Slide 1: NVIDIA
1 NVIDIA
Slide 2: Forward-looking statements
@GANVIDIA 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
Slide 3: Customer and partner logos
accenture ADEPT I,\ Adobe C@Alibaba Cloud amazon Argonne Ansys ANTHROP\C Sb anyscale 7 AUTODESK Baltes BlackRock BM BNP PARIBAS sostondynamics Pd Bristol Myers Squibb e.ai cadence Bicanonical Cana capital Oe Oy & EAR ML & cohere CU CoreWeave @ntinental+. <databricks Fal vatavoc data P dataloop @ Dat
Slide 4: Customer and partner logos
amazon a BlackRock BMW BNP PARIBAS UU Bristol Myers Squibb capitate aalotey Gntinentals “Denep Research ebay ExonMobil E> Foster + Partners @Q«- theore gettyimages’ Genentech Goldman GSK Honeywell JR (@ Jonn Deere Heese Means JPMORGAN CHASE & Co. H2Kawasaki KPF kt Linked] = LockHeep manTiINZe LOREAL
Slide 5: A new industrial revolution timeline
1964 IBM S/360 2022 OPEN AI CHATGPT Cm FINE Tu! GUARDRAILING 1995 WINDOWS 95 PENTIUM LEARN EVERYTHING PROTEIN LANGUAGE SOUND PHYSICS neg 80 MANIPULATION IMAGES GESTURE NING ALIGNMENT PROMPT ENGINEERING — A vector pp RAG MULTI-MODAL COT & TOT AGENTS, GENERATIVE Al ACCELERATED COMPUTING 1993 2003 2006
Slide 6: Slide 6
Slide 7: Slide 7
Slide 8: Ansys and NVIDIA
Ansys @INvIDIA Building a New Era of CAE with Accelerated Computing & Generative Al ‘ANSYS SOLVERS NVIDIA Al ACCELERATED. PHYSICS-BASED NVIDIA POWERED ‘TRANSFORM 6G RESEARCH ACCELERATED NVIDIA OPTIMIZED FOR ANSYS SIMULATION DIGITAL TWINS ANSYS LLMS. ‘WITH ANSYS PERCEIVE EM DEVELOPMENT NVIDIA GPUS IN
Slide 9: Synopsys and NVIDIA
ted oe Synopsys’ GANvVIDIA SYNOPSYS: MISSION CRITICAL FOR NVIDIA SILICON SUCCESS DECADES OF COLLABORATION ACROSS FULL EDA SUITE POWERS ACCELERATED COMPUTING we Verification Functional Verification + Synopsys ves + NVIDIA Lao) + NVIDIA Grace Hopper ‘Pareemance Spee ip tse on Pretec eats Design Place
Slide 10: Cadence and NVIDIA
cadence + G§nvipDiIA Deep collaboration on EDA, SDA, Digital Biology and Al cadence.al NVIDIANNeMo NVIDIA Modulus Al GenAl Co- Pilots Principled Simulation + Optimization Accelerated Compute
Slide 11: Training Compute (petaFLOPs)
Training Compute PFLOPs GPT-MoE-1.8T 1010 MT NLG 530B 109 Chinchilla GPT3-175B BLOOM Training Compute (petaFLOPs) PaLM 108 Microsoft T-NLG GPT-2 1.5B 107 Megatron-NLG Wav2Vec 2.0 106 XLNet MoCo ResNet50 Xception InceptionV3 BERT Large 105 GPT-1 VGG-19 Resnet Transformer 104 ResNeXt Seq2Seq ELMo Dens
Slide 12: Selene to Eos supercomputers
4,480 A100 GPUs 10,752 H100 GPUs SELENE EOS 3 EF AI Compute 43 EF AI Compute 2021 2023 112 TB/s Interconnect BW 1,100 TB/s Interconnect BW
Slide 13: Slide 13
Slide 14: Announcing NVIDIA Blackwell Platform
