Slide 1: Forward-looking statements
Except for the historical information contained herein, certain matters in this presentation are forward-looking statements. These forward-looking statements and any other forward-looking statements that go beyond historical facts that are made in this presentation are subject to risks and uncertain
Slide 2: NVIDIA
G1 NVIDIA
Slide 3: Partner ecosystem logos
acer A ADLINK AD\ANTECH aetina ‘AlMobile aira > APMIC AppWorks #E ASE MSReck ® AVerMedia O/4 rxomrex BIgGo CHENBRO ChoS2Mo Tare ® COLORFUL COMPAL ae Coretronic DeepMentor D DeepRad.Al Avneuta 9 DMKTZ @inom BdeverFocus roves Bl poxconn Saianr = GIGABYTE— @cuiactouv = GGMI Q2Homec.al AS Vrrninies Cing
Slide 4: Slide 4
Slide 5: A new computing age is starting
1964 1995 2007 2007 IBM 5/360 WINDOWS 95 IPHONE INTERNET PENTIUM MOBILE CLOUD AT BI3E SS CLE Ce A NEW COMPUTING AGE IS STARTING
Slide 6: CPU scaling slows, GPU-accelerated computing
CPU SCALING SLOWS GPU- ACCELERATED COMPUTING . AND 200€ COMPUTE DEMAND GROWS. CUbs EXPONENTIALLY IS AEB t ACCELERATE EVERY APPLICATION
Slide 7: GPU speed-up, power and cost savings
GPU ~100X SPEED-UP = 3X POWER -1.5X COST ©OX PERF /S OR 98% SAVINGS 3OXPERF/W oR 97% SAVINGS BRS BRS “THE MORE YOU BUY... THE MORE YOU SAVE
Slide 8: NVIDIA CUDA libraries open new markets
cuDF cuOPT cuQUANTUM Parabricks Data cuLITHO Combinatorial QC Processing Aerial RAN Gene Simulation Computational Optimization AI Radio Sequencing Modulus Lithography AI Physics cuDNN Deep Learning NVIDIA CUDA NVIDIA CUDA LIBRARIES OPEN NEW MARKETS
Slide 9: NVIDIA announces pandas-cuDF in Google Colab
NVIDIA GOOGLE COLAB PANDAS-CUDF NVIDIA ANNOUNCES PANDAS-CUDF IN GOOGLE COLAB
Slide 10: mort
mort Af = pa. read parguet|
Slide 11: CUDA achieves virtuous cycle
CUDA INSTALLED BASE BIGGER SPEED-UP R&D SCALE END USERS t DEVELOPERS & APPLICATIONS BIG SPEED-UP CUDA S38 RMEEIR CUDA ACHIEVES VIRTUOUS CYCLE
Slide 12: Slide 12
Slide 13: AlexNet 2012: first contact
2012 ALEXNET “FIRST CONTACT"
Slide 14: First DGX delivered 2016
2016 FIRST DGX DELIVERED DATA
Slide 15: Transformer 2017
2017 TRANSFORMER
Slide 16: LLM: the big bang of AI
LLM Large Language Model 2029 CHAT GPT oodo0oo000000 i : a THE BIG BANG OF AT’ DATA chatoPr as < r) ° iz es 9 Gere: osama om Peni oe Sento Geen een Message chat . ° iss
Slide 17: A new industrial revolution
$100T I MANUFACTURING TRANSPORTATION HEALTHCARE COMPUTING ) OOooOoo000000 ——<—_—_>»> ERS “A NEW INDUSTRIAL REVOLUTION”
Slide 18: Generative AI drives full stack reinvention
SOFTWARE FACTORY Al FACTORY ——_—_ TOOLS SKILLS RETRIEVAL GENERATIVE INSTRUCTIONS LLMs cpu GPU EmtA LS SHS eee GENERATIVE AI DRIVES FULL STACK REINVENTION
Slide 19: NVIDIA Inference Microservice
MA NVIDIA INFERENCE MICROSERVICE Pre-Trained Al Models Packaged and Optimized to Run Across CUDA Installed Base
Slide 20: Industry-standard APIs
Industry-Standard APIs Text, Speech, Image, Video Biology Triton Inference Server cuDF, CV-CUDA, DALI, NCCL, Post-Processing Decoder Cloud -Native Stack GPU Operator, Network Operator Enterprise Management Health Check, Identity, Metrics, Monitoring, Secrets Management Kubernetes TensorRT-LLM and Tr
Slide 21: AI.NVIDIA.COM
