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 and product roadmaps including the AMD InstinctTM MI350 Series, AMD RO
Slide 2: Advancing AI 2025
Slide 3: High Performance & Adaptive Computing
High Performance & Adaptive Computing Solving the World’s Most Important Challenges
Slide 4: AI Everywhere, for Everyone
AI Everywhere, for Everyone Financial Healthcare and Robotics Manufacturing Communications Automotive Productivity Services Life Sciences
Slide 5: AI Innovation is Accelerating
AI Innovation is Accelerating Inference Scaling Reasoning & Training is Evolving Accelerates Explosion of Models Agents Surge
Slide 6: Reasoning & Agents Fuel Compute Surge
Reasoning & Agents Fuel Compute Surge GPU Enterprise CPU Leadership GPU Leadership CPU Openness Lowers TCO Powers apps Accelerates Innovation
Slide 7: Data Center AI Accelerator TAM
Data Center AI Accelerator TAM Inference Growing at >80% CAGR >$ 500B >60% CAGR $45B 2023 2024 2025 2026 2027 2028 Training Inference Source: AMD Internal Data
Slide 8: AI Beyond the Data Center
AI Beyond the Data Center Inference Scaling Across Cloud to Edge to Client Cloud Edge Client Driven by domain-specific compute engines & open software stack
Slide 9: AI Strategy
Leadership Open Full Stack AI Strategy Compute Engines Ecosystem Solutions
Slide 10: Best End-to-End AI Compute Portfolio in the Industry
Best End-to-End AI Compute Portfolio in the Industry AMD Ryzen AI AMD EPYC AMD Instinct AMD Pensando AMD Versal Processors Accelerators Networking AMD Radeon AI Adaptive SOCs Processors Premier programmable Most powerful client Leadership AI Leading server CPU World’s best GPU accelerator DPUs & AI
Slide 11: Open Development Drives Value & Innovation
Open Development Drives Value & Innovation Open Hardware Open Software Open Ecosystem Choice Flexibility Rapid Co-Innovation Portability Proven
Slide 12: Investing in Full-Stack Solutions
Investing in Full-Stack Solutions Acquisitions Span Entire AI Value Chain Over 25 AI Acquisitions & Investments in the Last Year Alone
Slide 13: EPYC Momentum Accelerates…
EPYC Momentum Accelerates… 15X Market >18x Server Share Growth CPU Market Share Growth Industry Leaders Run on EPYC 40% 1Q25 Cloud 36% 24% Digital 8% Enterprise 2% 2018 2020 2022 2024 OEM Source: Mercury
Slide 14: Delivering on Leadership GPU Commitment
Delivering on Leadership GPU Commitment Powering Top Supercomputers Trusted by Leading AI Practitioners AMD Instinct AMD Instinct AMD Instinct AMD Instinct MI250X MI300A MI300X MI325X 2021 2025
Slide 15: Growing Industry Adoption
Growing Industry Adoption 7 of 10 Largest AI Companies Use AMD Instinct
Slide 16: Rapidly Advancing Open Software Capabilities
Rapidly Advancing Open Software Capabilities Day-0 Support Accelerated Broadening Developer First for Leading Models Pace of Innovation Ecosystem Partnerships Approach to Enablement A
Slide 17: Today at Advancing AI
Today at GPUs Software Solutions Advancing AI
Slide 18: Today at Advancing AI
Today at GPUs Software Solutions Advancing AI
Slide 19: Delivering on Annual Roadmap Commitment
Delivering on Annual Roadmap Commitment AMD Instinct AMD Instinct MI400 SERIES AMD Instinct MI350 SERIES AMD Instinct MI325X MI300A/X 2026 2025 2024 2023 Roadmap subject to change
Slide 20: AMD Instinct MI350 Series
Launching Today at Advancing AI 2025 AMD Instinct MI350 Series Continued Generative AI Leadership
Slide 21: AMD Instinct MI350 Series Architecture
4th Gen Instinct architecture 3nm process node 185 billion transistors FP4 & FP6 new gen AI datatypes AMD Instinct MI350 Series HBM3E leadership capacity
Slide 22: MI350 Series Accelerates Your Gen AI Outcomes
