE “HOST NODE” “SANDBOXES” CPU SERVERS “Venice” HF RDIMM / MRDIMM “Venice” 256c “Venice” SP7 “Venice” SP8 “Venice -X” “Verano” LPDDR High Frequency High Core Count High Performance Balanced Technical CPU to GPU Ba
Slide 37: AMD EPYC Venice Delivers More Tokens, Agents & Performance vs Competition
AMD EPYC “Venice” Delivers More Tokens, Agents & Performance vs Competition GPU SERVERS AGENTIC CPU SERVERS GENERAL PURPOSE “HOST NODE” “SANDBOXES” CPU SERVERS 1.8x Up to Tokens/second for Frontier Models Up to 2.1x Agents per Watt Up to 2.3x Performance per Watt 1.7 1.6 1.4 1.0 1.0 0.9 1.0 0.8 0.6
Slide 38: AMD EPYC Venice vs ARM-based CPUs
AMD EPYC “Venice” Delivers More Agents & Performance vs ARM -based CPUs AGENTIC CPU SERVERS GENERAL PURPOSE “SANDBOXES” CPU SERVERS Up to 2.8x Agents per Watt Up to 3.3x Performance per Watt 1.7 1.4 1.0 1.0 0.9 0.9 0.6 0.4 NVIDIA Vera ARM AGI 5th Gen EPYC 6th Gen EPYC
Slide 39: AMD EPYC Venice vs Nvidia Vera
AMD EPYC “Venice” vs Nvidia Vera Generational Lead Single Core Performance & Throughput SINGLE CORE PERFORMANCE THROUGHPUT PERFORMANCE 1.2x Up to vs Vera Up to 2.2x vs Vera 2.2 1.2 1.0 1.0 NVIDIA Vera AMD “Venice” NVIDIA Vera AMD “Venice” 88c/Socket, 2 Socket 96c/Socket, 2 Socket 88c/Socket, 2 Socke
Slide 40: 6th Gen AMD EPYC Venice: World's Best Agentic CPU Racks
6th 6thGen GenAMD AMD EPYC EPYC “Venice” “Venice” World’s World’s Best Best Agentic Agentic CPU CPU Racks Racks 24,576 36,864 43,008 49,152 49,152 22,528 Cores Cores Cores Cores Cores Cores “Venice” 256c OEM Vera 88c MGX
Slide 41: AMD EPYC Venice In Production Today
AMD EPYC “Venice” In Production Today HYPERSCALERS OEM PROVIDERS CLOUD INSTANCES & OEM PLATFORMS AVAILABLE STARTING IN Q4
Slide 42: AMD and Meta
AMD¢\ « © Meta
Slide 43: Inference Market Is Segmenting
Inference Market Is Segmenting High Batch, Cost Optimized Balanced Throughput & Latency Ultra-Low Latency Offline Tasks Chatbots Code Completion Elastic Workloads Copilots Real-Time Copilots Deep Research AI Assistants Agentic Workflows
Slide 44: Disaggregated Inference Is Emerging
Disaggregated Inference Is Emerging As A Powerful Way To Deliver Ultra Low Latency REQUEST PREFILL DECODE RESPONSE COMPUTE INTENSIVE MEMORY BANDWIDTH INTENSIVE INFERENCE WORKFLOW
Slide 45: AMD and Cerebras
od AMD«¢l < Eerebras
Slide 46: The Most Powerful Solution For Ultra Low Latency Inference
The Most Powerful Solution For Ultra Low Latency Inference High-Throughput Scalable Engine Wafer -scale, Ultra Low Latency Engine 432 GBHBM4 Capacity 44 GBon-chip SRAM 23.3 TB/sMemory Bandwidth 21,000 TB/sMemory Bandwidth 72GPU Domain 900,000AI Cores
Slide 47: Ultra-Fast Inference, Higher TPS/W on Leading AI Models
Ultra-Fast Inference, Higher TPS/W on Leading AI Models Helios Cerebras ULTRA LOW LATENCY INTERACTIVITY (TOKENS / SECOND / USER) Kimi 2.6 1T Model
Slide 48: Ultra-Fast Inference, Higher TPS/W on Leading AI Models
Ultra-Fast Inference, Higher TPS/W on Leading AI Models Up to 5x TPS/kW Helios + Cerebras Cerebras ULTRA LOW LATENCY INTERACTIVITY (TOKENS / SECOND / USER) Kimi 2.6 1T Model
Slide 49: Unlock Higher Throughput Ultra Low Latency Inference
