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, the evolving AI landscape, AMD’s ability to adva
Slide 2: Advancing AI 2024
ADVANCING AI2024 AMD«\ together we advance_Al
Slide 3: Computing at the heart of modern life
Computing at the heart of modern life Cloud Healthcare Industrial Automotive Connectivity PCs Gaming AI
Slide 4: Most transformational technology in years
AI Most transformational technology in 50 years Agents Smarter Cities Robotics Healthcare Research Supply Chain
Slide 5: Unmatched AI ecosystem
Unmatched AI ecosystem AI Platforms portfolio of training and inference compute engines Open software solutions with deep co-innovation Cluster level systems design
Slide 6: Evolving AI Landscape
Evolving AI Landscape Innovation moving from silicon to nodes to racks to clusters Silicon Server Rack Data Center
Slide 7: Advancing the AI Data Center
Advancing the AI Data Center CPUs GPUs Networking Software Solutions Cluster Level AMD EPYC AMD Instinct DPUs, UALink + Ultra Ethernet Open Software Stack Systems Design Announced acquisition expected to close 1H 2025
Slide 8: Leadership Engines for Enterprise AI Workloads
Leadership Engines for Enterprise AI Workloads Data Data Model Input Cleaning Pre-Processing Training Deployment x86 CPU GPU Data from LLM Training x86-Based Systems Agentic AI and Inference From analytics to generative AI to agentic AI
Slide 9: AMD EPYC record market share…and growing
AMD EPYC record market share…and growing 4th Gen Processor 34% Fami ly 31% H1’24 3rd Gen 27% Processor Fami ly 2nd Gen Processor 1st Gen Processor Fami ly 8% Fami ly 2% 2018 2019 2020 2021 2022 2023 2024 2025 >350 OEM Platforms >950 Cloud Instances Source: Mercury Research Sell-in Revenue Shipment E
Slide 10: #1 CPU for hyperscalers
#1 CPU for hyperscalers Hyperscale leaders power internal workloads with AMD, serving billions worldwide
Slide 11: Trusted by industry leaders on-prem
Trusted by industry leaders on-prem Use of third-party marks/l ogos/products is for informati onal purposes only and no endorsement of or by AMD i s i ntended or impl ied GD-83
Slide 12: AMD Instinct MI300 Series
AMD Instinct MI300 Series Powering the most popular Gen AI platforms
Slide 13: Solutions from leading OEMs and cloud
Right and Left Screen Solutions from leading OEMs and cloud
Slide 14: AMD end-to-end AI infrastructure leadership
Today at Advancing AI 2024 AMD end-to-end AI infrastructure leadership Data Center CPUs Data Center GPUs Networking AI PCs 5th Generation AMD Instinct AMD Pensando 3rd Generation AMD EPYC “Turin” MI325X and MI350 Series Pollara 400/Salina 400 AMD Ryzen AI PRO
Slide 15: AMD end-to-end AI infrastructure leadership
Today at Advancing AI 2024 AMD end-to-end AI infrastructure leadership Data Center CPUs 5th Generation AMD Instinct AMD Pensando 3rd Generation AMD EPYC “Turin” MI325X and MI350 Series Pollara 400/Salina 400 AMD Ryzen AI PRO
Slide 16: 5th Gen AMD EPYC
Launching Today: “Turin” 5th Gen AMD EPYC World’s best CPU for Cloud, Enterprise and AI 3nm 150 billion Up to 192 cores 17% IPC uplift Up to 5 GHz 4nm transistors 384 threads Full AVX512 *~17% Across 36 cloud and enterprise workloads As of 10/1/2024. See endnotes 9xx5-001, EPYC-029C
Slide 17: Turin Continues AMD EPYC Leadership
Consistent x86 ISA Consistent IPC SP5 Socket “Genoa” Compatible “Turin” Continues 8 to 192 cores AMD EPYC Leadership 125W to 500W Up to 12Ch DDR5-6400 128 PCIe® 5.0/CXL® 2.0 Confidential Compute with Trusted I/O As of 10/1/2024. See endnotes 9xx5-048, 9xx5-083: 5th Gen EPYC processors support DDR5-6
Slide 18: Scale-Up
Scale-Up Up to 16 “Zen 5” CCDs 128 Cores 256 Threads •
Slide 19: Scale-Out
Scale-Out Up to 12 “Zen 5c” CCDs 192 Cores 384 Threads •
Slide 20: Fifth Generation of CPU Leadership Starts Today
Fifth Generation of CPU Leadership Starts Today 6x Core Count Across 5 Generations 10.9x Performance Across 5 Generations Relative SPECrate2017_int_base best score by end of launch year 192 Cores 384 vCPU 10.9x 128 Cores 256 vCPU 6.6x 64 Cores 64 Cores 128 vCPU 128 vCPU 3.2x 32 Cores 2.6x 64 vCPU 1.