Speech Digital Human Computer Vision Biology Simulation Language Regional Vision RAG Language Language AI.NVIDIA.COM
Slide 22: Generative AI enables digital nurse, customer service agents, tutors
NeMo Retriever Embedding Model Medical Vector Electronic Health Specialist Agent Database Records User Speech Agent Checklist Physician-Approved Specialist Agent Reference Tables Audio2Face LLM Privacy and Compliance Specialist Agent GENERATIVE AI ENABLES DIGITAL NURSE, CUSTOMER SERVICE AGENTS, TUTO
Slide 23: Slide 23
Slide 24: Announcing new RTX AI PCs
ASUS TUF A14 / A16 ASUS Zephyrus G16 ASUS ProArt PX13 / P16 MSI Stealth A16 AI+ RTX AI 200 RTX AI | 700 AI TOPS | 7X ANNOUNCING NEW RTX AI PCs Now Over 200 RTX AI Laptops | Up to 700 AI TOPS | 7X Generative AI
Slide 25: Model compute scale grows exponentially
PHYSICAL AI 2X/ 3 MON o- TRANSFORMER 2X%/ 6 MON 23> [] ALEXNET 2Xx/YR BARE RRL RER MODEL COMPUTE SCALE GROWS EXPONENTIALLY
Slide 26: NVIDIA Blackwell
AI SUPERCHIP TRANSFORMER ENGINE SECURE AI 5th GENERATION NVLINK RAS ENGINE DECOMPRESSION ENGINE 208B Transistors FP4/FP6 Tensor Core Full Performance Scales to 576 GPUs 100% In-System Self-Test 800 GB/sec Encryption and TEE NVIDIA BLACKWELL
Slide 27: 1,000X AI compute in 8 years
Blackwell 20,000 TFLOPS FP4 TFLOPS Hopper 4,000 TFLOPS FP8 Ampere 620 TFLOPS Pascal Volta Moore’s Law FP16 19 TFLOPS 130 TFLOPS FP16 FP16 2016 2018 2020 2022 2024 8 1,000X 1,000X AI COMPUTE IN 8 YEARS
Slide 28: 350X energy reduction to train GPT-4
Pascal 1,000+ GWh TFLOPS Volta 140 GWh Ampere 40 GWh Hopper Blackwell 13 GWh 3 GWh 2016 2018 2020 2022 2024 8 1/350 GPT4-1.8T 350X ENERGY REDUCTION IN 8 YEARS TO TRAIN GPT4-1.8T
Slide 29: 45,000X energy reduction to generate tokens
Pascal 17,000 Joules/Token TFLOPS Volta 1,200 Joules/Token Ampere 150 Joules/Token Hopper 10 Joules/Token Blackwell 0.4 Joules/Token 2016 2018 2020 2022 2024 8 1/45,000 GPT4-1.8T 45,000X ENERGY REDUCTION IN 8 YEARS TO GENERATE GPT4-1.8T TOKENS
Slide 30: DGX Blackwell
DGX DGX BLACKWELL HOPPER NVLink Domain 72 8 9X NVLink BW (TB/s) 130 7 18X AI FLOPS (PF) 1,440 32 45X Power (kW) 100 10 10X
Slide 31: NVLink Switch Chip
NVLink Switch Chip 50B Transistors in TSMC 4NP 72-Ports 400G SerDes 4 NVLinks at 1.8TB/sec 7.2TB/sec Full-Duplex Bandwidth SHARP In-Network Compute – 3.6 TFLOPS FP8
Slide 32: if ul
if ul
Slide 33: NVIDIA Spectrum-X
BlueField-3 400G SuperNIC Spectrum-X800 Ethernet Switch 1.6X Effective Bandwidth NVIDIA SPECTRUM-X Supercharging Ethernet for Generative AI
Slide 34: Annual Spectrum-X rhythm
Spectrum-X1600 102.4T – 512 Radix CX9 1600G SuperNIC Millions of GPUs Spectrum-X800 Ultra TFLOPS 51.2T – 512 Radix CX8 800 SuperNIC 100’s of Thousands of GPUs Spectrum-X800 51.2T – 256 Radix BF3 400G SuperNIC 10’s of Thousands of GPUs 2016 2024 2018 2020 2025 2022 2024 2026 SPECTRUM-X GPU ANNUAL SPE
Slide 35: Slide 35
Slide 36: Customer and partner logos
ews Google Cloud BB Microsoft ADEPT Al21labs B character.al S&cohere © essential A @® rwgeing race Inflection 00Meta Meee GoOpenat if perplexity fo] ecursion. ap together.ai xX n = iT — tthe AiV2ES Ty A2PUED OAL ASRock 7—*} dbf CY CoreWeave @ crusoe Ma 1 SVIDEN [eexcenn FUJITSU GIGABYTE Few Packre I