Unlock Higher Throughput Ultra Low Latency Inference Fastest Tokens In The World For Leading AI Models AMD HELIOS AI RACK CEREBRAS AVAILABLE IN 2H 2026 STARTING IN CEREBRAS INFERENCE CLOUD
Slide 50: AMD and Cerebras
od AMD«¢l < Eerebras
Slide 51: Advancing AI 2026: AI Software
Advancing AI 2026 AI Software
Slide 52: Accelerating Open Innovation
Accelerating Open Innovation
Slide 53: Relentless Focus on Developers
Relentless Focus on Developers Feature Performance Velocity Optimization Deep Open Ecosystem Richer Out -of-the-box Partnerships Experience 3M+ Models on Hugging Face Running out -of-the-box TOP 10 AI Open-Source Projects Natively Supported on AMD >10x Increase in OSS Contributions
Slide 54: ROCm Strategy and Evolution
ROCm Strategy and Evolution Open Source Abstraction Driving Rapid Innovation Enabling Higher Productivity
Slide 55: Open Source Strategy Accelerating Progress
Open Source Strategy Accelerating Progress OPEN SOURCE COMMUNITY WORLD’S MOST IMPORTANT ENABLING AMD WORKLOADS RUN ON AMD VERL
Slide 56: Abstraction Makes Deployment Easy on AMD
Abstraction Makes Deployment Easy on AMD Program at Every Level. Leveraging Rich Ecosystem FRAMEWORKS Rapid innovation at the highest level PYTHONIC DSLS Express intent in Python LIBRARIES Performance with productivity LOW - LEVEL Maximum control and performance PROGRAMMING
Slide 57: The Next Evolution: AI Assisted GPU Programming
The Next Evolution: “AI Assisted” GPU Programming AI Assist Automating GPU Programming Open Source Abstraction Driving Rapid Innovation Enabling Higher Productivity
Slide 58: Introducing AMD ROCm.AI
Claude Codex Cursor Gemini Introducing AMD ROCm.AI AI -Driven Platform for Developers
Slide 59: AI-Driven Development Platform
AI -Driven Development Platform Accelerating development velocity and performance optimization AI SKILLS Skills Claude Codex Cursor Gemini Popular agents become ROCm superusers AI ASSISTED OPTIMIZATION Hyperloom End-to-end AI Workload Optimization AI that optimizes workloads ROCm CORE Framework Inte
Slide 60: Accelerating Inference with ROCm.AI
Accelerating Inference with ROCm.AI Parallelization & Scheduling RELEASE -TO - RELEASE IMPROVEMENT Expert parallelism, DP-attention, Compute communications overlap INFERENCE Memory Management 3.3x average Paged attention, KV cache quantization, performance Unified attention (prefill + decode) improv
Slide 61: Accelerating Training with ROCm.AI
Accelerating Training with ROCm.AI Parallelization & Scheduling RELEASE -TO - RELEASE IMPROVEMENT Pipeline parallelism, Sharded data parallel, Gradient AllReduce overlap TRAINING Memory Management 2.4x average Activation checkpointing, FP8 mixed precision, performance Fused gradient accumulation imp
Slide 62: ROCm.AI Delivers Leadership Performance on MI455X
ROCm.AI Delivers Leadership Performance on MI455X MEMORY COMPUTE SCALE UP NETWORKING SCALE OUT NETWORKING 20 TB/s Measured MLA (Decode, FP8) 20 PF Measured FP4 Compute 3.2 TB/s Measured Bandwidth 190 GB/s Measured Bandwidth 3.8x 3.3x 3.5x 2x More Performance More AI Compute More Scale Up Bandwidth M
Slide 63: ROCm.AI Ready for MI455X Day 0 Support