Slide 21: Industry's Highest Performing Server CPU
Industry’s Highest Performing Server CPU AMD EPYC 192 cores 2.7 2.7x 5th Gen 9965 AMD EPYC 4th Gen 9754 128 cores 1.7 Intel Xeon® 5th Gen 8592+ 64 cores 1.0 vs. top-of-stack “Emerald Rapids” SPECrate®_2017_int_base As of 10/01/2024. See endnote 9xx5-002c.
Slide 22: 60% More Performance at the Same Licensing Cost
60% More Performance at the Same Licensing Cost AMD EPYC 64 cores 1.6 1.6x 5th Gen 9575F up to AMD EPYC 4th Gen 9554 64 cores 1.2 Intel Xeon® 5th Gen 8592+ 64 cores 1.0 Performance per core in virtualized infrastructure Virtualized Infrastructure VMmark® 4.0 As of 10/4/2024. See endnote 9xx5-071.
Slide 23: Grow Your Database and Media Processing Capabilities
Grow Your Database and Media Processing Capabilities 3.9x 4x 1.9x 1.9x up to 4x Open-Source Database Video Transcoding Throughput performance MySQL OLTP FFMPEG 5th Gen Intel® 4th Gen AMD 5th Gen AMD Xeon® 8592+ 64C EPYC 9654 96C EPYC 9965 192C As of 10/1/2024. See endnote 9xx5-005A and 9xx5-006.
Slide 24: Fastest CPU for the Most Challenging HPC Problems
Fastest CPU for the Most Challenging HPC Problems 3.9x 3.1x 1.1x 1.9x up to 3.9x Dense Linear Solver Molecular Dynamics Improved time to insight HPL GROMACS 5th Gen Intel® 4th Gen AMD 5th Gen AMD Xeon® 8592+ 64C EPYC 9654 96C EPYC 9965 192C As of 9/30/2024. See endnote 9xx5-038, 9xx5-039 and 9xx5-02
Slide 25: End-to-End AI and Inference Performance
End-to-End AI and Inference Performance 3.8x 3.0x 1.2x 1.7x up to 3.8x Machine Learning End-to-End AI AI performance on CPU XGBoost (Higgs) Workload Derived from TPCx-AI 5th Gen Intel® 4th Gen AMD 5th Gen AMD Xeon® 8592+ 64C EPYC 9654 96C EPYC 9965 192C As of 9/30/2024. See endnote 9xx5-040a and 9xx
Slide 26: 1,000 legacy servers
1,000 legacy servers 2P Intel® Xeon® Platinum 8280 servers To deliver 391,000 unit of integer performance Servers required to achieve a total of 391,000 SPEC®rate_2017_ int_base performance score. AMD EPYC 9965 SPEC®rate_2017_int_base score is estimated
Slide 27: 131 modern servers
131 modern servers 2P AMD EPYC 9965 To deliver 391,000 unit of integer performance
Slide 28: 7:1 consolidation
7:1 consolidation Use the savings, space and power to grow your business 87% fewer servers ~ ~67% lower TCO 68% less power ~ 2P EPYC 9965 vs 2P Intel Xeon® 8280 to deliver 391,000 unit of integer performance As of 10/4/2024. See endnote 9xx5TCO-002a.