Slide 37: Hopper platform
Hopper Platform GPU Hopper GPU Hopper+ GPU 6S HBM3 6S HBM3e CPU Grace CPU NVLINK NVLink Switch 900 GB/sec NIC CX7 BF3 SuperNIC SuperNIC SWITCH Quantum-X400 InfiniBand Switch DATACENTER SCALE • ONE-YEAR RHYTHM • TECHNOLOGY LIMITS • ONE ARCHITECTURE
Slide 38: Blackwell platform
Blackwell Platform GPU Blackwell GPU 8S HBM3e CPU NVLINK NVLink 5 Switch 1800 GB/sec NIC CX8 SuperNIC SWITCH Spectrum-X800 Quantum-X800 Ethernet Switch Switch DATACENTER SCALE • ONE-YEAR RHYTHM • TECHNOLOGY LIMITS • ONE ARCHITECTURE
Slide 39: Blackwell and Blackwell Ultra
Blackwell Platform GPU Blackwell GPU Blackwell Ultra GPU 8S HBM3e 8S HBM3e 12H CPU NVLINK NVLink 5 Switch 1800 GB/sec NIC CX8 SuperNIC SWITCH Spectrum-X800 Quantum-X800 Spectrum Ultra X800 Ethernet Switch Switch Ethernet Switch 512-Radix DATACENTER SCALE • ONE-YEAR RHYTHM • TECHNOLOGY LIMITS • ONE A
Slide 40: Rubin platform
Rubin Platform GPU Rubin GPU Rubin Ultra GPU 8S HBM4 12S HBM4 CPU Vera CPU NVLINK NVLink 6 Switch 3600 GB/sec NIC CX9 SuperNIC 1600 Gb/sec SWITCH X1600 IB/Ethernet Switch DATACENTER SCALE • ONE-YEAR RHYTHM • TECHNOLOGY LIMITS • ONE ARCHITECTURE
Slide 41: Hopper, Blackwell and Rubin roadmap 2022-2027
Blackwell Platform Rubin Platform Hopper Platform GPU Blackwell GPU Blackwell Ultra GPU Rubin GPU Rubin Ultra GPU Hopper GPU Hopper+ GPU 8S HBM4 12S HBM4 6S HBM3 6S HBM3e 8S HBM3e 8S HBM3e 12H CPU Vera CPU Grace CPU NVLINK NVLink Switch NVLink 5 Switch NVLink 6 Switch 900 GB/sec 1800 GB/sec 3600 GB/
Slide 42: NVIDIA Blackwell platform
Spectrum-X800 Switch Quantum-X800 Switch BlueField-3 SuperNIC GB200 Superchip ConnectX-8 SuperNIC Compute Node NVLink Switch HGX B200 NVIDIA BLACKWELL NVIDIA BLACKWELL PLATFORM
Slide 43: Slide 43
Slide 44: Slide 44
Slide 45: Warehouse ecosystem and integrators
WAREHOUSE ECOSYSTEM INTEGRATOR AUTODESK ‘AutoCAD Coretronic YASKAWA @emustess DATALOGIC aetina — [- <4 = eos merin Lock A AUTODESK Inventor QsPiNcence a SOLOMON (UNIVERSAL Ie inter vision © MetAl SICK GIGABYTE Obrssuss ISAAC DRIVE HOLOSCAN METROPOLIS [ DIA OMNIVERSE
Slide 46: Slide 46
Slide 47: Factory software and ODM ecosystem
FACTOR eee wreesice FOXCONN a > SOFTWARE ODMS Asm —\ EPSON Ansys | ASPEED 771 AUTODESK cadence Ennoconn ~~ Cingrasys @FARobot a G creo aoc f kurtz ersa i ) 4 ie 1 — TRI SIEMENS SketchUp —V, Visual SIEMENS ISAAC DRIVE HOLOSCAN METROPOLIS © [ NVIDIA OMNIVERSE j
Slide 48: Slide 48
Slide 49: Slide 49
Slide 50: Slide 50
Slide 51: Slide 51
Slide 52: Slide 52
Slide 53: Isaac ROS 3.0 available now on GitHub
NVIDIA Isaac Manipulator NVIDIA Omniverse Isaac ROS 3.0 Available Now on Github NVIDIA AI
Slide 54: SIMATIC Pick AI with NVIDIA Isaac Manipulator
SIMATIC Pick AI NVIDIA Isaac Manipulator NVIDIA Omniverse NVIDIA AI
Slide 55: ArcBest with NVIDIA Omniverse
ArcBest NVIDIA Omniverse | WA NVIDIA Al
Slide 56: NVIDIA DRIVE AV and Omniverse
NVIDIA DRIVE AV NVIDIA Omniverse NVIDIA AI
Slide 57: GR00T stack with Omniverse
GR00T Stack NVIDIA Omniverse NVIDIA AI
Slide 58: Slide 58
Slide 59: AI factory, Blackwell, NIMs, Omniverse, robotics
Al FACTORY BLACKWELL NIMS. OMNIVERSE/ ROBOTICS MER Sop “A NEW INDUSTRIAL REVOLUTION”
Slide 60: Slide 60
Slide 61: NVIDIA
G1 NVIDIA