ROCm.AI Ready for MI455X Day 0 Support PyTorch Core tests We tested the vLLM is working well We are thrilled run successfully on MI455X and love on Helios and we are to have SGL ang MI455X, and full what we’re seeing. excited to see what enabled on day support is on track. Models run out of the our
Slide 64: Slide 64
Slide 65: AMD and OpenAI
od AMDZ1 « G OpenAl
Slide 66: Open Software Stack Built for Every Workload
AI SCIENCE/HPC Training, Inference, Intelligence Simulations, Discoveries, Impact Open Software Stack Built for Every Workload
Slide 67: Leadership Platforms Powering AI & HPC
Leadership Platforms Powering AI & HPC Require Breadth of Accelerated Compute AT SCALE AI TRAINING & INFERENCE SOVEREIGN AI & HPC AMD Instinct MI455X AMD Instinct MI430X AI COMPUTE HARDWARE BASED FP64 HBM4 MEMORY HBM4 MEMORY CAPACITY NETWORKING PERFORMANCE HBM4 MEMORY BANDWIDTH
Slide 68: AMD Instinct MI430X
SHIPPING IN 1H 2027 AMD Instinct MI430X Most Advanced HPC & Sovereign AI Accelerator HPC FP64 COMPUTE HBM4 CAPACITY HBM4 BANDWIDTH 288TF 432 GB 23.3 TB/s 8.7x vs Vera Rubin +50% vs Vera Rubin +6% vs Vera Rubin
Slide 69: First Exascale Sovereign AI Factories Are Deploying MI430X
US and Europe’s First Exascale Sovereign AI Factories Are Deploying MI430X Discovery Alice Recoque
Slide 70: The Future of GPU Development
The Future of GPU Development Open Higher Level AI -Assisted ROCm Ecosystem Abstractions Development Everywhere ROCm.AI AVAILABLE AUGUST 2026
Slide 71: Slide 71
Slide 72: Advancing AI 2026: Powering AI Everywhere
Advancing AI 2026 Powering AI Everywhere
Slide 73: Powering AI Everywhere
Powering AI Everywhere ENTERPRISE AI PERSONAL AI PHYSICAL AI Enterprise Individual Productivity Robotics, Autonomous Vehicles, Data Center & Development Industrial AI, Healthcare
Slide 74: EPYC is the Established Leader for Enterprises
EPYC is the Established Leader for Enterprises AWS = EE Microsoft ©) GoogleClud ORACLE FEE 60% Tut Th Ueto eer Gi ube BP NETFLIX €} zoom 6. -” ay € WO) BL enim © 4¢ AMD Is the Company to Beat for Enterprise Al Server CPUs 99
Slide 75: Enterprise Workloads are Diverse
Enterprise Workloads are Diverse MEDIA REAL TIME WEB BUSINESS DATA PROCESSING SERVICES APPLICATIONS APPLICATIONS ENCRYPTION Thread Count Memory Bandwidth I/O Performance Cache Efficiency Core Performance
Slide 76: Venice Extends Enterprise Leadership
“Venice” Extends Enterprise Leadership MEDIA REAL TIME WEB BUSINESS DATA PROCESSING SERVICES APPLICATIONS APPLICATIONS ENCRYPTION Up to 2.7x Up to 2.9x Up to 3.7x Up to 2.6x Up to 2.5x FFmpeg Redis NGINX MySQL OpenSSL 6th Gen Intel 6th Gen EPYC
Slide 77: Agentic AI is the Force Multiplier for Enterprise
Agentic AI is the Force Multiplier for Enterprise But it Comes with New Challenges for CIOs 95% CIO Considerations Cost Data Governance of enterprises surveyed expect to use agentic AI by 2027 Integration
Slide 78: Agentic AI Deployment Will Be Distributed
Agentic AI Deployment Will Be Distributed to Address These Challenges FRONTIER CLOUD ON PREM CLIENT
Slide 79: Agentic AI Deployment Will Be Distributed: Constraints
Agentic AI Deployment Will Be Distributed to Address These Challenges CONSTRAINTS ON PREM Power & Cooling Cost Ease of Deployment