Slide 29: AMD end-to-end AI infrastructure leadership
Today at Advancing AI 2024 AMD end-to-end AI infrastructure leadership Data Center GPUs 5th Generation AMD Instinct AMD Pensando 3rd Generation AMD EPYC “Turin” MI325X and MI350 Series Pollara 400/Salina 400 AMD Ryzen AI PRO
Slide 30: Data Center AI Accelerators market
45B $ >60% CAGR Data Center AI Accelerators 500B $ Source: AMD 2023 2028
Slide 31: AMD Instinct MI300 GPU
AMD Instinct MI300 GPU Fastest ramping product in AMD history
Slide 32: Advancing ROCm performance and ecosystem
Advancing ROCm performance and ecosystem ~ 2x ~ 3x * Improvement in inference ROCm now supports 1M+ Deepening partnership Growing AI expertise and and training performance models out of box* with AI ecosystem customer implementations See endnote MI300-62. ROCm 6.2 vs. ROCm 6.0. HuggingFace Models to
Slide 33: Demonstrated inference leadership at key customers
Demonstrated inference leadership at key customers Nvidia 1.3x H100 up to Conversational AI 1.3x Content Generation 1.3x Higher performance across key workloads AI Agent & Chatbot 1.2x Mistral Mixtral Summarization 1.1x CommandR Llama 3.1 • 405B Latency Improvement See endnotes MI300-053, MI300-054,
Slide 34: AMD Instinct MI325X GPU
Launching Today AMD Instinct MI325X GPU Extending generative AI leadership 256GB HBM3E 6TB/s 1.3 PF 2.6 PF 1.8x memory 1.3x bandwidth 1.3x FP16 1.3x FP8 Compared to Nvidia H200. See endnotes MI325-001a, MI325-002.
Slide 35: AMD Instinct MI325X GPU Leadership inference performance
AMD Instinct MI325X GPU Leadership inference performance 1.4x ~ ~ 1.2x ~1.3x Mixtral Meta Llama-3.1 Mistral 8x7B 70B 7B Throughput Latency improvement Latency improvement Nvidia AMD Instinct H200 MI325X See endnotes:MI325-004, MI325-005, MI325-006
Slide 36: MI325X Platform
2 TB HBM3E 1.8x memory vs. H200 HGX 48 TB/s Memory Bandwidth 1.3x memory bandwidth vs. H200 HGX 10.4 PF FP16 1.3x compute flops vs. H200 HGX AMD Instinct 20.8 PF FP8 MI325X Platform 1.3x compute flops vs. H200 HGX See endnotes MI325-001A, MI325-002 *Dense flops
Slide 37: AMD Instinct MI325X Platform
AMD Instinct MI325X Platform Leadership Inference performance using 8x MI325X ~ 1.4x ~ 1.2x up to 1.4x Inference performance Meta Llama-3.1 Meta Llama-3.1 405B 70B Throughput Latency improvement Nvidia AMD Instinct H200 HGX MI325X Platform See endnotes MI325-015, MI325-014
Slide 38: World-Class Training Performance
World-Class Training Performance Single GPU and 8 GPU Training 1x GPU 8x GPU ~ 1.1x 1x ~ H200 MI325X H200 MI325X HGX Instinct Platform Meta Llama-2 Meta Llama-2 7B 70B Nvidia AMD Instinct H200 HGX MI325X Platform See endnotes MI325-013, MI325-012.