MI350 Series Accelerates Your Gen AI Outcomes Faster AI Inference & Training Larger AI Model Support Rapid AI Infrastructure Deployment 20PF 288GB UBB8 FP4 & FP6 HBM3E Industry Standard GPU Node 4x Gen-on-Gen AI Supports up to 520B Available in Air Cooled Compute Increase Parameter AI Model & Direct
Slide 23: MI350 Series Specifications
Instinct MI355X AMD Instinct MEMORY 288 GB HBM3E MI350 Series MEMORY BANDWIDTH 8 TB/s FP64 79 TF FP16 5 PF FP8 10 PF FP6 20 PF FP4 20 PF TBP 1400W
Slide 24: Instinct MI350 Series Advantage
Instinct MI350 Series Advantage vs. GB200 vs. B200 MEMORY 1.6x 1.6x MEMORY BANDWIDTH 1.0x 1.0x FP64 2.0x 2.1x FP16 1.0x 1.1x FP8 1.0x 1.1x FP6 2.0x 2.2x FP4 1.0x 1.1x See endnote: MI350-008, 009
Slide 25: MI350 Series Platform
AMD Instinct MI350 Series Platform Instinct MI355X • 8x MEMORY 2.3 TB HBM3E MEMORY BANDWIDTH 64 TB/s FP64 0.63 PF FP8 81 PF FP6 161 PF FP4 161 PF
Slide 26: Air Cooled Liquid Cooled
Air Cooled Liquid Cooled AMD Instinct MI350 Series AMD Instinct MI350 Series
Slide 27: MI355X Delivers Generational Performance Leap
MI355X Delivers Generational Performance Leap For Ultra Low Latency Inference AMD Instinct MI355X Up to 35x AMD Instinct MI300X Llama 3.1 405B Inference performance, throughput • MI355X (FP4) and MI300X (FP8) See endnote: MI350--044
Slide 28: Over 3x Generational Inference Improvement
Over 3x Generational Inference Improvement For Broad AI Use Cases Up to 4.2x 3.8x 2.9x 2.6x AI Agent & Chatbot Content Generation Summarization Conversational AI Llama 3.1 405B • MI355X vs. MI300X Inference performance, throughput • MI355X (FP4) and MI300X (FP8) See endnote: MI350-042
Slide 29: Meeting Next-Gen Inference Demands
Meeting Next-Gen Inference Demands For the Most Popular AI Models 3.3x 3.2x Up to 3.0x DeepSeek R1 Llama 4 Maverick Llama 3.3 70B MI355X vs. MI300X Inference performance, throughput • MI355X (FP4) and MI300X (FP8) See endnote: MI350-043, 041, 048
Slide 30: MI355X Delivers the Highest Inference Throughput
MI355X Delivers the Highest Inference Throughput For Large Models Up to 1.2x 1.3x 1.0x DeepSeek R1 • FP4 Llama 3.1 405B • FP4 Llama 3.1 405B • FP4 Nvidia Nvidia AMD Instinct B200 GB200 MI355X Inference performance, throughput See endnote: MI350-038, 039, 040
Slide 31: Up to 40% More Tokens / $
Up to 40% More Tokens / $ Using AMD Instinct MI355X vs. B200 See endnote: MI350-049
Slide 32: Meta Runs AI Recommendation Models on AMD MI300X
In addition to Llama Inference, Meta’s AI Recommendation Inference & Training models run on AMD MI300X GPUs
Slide 33: Accelerating Model Training by 3x
Accelerating Model Training by 3x For Pre-Training & Fine-Tuning 3.5x Up to 2.6x 2.7x 2.9x Llama 3 8B • FP8 Llama 3 70B • FP8 Llama 3 70B • BF16 Llama 2 70B • FP8 Pre-Training Pre-Training Pre-Training Fine-Tuning MI355X vs. MI300X Training: Throughput • Fine Tuning: Time To Train See endnote: MI350
Slide 34: World Class Training & Fine-Tuning Performance
World Class Training & Fine-Tuning Performance Across Models & Data Types Up to 1.0x 1.0x 1.13x 1.1x Llama 3 8B • FP8 • BF16 Llama 3 70B • FP8 • BF16 Llama 2 70B • FP8 Llama 2 70B • FP8 Pre-Training Pre-Training Fine-Tuning Fine-Tuning Nvidia Nvidia AMD Instinct B200 GB200 MI355X Pre-Training: Throu
Slide 35: Proven Liquid Cooled Infrastructure
AMD Instinct MI350 Series Proven Liquid Cooled Infrastructure “Turin” 128/96 Up to 36 TB Up to 2.6 EF Up to 1.3 EF Up to MI350 Series GPUs HBM3E FP4 FP8 Pollara 400 Specs available on 2ou/3ou server
Slide 36: Proven Air Cooled Infrastructure
AMD Instinct MI350 Series Proven Air Cooled Infrastructure “Turin” 64 18 TB 1.3 EF 0.6 EF MI350 Series GPUs HBM3E FP4 FP8 Pollara 400 Specs available on 6u server
Slide 37: Driving AI Innovation, Trusted by Enterprises