Slide 80: AMD Instinct MI350P: The Easiest Way to Add AI to Your Data Center
AMD Instinct MI350P The Easiest Way to Add AI to Your Data Center PERF/$ LARGE MODEL STANDARD SERVER LEADERSHIP CAPABLE COMPATIBLE Up to 4.2x Up to 260 Billion Air Cooled Tokens/second/$ vs RTX Pro 6000 Model parameters @ FP4 450-600W TBP
Slide 81: AMD Instinct MI350P Delivering Higher Productivity
AMD Instinct MI350P Delivering Higher Productivity Tokens/Second vs RTX Pro 6000 WORKFLOW AUTOMATION Llama 3.3 70B Instruct 5.1x CUSTOMER SUPPORT Mistral Small 3.2 24B 2.5x CONTRACT ANALYSIS GPT-OSS 120B 2.1x
Slide 82: AMD IT Optimizes Costs by Routing to the Right Model
Case Study AMD IT Optimizes Costs by Routing to the Right Model 24/7 Autonomous Frontier 43% Cost Savings Threat Detection vs 100% Frontier tokens On Prem EPYC +MI350P Qwen 72B Personal Agents (OpenClaw) TOKEN ROUTER On Prem EPYC +MI350P 2.9x Faster Responses Gemma 31B vs Frontier
Slide 83: A Complete Ecosystem for Enterprise AI
A Complete Ecosystem for Enterprise AI AI ISVs GLM Models Open Software VERL Platform ISVs OEMs/ODMs
Slide 84: Slide 84
Slide 85: AMD and Sater
AMD2Z1 « Sater
Slide 86: Delivering AI at Scale
Delivering AI at Scale >45B #1 122 Tokens Per Day Leadership on 7 Key AI Benchmarks Production RAG+FT Pipelines 40% #1 5x Announcing OTel 2.0 In-year Free Cash Flow Specialized Models on AMD Trained on AMD Impacting ROI
Slide 87: AMD Delivers End-to-End Solutions for Enterprise AI
AMD Delivers End-to-End Solutions for Enterprise AI CLOUD ON PREM SERVER CLIENT Full Ryzen AI Racks Laptops Instinct PCIe Agent Cards Computers Radeon Pro Cloud Instances OEM Servers PCIe Cards
Slide 88: AI Agents Are the Next Great Productivity Multiplier
AI Agents Are the Next Great Productivity Multiplier PERSONAL AGENT AGENT TEAM MULTI - CAPABILITY AGENTS 2x Force Multiplier 10x Force Multiplier 100x Force Multiplier
Slide 89: Why Local AI is Taking Off
Why Local AI is Taking Off SECURITY & PRIVACY “FREE TOKEN” INFERENCE RELIABLE ANYWHERE Full control over data and Optimize recurring Continue working no third-party handling inference costs without connectivity
Slide 90: Small Models are Becoming Incredibly Efficient
Small Models are Becoming Incredibly Efficient GPT-OSS Qwen 3.5 Number of Parameters 120B Number of Parameters 9B GPQA Score* Graduate-Level Google-Proof Q&A 80.1 GPQA Score* Graduate-Level Google-Proof Q&A 81.7 August 2025 March 2026
Slide 91: Smaller Models are Closing The Frontier Gap
Smaller Models are Closing The Frontier Gap GPQA SCORE* 87.8 Smaller open models can now outperform yesterday's frontier models while using a fraction of the cost. Sonnet 4.5 Opus 4.5 Qwen 3.6 27B
Slide 92: Efficient Models Unlock Personal AI
Efficient Models Unlock Personal AI Small Efficient Models Can Fit Easily Across AMD Agentic PCs 9B Model Size 24B Model Size Limit Ryzen AI 400 27B Model Size 200B Model Size Limit Ryzen AI MAX
Slide 93: 128GB Unified Memory, 200B Models Natively
128GBUnified Memory 200BModels Natively + Speed to Productivity + Simplicity of Setup + Efficiency of Cost + Continuous Developer Support
Slide 94: One Software Stack. Every AMD Device.