Slide 39: AMD Instinct MI325X GPU
AMD Instinct MI325X GPU Production starting in Q4 2024 Available from leading system and infrastructure solution partners starting Q1 2025
Slide 40: AMD Instinct Annual Roadmap Cadence
AMD Instinct Annual Roadmap Cadence AMD Instinct MI350 AMD Instinct SERIES AMD Instinct MI325X MI300X 2025 2024 2023 Roadmap subject to change
Slide 41: AMD Instinct MI350 Series
Previewing today AMD Instinct MI350 Series Continued Gen AI Leadership 3nm Up to 288GB FP4 / FP6 Process Node HBM3E Datatype Support Planned availability 2H 2025
Slide 42: Generational performance leap
Generational performance leap 35x AMD Instinct CDNA 4 ~ AMD Instinct CDNA 3 Inference projected performance – Throughput / GPU ~ 7x ~ 1.5x FP4 and FP6 Improved AI Compute Memory Capacity / Bandwidth New Datatypes Network Efficiency 7x AI compute compares MI300x FP16 to MI355x FP4. See endnotes MI355
Slide 43: MI355X vs MI300X new datatypes and compute
New Datatypes ~ 3.5x AI Flops vs. FP8 AI Compute ~ 1.8x FP16 / FP8 memory capacity HBM3E Memory ~ 1.5x memory bandwidth AMD Instinct MI355X vs AMD Instinct MI300X. See endnotes: MI355-005, MI355-002. Actual result s may vary.
Slide 44: AMD Instinct MI355X Accelerator
AMD Instinct MI355X Accelerator Leadership performance for Gen AI 1.8x ~ ~ 1.8x 9.2 PF Introducing FP6 and FP4 AMD Instinct AMD Instinct 1.3 PF 2.6 PF MI325X MI355X FP16 FP8 See endnotes:MI355-005
Slide 45: Leadership roadmap commitment continues
Leadership roadmap commitment continues 2026 2025 2024 2023 MI400 SE RIES MI350 SE RIES MI325X Next-Gen Architecture MI300X Compute and HBM3E Memory Memory Leadership Increased Compute Gen AI Leadership Roadmap subject to change
Slide 46: Enabling open innovation at scale
Enabling open innovation at scale
Slide 47: Open, modular software stack
Open, modular software stack AI Models and Algorithms Support all major frameworks and models Libraries Expanded Gen AI optimizations New algorithms Compilers and Tools New libraries Runtime Expanding platform support AMD GPUs Extended developer support Use or mention of third-party marks, logos, pr
Slide 48: Deep collaboration with developer community
Deep collaboration with developer community vLLM SGLang JAX Tensor Flow Day 0 support for Vendor agnostic Nightly CI/CD ensuring all Increasing open-source contributions latest features compiler support models work out-of-box and expanding footprint Use or mention of third-party marks, logos, produc
Slide 49: AMD Instinct MI300X Accelerator
AMD Instinct MI300X Accelerator Performant out-of-box support on popular generative AI models Llama 3 405B 1M+ models supported latency out of the box improvement MI300X vs. H100 Extended support Day 0 support Leadership performance for leading models for AMD GPUs on popular models
Slide 50: AMD Instinct MI300X Accelerator
AMD Instinct MI300X Accelerator Expanding out-of-box support on popular generative AI models GPT-4 GPT-4o Mixtral Grok 1 Zamba Flux CommandR MPM4 OLMo Aria Use or mention of third-party marks, logos, products, serv ices, or solutions herein is for informational purposes only and no endorsement by AM
Slide 51: Generational inference improvement
Generational inference improvement ROCm 6.2 vs. ROCm 6.0 ~ 2.6x 2.6x ~ 2.8x ~ 2.1x ~2.4x average ~1.9x ~ performance improvement Mixtral 8x22B Mixtral 8x7B Qwen2 72B Llama3.1 70B Llama3.1 8B Runtime Optimization Kernel Fusion Collective Communication Subgraph As of 9/30/2024. See endnote MI300-62.