Driving AI Innovation Trusted by Enterprises Powering Enterprise Software Driving AI Innovation Trusted by Enterprises Powering Enterprise Software
Slide 38: AMD Instinct MI350 Series Solution Partners
AMD Instinct MI350 Series Solution Partners Available starting Q3
Slide 39: Harnessing AI Across Nations
Harnessing AI Across Nations End-to-End Infrastructure Diverse Ecosystems Open Architectures
Slide 40: Advancing National Economies
Advancing National Economies USA USA Canada EU Finland Switzerland France France Germany Germany Italy Czech Republic Sweden Austria Singapore UAE Japan Chile Australia Saudi Arabia
Slide 41: AMD Silo AI: Europe's AI Solution Factory
AMD Silo AI: Europe’s AI Solution Factory Working with European AI Stakeholders Across Public & Private Sectors OpenEuroLLM Deploy AI MINERVA Multilingual LLMs Partnerships
Slide 42: AMD Instinct MI350 Series
AMD Instinct MI350 Series Generative AI Leadership
Slide 43: Today at Advancing AI
Today at GPUs Software Solutions Advancing AI
Slide 44: Enabling Open Innovation at Scale
Enabling Open Innovation at Scale
Slide 45: Relentless Progress, Focused on Developers
Relentless Progress, Focused on Developers Accelerated inference Expanded support Richer out-of-the-box Developer-first approach capabilities for training experience to enablement Accelerated release Day-0 support for Deepening ecosystem Industry benchmarks cadence leading models partnerships
Slide 46: Deepening Ecosystem Collaboration
Deepening Ecosystem Collaboration v1 Pytorch Triton v3.3 Hugging Face 1.8 million models Serving leadership Day 0 support Nightly CI/CD, Support for Expanding open- Performance focus Distributed daily performance CI finetuning support SOTA models source footprint inference
Slide 47: AI Software Innovation Continues at Rapid Pace
AI Software Innovation Continues at Rapid Pace Models & Applications Breakthrough Algorithms AI at Scale Hardware Innovation Sparse Mixture Lower Precision Distributed Reinforcement Reasoning Agentic AI Attention of Experts Compute Inference Learning
Slide 48: Introducing AMD ROCm 7
Introducing AMD ROCm 7 Accelerating AI Innovation & Developer Productivity Latest Algorithms Advanced Features MI350 Series Cluster Enterprise & Models for Scaling AI Support Management Capabilities
Slide 49: Growing Inference Capabilities
Growing Inference Capabilities New AMD ROCm Features Enhanced Frameworks vLLM v1 llm-d SG Lang Distributed Serving Optimization Inference Prefill Disaggregation Python-Based GEMM Kernels & Algorithms Autotuning MoE Attention Kernel Authoring Communication rocSHMEM GPU Direct Access RCCL Advanced Dat
Slide 50: Accelerating Inference Performance
Accelerating Inference Performance 3.4x 3.8x 3.2x 3.5x average performance improvement Llama 3.1 70B Qwen2-72B DeepSeek R1 ROCm 7 vs. ROCm 6 See endnote: MI300-080
Slide 51: DeepSeek R1 Performance
Open Source: Feature Velocity & Leadership Performance Up to 1.3x DeepSeek R1 FP8 Throughput FP8 Model Support TRT-LLM Nvidia AMD Instinct B200 MI355X See endnote: MI350-025
Slide 52: AI Assistants, Agentic Solutions, Multi-Modal Training
Azure OpenAI Service M365 Copilot App Azure AI Foundry AI Assistants Agentic Solutions Multi-Modal Training
Slide 53: Growing Training Capabilities
Growing Training Capabilities New AMD ROCm Features Text-to- Text-to- European Game AMD Open-Source Models Text Image Models Multimodal Agent JAX Enhanced Frameworks PyTorch Maxtext Torchtune Torch-titan Parallelization DP PP TP FSDP CP EP Kernels and Algorithms GEMM Attention Advanced Data Types BF
Slide 54: Accelerating Training Performance
Accelerating Training Performance 3.0x 3.0x 3.1x 3x average performance improvement Llama 2 70B Llama 3.1 8B Qwen 1.5 7B ROCm 7 vs. ROCm 6 See endnote: MI300-081