One Software Stack. Every AMD Device . Write Once. Run Anywhere. Agent Frameworks Lemonade · OpenClaw · LM Studio · Ollama · CrewAI AI Frameworks PyTorch · JAX · vLLM · ONNX ROCm Libraries rocBLAS · MIOpen · hipBLAS · Triton ROCm Runtime & Compiler HIP · HIPCC · ROCm-CPU AMD Silicon Ryzen AI · Radeo
Slide 95: Deepening our Partnership with the World's Premier Open-Source AI Community
Deepening our Partnership with the World’s Premier Open-Source AI Community Native Halo support in Hugging Face Co-engineered Halo Local AI experiences Every Halo Box Ships with with optimized models and workflows + Toolkits to Accelerate Development 1 Year Hugging Face PRO
Slide 96: Pushing inference further with Gorgon Halo
Pushing inference further with “GORGON HALO ” Largest unified memory from128GB to192GB Extending AI models from200Bto300B Run cloud-scale AI Locally on Your Desktop
Slide 97: Slide 97
Slide 98: AMD and Cisco
AMD¢l ~ cisco
Slide 99: Agents Operate at Machine Speed
Agents Operate at Machine Speed Bursty —demand spikes, then idles Steady state —always on, at machine speed INFRASTRUCTURE DEMAND INFRASTRUCTURE DEMAND CHATBOTS AGENTIC & PHYSICAL AI
Slide 100: AMD and Cisco
AMD¢l ~ cisco
Slide 101: Network Infrastructure, Tokenomics, Agent Behavior, Security
NETWORK INFRASTRUCTURE TOKENOMICS AGENT BEHAVIOR SECURITY
Slide 102: A Full Stack Inside Each Device
Cisco security, network & control A Full Stack Inside Each Device AMD hardware & software Unified policy & control (Cisco Cloud Control) Observability (Splunk) Model & agent security (AI Defense) Security policy enforcement (DefenseClaw) Model integration Model routing & token limits Local inference
Slide 103: Security & Governance for Agentic AI
Security & Governance for Agentic AI |Limit agent access to approved resources |Apply consistent security policies |Identify and isolate activity in real time
Slide 104: Intelligence for Enterprise Inference
Intelligence for Enterprise Inference |Gain visibility into AI costs and savings |Improve model and token efficiency |Identify adoption and usage trends
Slide 105: AMD and Cisco
AMD ~ cisco
Slide 106: Powering AI Everywhere
Powering AI Everywhere ENTERPRISE AI PERSONAL AI PHYSICAL AI Enterprise Robotics, Autonomous Vehicles, Individual Productivity & Development Data Center Industrial AI, Healthcare
Slide 107: Physical AI: The Ultimate Application of Agentic AI
Physical AI The Ultimate Application of Agentic AI
Slide 108: Physical AI that Amplifies Human Potential
Physical AI that Amplifies Human Potential From today's deployment to tomorrow’s possibilities ROBOTIC SMART INTELLIGENT FACTORY WAREHOUSING MULTI - PURPOSE SURGERY MANUFACTURING AGRICULTURE AUTOMATION
Slide 109: AMD is a Trusted Leader in Robotics
AMD is a Trusted Leader in Robotics Enabling sensing , safety, and control systems for over 20 years
Slide 110: The Next Era of Robotics is Autonomous
The Next Era of Robotics is Autonomous HIGH-LEVEL GOALS CPU MULTI - AGENT PLANNING & ORCHESTRATION FPGA NPU GPU CPU GPU FPGA SENSORS VISION PERCEPTION CONTROL & SELF CORRECTION & STATE ESTIMATION LEARNING ACTION REASONING & ACTION PLANNING Plans, reasons, learns and adapts to achieve desired outcome
Slide 111: AMD KRIA AI System on Module