Slide 52: Generational training improvement
Generational training improvement ROCm 6.2 vs. ROCm 6.0 1.7x ~ ~ 1.9x ~1.9x ~1.8x average performance improvement Llama2 70B Qwen 1.5 14B Llama2 7B Flash Attention 3 Parallelization Strategy Kernels Scale out Efficiency As of 9/30/2024. See endnote MI300-63.
Slide 53: Solving the last mile of customer AI
200+ AI implementations Helping clients succeed in building AI driven products and solutions Solving the last mile Open-source base models of customer AI European language LLMs trained on LUMI supercomputer with AMD Instinct accelerators
Slide 54: Enabling open innovation at scale
Enabling open innovation at scale
Slide 55: Advancing Data Center Solutions
Advancing Data Center Solutions Data Center CPUs Data Center GPUs Networking
Slide 56: CPU Performance Enhances GPU Performance
CPU Performance Enhances GPU Performance AMD EPYC 5th Gen 9575F 5.0GHz 28% Faster processing for GPU Intel Xeon® 3.9GHz orchestration tasks 5th Gen 8592+ CPU Host Processor Speeds up data prep, memory copies, Max Frequency kernel launch and task orchestration
Slide 57: AMD EPYC 9575F
~ 700,000 more inference tokens/s on 1K node AI cluster running Llama3.1-70B AMD EPYC 9575F Purpose built for GPU host nodes 20% Up to faster training with Stable Diffusion XL V2 Llama 3.1: 8.8% more perf on 1000 Node Cluster of Turin + 8xMI300X vs Emerald Rapids + 8xMI300X on Llama3.1-70B with 128
Slide 58: Programmable DPU
Programmable DPU Evolving front-end network Enables: Enables: Enables: Faster Data Integration SDN and Security Zero CPU Overhead Storage offload Evolving network DPU accelerates infrastructure and acceleration infrastructure services services at line rate Secure multi-tenant access Dedicated CPU fo
Slide 59: Back-end networks drive AI system performance
Back-end networks drive AI system performance 75% % of time spent in networking 57% 38% 40% 35% 18% AI Model 1 AI Model 2 AI Model 3 AI Model 4 AI Model 1 AI Model 2 At an average 30% of training cycle time Communication accounts for 40%-75% of time is elapsed in waiting for networking with Training
Slide 60: Advancing Data Center Solutions
Advancing Data Center Solutions Data Center CPUs Data Center GPUs Networking
Slide 61: Ethernet is always the preferred choice
Ethernet is always the preferred choice 1,000,000+ GPU > 50% up to TCO Saving 48,000 GPU InfiniBand Ethernet InfiniBand Ethernet RoCEv2 RoCEv2 Total Cost of Ownership1 Scalability Lower is better Higher is better Sources: 1) 650Group Datacenter AI Networking and Server SmartNIC Forecast Reports 2Q24
Slide 62: The challenge of high network utilization
~ 95% The challenge of high network utilization < 50% AI backend networks drive sustained data transfers Intelligent Load Congestion Fast Failover and General Balancing Management Loss Recovery Purpose AI Back-End Network Network Network Utilization Source: Alibaba Cloud Predictable network: Built f
Slide 63: Performant Scalable Cost Effective
Evolve ethernet as an open, interoperable, high performance, full-communications stack architecture to meet the growing network demands of AI and HPC at scale UEC 1.0 Specification - Q1CY25 Performant Scalable Cost Effective
Slide 64: Ultra Ethernet Consortium Steering Members
Ultra£=thernet —=—— Consortium Steering Members AMDZ ARISTA @BROADCOM ably Pee eas intel © Meta BE Microsoft Enterprise eVIDEN an ates business ORACLE
Slide 65: General Members
General Members • Total 97 Members •