Slide 55: Distributed Inference at Scale with Open Ecosystem
Distributed Inference at Scale with Open Ecosystem AI Serving Throughput, Multiplied Orchestration Framework Key Functions PD KVcache Transfer Cross-node Communication Cross-PD Group Schedule Key Technologies Mooncake GPU Direct Access DeepEP Distributed Triton SHMEM
Slide 56: Lowering the Cost of Token Generation
Lowering the Cost of Token Generation
Slide 57: Extending AMD ROCm for Enterprise AI
Extending AMD ROCm for Enterprise AI Enterprise Ready End-to-End Solutions Secure Data Integration Ease of Deployment
Slide 58: AMD ROCm Enterprise AI
MLOps AI Workload & Quota Management AMD ROCm Enterprise AI Kubernetes & Slurm Integration Operations Platform | Cluster Management Cluster Provisioning & Telemetry Compiler Libraries Profiler Runtime AMD ROCm 7 GPUs CPUs DPUs Data Center Infrastructure
Slide 59: Empowering Developers
Empowering Developers Ease of Use Collateral for Rapid Adoption Tutorials Blogs Videos
Slide 60: Engaging Developers
Engaging Developers Developer-Focused Events to Strengthen the Community Hackathons Hands-on Workshops Developer Contests Meet-Ups Community CI
Slide 61: Announcing AMD Developer Cloud & Developer Credits
Announcing AMD Developer Cloud & Developer Credits Available for Developers & Open-Source Contributors
Slide 62: Expanding AMD ROCm on Client
Expanding AMD ROCm on Client AI-Assisted Coding Customization Automation Advanced Reasoning Model Fine-Tuning
Slide 63: In-Box Linux Support, Full Windows Support
In-Box Linux Support Full Windows Support Red Hat EPEL PyTorch Preview Q3 2025 Ubuntu ONNX-EP Preview July 2025 OpenSUSE HIP SDK Fedora Linux in Windows WSL
Slide 64: The Ultimate Client AI Solutions for Every Need
The Ultimate Client AI Solutions for Every Need AMD Threadripper AMD Ryzen AI 300 AMD Ryzen AI Max Radeon AI Up to 24B parameters Up to 70B parameters Up to 128B parameters
Slide 65: Developer Track: Meet the AI Experts
Developer Track: Meet the AI Experts Andrew NG Chris Lattner Ashish Vaswani Co-Founder, Coursera, CEO, Modular Inc. CEO, Essential AI Landing AI, DeepLearning AI Robert Shaw Mattias Reso Daniel Han Director of Engineering, Red Hat AI AI Partner Engineer, Meta CEO, Unsloth AI Kunlun Zhu Mark Saroufim
Slide 66: Accelerating Innovation
Accelerating Innovation Open by Design Trusted at Scale Empowering Developers
Slide 67: Today at Advancing AI
Today at GPUs Software Solutions Advancing AI
Slide 68: Transforming Business & Society
Transforming Business & Society Optimize Business Processes Accelerate Innovation Legal Contracts Material Science Inventory Management Drug Discovery Product Pricing Mechanical Design
Slide 69: Query Business Process — Outcome
Query Business Process — Outcome v SUPPORT AGENT SUPPORT AGENT SUPPORT AGENT SUPPORT AGENT Data Retrieval Human Feedback External Research Internal Tools a # A a &® a # a #” Gateway Gateway Gateway Gateway Servers Servers Servers Servers a B®” a B® a B®” a B® | Enterprise Applications, Databases and
Slide 70: Agentic Execution
Agentic Execution Initial Prompt Actions Results Prompt, Multiple Passes PRE-PROCESSING MODEL INFERENCE POST-PROCESSING “Find the Largest 0 ity Li tunity List Opportunities and Build —> — } see — ener > and Sales Plan. a Strategy to Win” ™ EPYC CPU Instinct GPU EPYC CPU Personal Enterprise _... | Co
Slide 71: End-to-End Integrated AI Platform
End-to-End Integrated AI Platform Front End Network CPU Node GPU Node Scale Up Network Scale Out Network Securely Integrate X86 Applications AI Model Training Connectivity for Training Enable Gigawatt into Enterprise & AI Execution & Inference & Distributed Inference Level Scaling
Slide 72: Front-End Networks Feed AI