INTRODUCING AMD KRIA AI System on Module Powered by Ryzen AI Embedded X100 Series INDUSTRY - LEADING SYSTEM PERFORMANCE CPU 125 µs Deterministic Control GPU 92ms AI Reasoning NPU <0.4ms Vision Classification Unified Memory OPEN STANDARDS DEPLOY FASTER COM-HPC Form Factor Pre-engineered Platform
Slide 112: The New Standard for the Robotics Brain
KRIA AI System On Module The New Standard for the Robotics Brain REACTION TIME AGENTIC AI CAPACITY ROBOT CAPABILITY 3.4x Better real-time results 2.3x More Concurrent agents 1.6x More CPU Capacity
Slide 113: AMD KRIA AI Robotics Developer Platform
INTRODUCING AMD KRIA AI Robotics Developer Platform Industry’s First Turnkey Open, Fully Integrated Platform From concept to prototype in days Enabled with Robotics SW Suite built on AMD ROCm and ROS 2 Faster deployment with open baseboard schematic & FPGA design
Slide 114: Only AMD Delivers the Complete Autonomous Robotics Solution
BRAIN High-level decision making, communication and human-machine interface Optimized for inference throughput Only AMD SPINE Time-critical communication coordination between brain and joints Delivers the Optimized for reliability Complete JOINTS Real-time multi-axis actuation Autonomous Optimized f
Slide 115: Advancing the Open Robotics Ecosystem
Advancing the Open Robotics Ecosystem AI MODELS & SOFTWARE ROBOTIC TECHNOLOGIES HARDWARE PARTNERS INDUSTRY GROUPS
Slide 116: Announced today at Advancing AI 2026
Announced today at Advancing AI 2026 “Helios” “Venice” MI430X MI350P ROCm.ai Kria AI Robotics “Gorgon Halo” SOM & DEVELOPER PLATFORM DEVELOPER PLATFORM
Slide 117: AMD EPYC Leadership Server CPU Roadmap
AMD EPYC Leadership Server CPU Roadmap “Ravenna” “Florence” 8th Generation EPYC “Venice” 7th Generation EPYC “Turin” 6th Generation EPYC 5th Generation EPYC “Zen 8 ” “Zen 7 ” “Zen 6 ” Family “Zen 5 ” “Zen 7c ” “Zen 6c ” “Zen 5c ” Next Gen Process Node First to 2nm & 512 Threads In Development Wider/
Slide 118: EPYC Florence Extends EPYC Leadership For the Agentic Era
COMING IN 2028 EPYC “Florence” Extends EPYC Leadership For the Agentic Era “Florence” “Florence” “Ferrara” “Fidenza” SP7 SP8 PERFORMANCE ENTERPRISE AI HOST NODE AGENTIC “SANDBOX” Leadership Core Count and Optimized for Performance Optimized for high- Optimized for Performance per System $ performanc
Slide 119: AMD Instinct Leadership GPU Roadmap
AMD Instinct Leadership GPU Roadmap MI600 Series MI500 Series MI400 Series MI350 Series Next Gen HBM4e Memory Increased HBM4 Capacity & Bandwidth In Development Increased HBM3e Capacity & Bandwidth Leadership GPU Scale Up Domain Expanded AI Formats with Higher Throughput Block Scaled AI Formats Copp
Slide 120: Instinct MI500: The Next Leap In AI Performance
Instinct MI500 The Next Leap In AI Performance AMD INSTINCT MI500 Series INFERENCE THROUGHPUT AMD INSTINCT >2,000X Increase in 4 Years MI455X AMD INSTINCT MI355X
Slide 121: Leadership AI Rack Roadmap on an Annual Cadence
Leadership AI Rack Roadmap on an Annual Cadence AMD HELIOS 600 AMD HELIOS 500 AMD HELIOS AMD EPYC AMD EPYC “Ferrara” AMD EPYC “Verano” “Venice” AMD Instinct AMD Instinct MI600 Series AMD Instinct MI500 Series MI455X AMD Pensando AMD Pensando “Palma” & “Levanzo” AMD Pensando “Como” & “Monza” Vulcano