Slide 66: RDMA outperforms RoCEv2
RDMA outperforms RoCEv2 6x faster message completion time 5x faster collective completion time Intelligent packet spray and Path aware Selective retransmission in-order message delivery congestion avoidance and fast loss recovery * Reference: STrack: A Reliable Multipath Transport for AI/ML Clusters
Slide 67: 3rd Gen AMD P4 Engines
3rd Gen AMD P4 Engines Deliver network innovation at the speed of AI 120M Packets/s 400Gb/s 5M Connections/s Fully Programmable Wire Rate Concurrent Services (SDN, Security, Storage Acceleration)
Slide 68: AMD Pensando Salina 400
Announcing Today: “Salina” AMD Pensando Salina 400 Best DPU for evolving front-end networks 3rd Gen 400G 232 P4 MPU 2x DDR5 16 N1 AMD Pensando Software PCIe® Gen 5 Multi-Services 102GB/S Memory ARM Cores Compatible 2x400GE Bandwidth DPU choice for hyperscalers Up to 128 GB DDR
Slide 69: AMD Pensando Pollara 400
Announcing Today: “Pollara” AMD Pensando Pollara 400 Industry’s first ultra ethernet consortium ready AI NIC Programmable Up to 6x 400 Gbps Open Ecosystem Hardware Performance UEC Ready RDMA Pipeline Boost* Reduction in Job Completion Times High Availability Based on comparison of RDMA vs. RoCEv2. *
Slide 70: Solutions you can rely on for your business
Solutions you can rely on for your business 350+ 950+ cloud instances server platforms
Slide 71: AMD end-to-end AI infrastructure leadership
Today at Advancing AI 2024 AMD end-to-end AI infrastructure leadership AI PCs 5th Generation AMD Instinct AMD Pensando 3rd Generation AMD EPYC “Turin” MI325X and MI350 Series Pollara 400/Salina 400 AMD Ryzen AI PRO
Slide 72: Unprecedented transformational experiences with next-gen AI PCs
Unprecedented transformational experiences with next-gen AI PCs Enterprise Immersive Revolutionary Personal AI Productivity Collaboration Creation and Editing Assistance Business LLMs for data processing Live captions with translation Automated content generation Smart information retrieval Multilin
Slide 73: AMD Ryzen AI Leads the AI PC Era
AMD Ryzen AI Leads the AI PC Era Q2 2024 Q4 2023 Q1 2023 3rd Gen “Strix Point ” 2nd Gen “Hawk Point ” 1st Gen “Phoenix Point” 16 50+ 10 TOPS NPU TOPS NPU TOPS NPU See endnote GD-243.
Slide 74: Best Processors for Copilot+ Enterprise PCs
Best Processors for Copilot+ Enterprise PCs Enterprise AI Experiences Pretrain Quantize Deploy Simplified Development Framework Open Platform Top-Tier Performance Multi-Day Battery Life Ryzen AI PRO Security Reliability Manageability Designed for Enterprise AMD PRO TECHNOLOGY AMD PRO TECHNOLOGY AMD
Slide 75: AMD Ryzen AI PRO 300 Series
Announcing today AMD Ryzen AI PRO 300 Series First Copilot+ laptops enabled for enterprise PCs See endnote STXP-05
Slide 76: 1.4x multithreaded performance
Up to 1.4x multithreaded performance Comparing AMD Ryzen AI 9 HX PRO 375 vs. Intel® Core® Ultra 7 165H vPro on Cinebench 2024 nT. See endnote STXP-12
Slide 77: Up to 23hrs multi-day battery life
Up to 23hrs Multi day battery life AMD Ryzen AI 9 HX PRO 375, VIdeo Playback. See endnote STXP-30
Slide 78: Up to 9hrs battery life with Microsoft Teams
Up to 9hrs battery life with Microsoft Teams AMD Ryzen AI 9 HX PRO 375. See endnote STXP-32
Slide 79: Enterprise AI PC Application Ecosystem
Enterprise AI PC Application Ecosystem
Slide 80: End-to-End AI Infrastructure Leadership