Front-End Networks Feed AI Securely bridge AI servers Boost AI server performance with DPU accelerates network functions to enterprise data & apps network & security offloads up to 40x versus CPU only Source: AMD Internal Estimate
Slide 73: High Performance CPUs Drive Agentic AI
High Performance CPUs Drive Agentic AI Leverages x86 ecosystem to 5Ghz operation boosts Robust Enterprise class reliability, power agentic use-cases GPU operating efficiency availability & serviceability
Slide 74: AMD EPYC Driving End-to-End System Performance
AMD EPYC Driving End-to-End System Performance Llama 3.1 8B Llama 3.1 70B Mixtral 70B Up to 6% Uplift Average 11% Uplift Average 17% Uplift Average Relative Performance Chatbot Content Summarization Translation Chatbot Content Summarization Translation Chatbot Content Summarization Translation Creat
Slide 75: Optimized for Performance
Optimized for Performance High efficiency protocol for Up to 800 Gbps per port with Leverages 200 G ethernet low power & low latency <1uS multiple ports per accelerator physical layer & infrastructure
Slide 76: Protocol Interface, Transaction Layer, Data Link Layer
Protocol Interface Transaction Layer Data Link Layer Industry Standard Ethernet PHY Management Cabling Connectors Re-Timers Software
Slide 77: Engineered to Scale
Engineered to Scale Scales from SMB to Supports virtual POD Robust & reliable Accelerating AI models with Gigawatt Data Center partitioning with fault resiliency in-network collectives
Slide 78: Ultra Accelerator Link
ULTRA ACCELERATOR LINK" <)>
Slide 79: Ultra Accelerator Link, the Truly Open Standard
Ultra Accelerator Link, the Truly Open Standard UALINK NVLINK FUSION LOW ROUND TRIP LATENCY Yes Yes HIGH SPEED I/O 224 Gbps 224 Gbps MAXIMUM SCALABILITY 1024 GPUs 576 GPUs CPU USE Any Only to Nvidia GPU GPU USE Any Only Nvidia CPU MANAGEMENT SOFTWARE Open Nvidia Proprietary SPECIFICATION Fully Open
Slide 80: Any CPU, Any Accelerator, Any Switch
Any CPU Any Accelerator Any Switch
Slide 81: Open Standard Drives Innovation & Choice
Open Standard Drives Innovation & Choice 100+ Consortium Members
Slide 82: Promoters and Select Contributors
Promoters Select Contributors
Slide 83: Purpose-Built Connectivity for AI Infrastructure
Switch Fabric Signal Conditioning Controllers More… Chiplets ICs Modules Boards Software Copper Optical Purpose-Built Connectivity for AI Infrastructure
Slide 84: Rack-Scale AI
Optics Connectivity Switching CPO Scale-Up Fabric CPC Custom Silicon Rack-Scale AI
Slide 85: Open Architecture for Performant AI Scale-Out
Open Architecture for Performant AI Scale-Out Modern RDMA transport with Optimized protocol for high fabric Unparallel scalability to automatic configuration & tuning utilization & congestion control over a million GPUs
Slide 86: 100+ Consortium Members
Steering Committee Select General Members 100+ Consortium Members
Slide 87: Scaling Out to Gigawatt Data Centers
Scaling Out to Gigawatt Data Centers Programmability enables Load balancing reduces Line-rate engines boost Improved reliability collective acceleration networking cost flexibility & performance & resiliency
Slide 88: World Class Open Platform
World Class Open Platform Highest memory bandwidth & Industry standard x86 for easy Ready to deploy rack capacity drives performance integration to enterprise applications infrastructure AMD Instinct GPU AMD Server CPU AMD Pensando NIC MI350 SERIES EPYC “TURIN” POLLARA 400
Slide 89: MI350 Series Rack Specifications vs GB200 NVL72
96 / 128 GPUs Up to 1.8x more than GB200 NVL72 36 TB HBM3E Up to 2.8x more than GB200 NVL72 2.6 EF FP4 1.3 EF FP8 Up to 1.6x more than GB200 NVL72 2.6 EF FP6 Up to 3.3x more than GB200 NVL72 x86 CPU 5th Gen EPYC “Turin” Ultra Ethernet Consortium Scale-Out NIC AMD Pollara NIC OCP Compliant Industry s