Slide 122: Announced today at Advancing AI 2026
Announced today at Advancing AI 2026 “Helios” “Venice” MI430X MI350P ROCm.ai Kria AI Robotics “Gorgon Halo” SOM & DEVELOPER PLATFORM DEVELOPER PLATFORM In Production In Production Available Available Available Available Available Now Now 1H 2027 Now August 2H Q4
Slide 123: Slide 123
Slide 124: Endnotes
Endnotes 9xx5-266: SPECrate®2017_int_base comparison based on published scores from www.spec.org as of 10/24/2025. 2P AMD EPYC 9965, 192C, 3230 SPECrate®2017_int_base, https://www.spec.org/cpu2017/results/res2025q2/cpu2017-20250324-47086.html SPEC®, SPEC CPU®, and SPECrate® are registered trademarks
Slide 125: Endnotes
Endnotes 9xx6-014: 9xx6-014: Comparison based on published Top-of-stack core counts and CPU W, Default CPU Power for 6th Gen EPYC, and TDPs for 5th Gen EPYC, Intel® Xeon®, and Nvidia Vera for estimated Highest Agents / CPU W across 6th Gen AMD EPYC 256C at 400W, AMD EPYC 9965 (500W TDP), Nvidia Vera
Slide 126: Endnotes
Endnotes 9xx6-036: FFmpeg video encoding and transcoding throughput comparison based on AMD internal testing as of 7/20/2026. 1P AMD EPYC 9996 (256C) delivers up to 2.7x the FFmpeg video encoding and transcoding throughput of 1P Intel Xeon 6980P (128C) Software: FFmpeg encode and transcode workload
Slide 127: Endnotes
Endnotes 9xx6-039: MySQL 8.4.10 online transaction processing throughput comparison (HammerDB TPROC-C, a workload derived from the TPC-C Benchmark Standard. Results are based on TPC-C derived workloads and are not comparable to published TPC benchmark results.) based on AMD internal testing as of 7/
Slide 128: Endnotes
Endnotes 9xx6-050: SPECrate®2026_int_base throughput comparison based on AMD internal estimates for the 2P 6th Gen AMD EPYC 9996 and 2P NVIDIA Vera as of 7/22/2026. Preliminary performance estimates based on AMD engineering projections and subject to change. 2P 6th Gen AMD EPYC 9996: GCC 15.2, estim
Slide 129: Endnotes
Endnotes MI350P-001: Based on AMD internal testing (July 2026) on a (1x) AMD Instinct MI350P GPU vs (1x) Nvidia RTX Pro 6000 Blackwell Server Edition, on the OpenAI GPT-OSS-120B (MXFP4) benchmark measuring online serving output-throughput (tok/s) at ISL/OSL 1024/1024 across concurrency levels 1, 4,
Slide 130: Endnotes
Endnotes MI400-001B: Based on engineering projections by AMD performance labs in July 2026, to estimate peak theoretical precision performance of an AMD Instinct MI430X GPU using FP64 (Vector) datatype vs. NVIDIA Rubin GPU with FP64 (Vector) datatype. Results subject to change when products are rele
Slide 131: Endnotes
Endnotes MI400-030: Based on testing by AMD Performance Labs in July 2026, on systems powered by AMD ROCm.ai and configured with an AMD Instinct MI455X GPU and AMD Pensando Vulcano 800 AI NIC and AMD Instinct MI355X GPU with AMD Pensando Pollara 400 AI NIC to determine bidirectional scale-out networ
Slide 132: General Disclaimer and Attribution
General Disclaimer and Attribution Timelines, roadmaps, and/or product release dates shown in these slides are plans only and subject to change. The information contained herein is for informational purposes only and is subject to change without notice. While every precaution has been taken in the p