Announced today at Advancing AI 2024 End-to-End AI Infrastructure Leadership Data Center CPUs Data Center GPUs Solutions AI PCs 5th Generation AMD Instinct AMD Pensando 3rd Generation AMD EPYC “Turin” MI325X Series Pollara 400/Salina 400 AMD Ryzen AI PRO Shipping Now Production Q4, Systems Q1 Coming
Slide 81: AMD
AMD«e\l
Slide 82: Endnotes
Endnotes 9xx5-001: Based on AMD internal testing as of 9/10/2024, geomean performance improvement (IPC) at fixed-frequency. 5th Gen EPYC CPU Enterprise and Cloud Server Workloads generational IPC Uplift of 1.170x (geomean) using a select set of 36 workloads and is the geomean of estimated scores for
Slide 83: Endnotes
Endnotes 9xx5-006: AMD internal testing as of 09/01/2024, on FFMPEG (Raw to VP9, 1080P, 302 Frames, 1 instance/thread, video source: https://media.xiph.org/video/derf/y4m/ducks_take_off_1080p50.y4m). System Configurations: 2P AMD EPYC 9965 reference system (2 x 192C) 1.5TB 24x64GB DDR5-6400 running
Slide 84: Endnotes
Endnotes 9xx5-039: AMD testing as of 09/18/2024.The detailed results show the average uplift of the performance metric (TFLOPS) of this benchmark for a 2P 192-Core AMD EPYC 9965 powered system compared to a 2P 64- Core Intel® Xeon® PLATINUM 8592+ powered system running select tests on Open-Source HP
Slide 85: Endnotes
Endnotes 9xx5-048: AMD EPYC 9005 Series processors require OEM enablement and a BIOS update from your server or motherboard manufacturer if used with a m otherboard designed for the SP5 socketed AMD EPYC 9004 Series processors. Contact your system manufacturer prior to purchase to determine compatib
Slide 86: Endnotes
Endnotes 99xx5TCO-002A: This scenario contains many assumptions and estimates and, while based on AMD internal research and best approxim ations, should be considered an example for information purposes only, and not used as a basis for decision making over actual testing. The AMD Server & Greenhous
Slide 87: Endnotes
Endnotes MI300-63: Testing conducted by internal AMD Performance Labs as of September 29, 2024 training performance comparison between ROCm 6.2 software with compared to ROCm 6.0 software both with Megatron-LM on systems with 8 AMD Instinct MI300X GPUs running Llama 2-7B, Llama 2-70B (4K), Qwen1.5-1
Slide 88: Endnotes
Endnotes MI325-02: Calculations conducted by AMD Performance Labs as of May 28th, 2024 for the AMD Instinct MI325X GPU resulted in 1307.4 TFLOPS peak theoretical half precision (FP16), 1307.4 TFLOPS peak theoretical Bfloat16 format precision (BF16), 2614.9 TFLOPS peak theoretical 8 -bit precision (F
Slide 89: Endnotes
Endnotes MI355-004: Calculations conducted by AMD Performance Labs as of September 26th, 2024 for the AMD Instinct MI300X GPU platform and AMD Instinct MI300X GPU platform performance comparing FP16, FP8 and FP4 datatypes. MI355X 8xGPU Platform Peak theoretical Half Precision (FP16) Performance - 18
Slide 90: Endnotes
Endnotes STXP-30: Video Playback Battery Life vs Intel Core Ultra 165H: Intel Core Ultra 7 165H: 13.1 or 13 hours and 6 min (100.0% Basel ine Score) - Score: AMD Ryzen AI 9 HX PRO 375: 23.6 or 23 hours and 36 minutes (23.6 / 13.1 = ~180% or 80%). Video Playback Battery Life vs Apple M3 Pro 12-Core:
Slide 91: Disclaimer and Attributions
[Public] DISCLAIMER AND ATTRIBUTIONS DISCLAIMER The information contained herein is for informational purposes only, and is subject to change without notice. While every precaution has been taken in the preparation of this document, it may contain technical inaccuracies, omissions and typographical