Slide 90: Advancing AI Infrastructure Solutions
Advancing AI Infrastructure Solutions 2026 2025 2024 Roadmap subject to change
Slide 91: The World's Best AI Rack Solution
Previewing Today at Advancing AI 2025 The World’s Best AI Rack Solution For At-Scale Training & Distributed Inference
Slide 92: AMD "Helios"
Previewing Today at Advancing AI 2025 AMD “Helios” Optimized AI Rack Solution Available in 2026
Slide 93: AMD "Helios" AI Rack
AMD “Helios” AI Rack Rack Scale AI Performance Leadership AMD Instinct MI400 Series vs. Vera Rubin “Helios” Oberon GPU DOMAIN 72 1.0x SCALE UP BANDWIDTH 260 TB/s 1.0x FP4 • FP8 FLOPS 2.9 EF • 1.4 EF 1.0x HBM4 MEMORY CAPACITY 31 TB 1.5x MEMORY BANDWIDTH 1.4 PB/s 1.5x SCALE OUT BANDWIDTH 43 TB/s 1.5x
Slide 94: AMD EPYC "Venice"
AMD EPYC “Venice” Highest Performance Server CPU Up to 256 cores 2.0x 1.7x 1.6 TB/s 2nm • Zen 6 CPU to GPU Bandwidth Gen vs. Gen Performance Memory Bandwidth Coming in 2026 See endnote: VEN-003
Slide 95: AMD Instinct MI400
AMD Instinct MI400 Leadership Gen AI Accelerator 40 PF 20 PF 432 GB 19.6 TB/s 300 GB/s FP4 • FP8 Flops HBM4 Memory Capacity Memory Bandwidth Scale Out Bandwidth / GPU Coming in 2026 Specs based on engineering projections
Slide 96: AMD Pensando "Vulcano"
AMD Pensando “Vulcano” Next Gen NIC for AI Clusters 3nm 800G Up to 8x UAL PCIe® Process Node Network Throughput Scale Out Bandwidth per GPU Host Interface Coming in 2026
Slide 97: AI Compute Performance
AI Compute Performance AMD Instinct MI400 SERIES AI Compute Performance AMD Instinct MI355X AMD Instinct AMD Instinct MI300X MI325X Time See endnote: MI350-047
Slide 98: Up to 10x More Performance with AMD Instinct MI400 Series
Up to 10x More Performance with AMD Instinct MI400 Series Compared to Instinct MI355X See endnote: MI400-001
Slide 99: Advancing AI Infrastructure on an Annual Cadence
Advancing AI Infrastructure on an Annual Cadence 2027 2026 2025 Next Gen AI Rack “Helios” Roadmap subject to change
Slide 100: Accelerating Momentum Across Our Broad AI Capabilities
Announced Today at Advancing AI 2025 Accelerating Momentum Across Our Broad AI Capabilities AMD Instinct AMD ROCm 7 Optimized Empowering AMD “Helios” AMD Threadripper and MI350 Series for Reasoning & Agents Developers AI Rack Radeon AI solutions In Production Now Available August Available Now Comin
Slide 101: AI Innovation is a Global, Collective Effort
AI Innovation is a Global, Collective Effort
Slide 102: Endnotes
Endnotes SHO-06: Testing as of Dec 2024 using the following benchmark scores compared to Intel Core Ultra 9 288V and Qualcomm Snapdragon X Elite X1E-84-100. Cinebench 2024 nT, 3Dmark Wildlife Extreme, and Blender.. Next gen AI PC defined as a Windows PC with a processor that includes a NPU with at l
Slide 103: Endnotes
Endnotes MI350-031: Based on calculations by AMD internal testing as of 6/4/2025. Using 8 GPU AMD Instinct MI355X Platform for overall GPU-normalized Training Throughput (processed tokens per second) for text generation using both LLaMA3-70B and LLaMA3-8B chat models running Torchtitan (BF16) or Meg
Slide 104: Endnotes
Endnotes MI350-042: Based on AMD internal testing as of 6/5/2025. Using 8 GPU AMD Instinct MI355X Platform measuring text generated offline inference throughput for Llama 3.1-405B chat model (FP4) compared 8 GPU AMD Instinct MI300X Platform performance with (FP8). MI355X ran 8xTP1 (8 copies of model
Slide 105: Endnotes
Endnotes MI350-54: Based on calculations by AMD internal testing as of 6/10/2025. Using 8 GPU AMD Instinct MI355X Platform for overall GPU-normalized Training Throughput (time to complete) for fine-tuning using the Llama2-70B LoRA chat model (FP8) compared to published 8 GPU Nvidia H200 Platform per