ANNOUNCING NVIDIA BLACKWELL PLATFORM FOR TRILLION-PARAMETER SCALE GENERATIVE AI AI SUPERCHIP 2nd GEN TRANSFORMER ENGINE 5th GENERATION NVLINK 208B Transistors FP4/FP6 Tensor Core Scales to 576 GPUs RAS ENGINE SECURE AI DECOMPRESSION ENGINE 100% In-System Self-Test Full Performance 800 GB/sec Encrypt
Slide 15: Blackwell GPU
Blackwell GPU FP8 20 PFLOPS 2.5X Hopper NEW FP6 20 PFLOPS 2.5X NEW FP4 40 PFLOPS 5X HBM Model Size 740B param 6X HBM Bandwidth 34T param/sec 5X NVLINK All-Reduce with SHARP 7.2 TB/s 4X
Slide 16: 1000X AI Compute in Years
1000X AI Compute in 8 Years Blackwell 20,000 TFLOPS FP4 TFLOPS Hopper 4,000 TFLOPS FP8 Ampere Pascal Volta 620 TFLOPS 19 TFLOPS 130 TFLOPS BF16/FP16 FP16 FP16 2016 2017 2020 2022 2024
Slide 17: NVLink Switch Chip
NVLink Switch Chip 50B Transistors in TSMC 4NP 72-Ports Dual 200 Gb/sec SerDes 4 NVLinks at 1.8TB/sec 7.2TB/sec Full-Duplex Bandwidth SHARP In-Network Compute – 3.6 TFLOPS FP8
Slide 18: DGX GB200 NVL72
DGX GB200 NVL72 1 Giant GPU Training FP8 720 PFLOPS 22X Inference FP4 1.44 ExaFLOPS 45X Multi-Node All-to-All 130 TB/sec 18X Multi-Node All-Reduce 260 TB/sec 36X
Slide 19: Slide 19
Slide 20: Train GPT-MoE-1.8T on Hopper
Train GPT-MoE-1.8T in 90 Days Hopper 8000 GPUs | 15MW LLM Training Workload: GPT-MoE-1.8T | H100 vs GB200 NVL72
Slide 21: Train GPT-MoE-1.8T on Blackwell
Train GPT-MoE-1.8T in 90 Days Blackwell GB200 NVL72 2000 GPUs | 4MW 1/4th the Power LLM Training Workload: GPT-MoE-1.8T | H100 vs GB200 NVL72
Slide 22: GPT-MoE 1.8T Inference
GPT-MoE 1.8T Throughput per GPU Inference (seqlen=32k/1k, FTL=5s) Multi-Dimensional Optimization: Tokens per Second • Tensor Parallel • Pipeline Parallel 160 • Expert Parallel • Data Parallel 140 TP2.EP8.DP4 TP2.EP16.PP2 120 GB200 FP4 100 TP4.EP16 80 60 40 20 0 0 10 20 30 40 50 Interactivity per Use
Slide 23: Blackwell 30X Hopper inference
GPT-MoE 1.8T Throughput per GPU Inference (seqlen=32k/1k, FTL=5s) Multi-Dimensional Optimization: Tokens per Second • Tensor Parallel • Pipeline Parallel 160 • Expert Parallel • Data Parallel 140 TP2.EP8.DP4 TP2.EP16.PP2 120 GB200 FP4 B200 FP8 100 H200 FP8 TP4.EP16 80 TP4.EP2.PP2.DP4 60 Blackwell 30
Slide 24: Blackwell adopters
®Meta AiNRES SVIDEN | Lambda ADEPT aws — Google Cloud Microsoft ORACLE are infrastructure Al21labs PB character.al Scohere © essential A BD rwseing Face ema @openat fh perplexity @ Recursion TT together.ai AI armors MERock 75uS pee CY CoreWeave @ crusoe , <o — f e [AY coxcenn FUJITSU GIGABYTE Hewlet
Slide 25: AWS and NVIDIA expand joint AI offerings
aws —) k>| NVIDIA. AWS and NVIDIA expand joint Al offerings OVER A DECADE OF CO-INNOVATION IN THE CLOUD OPS | 20,736 68200 | GEN 4 AWS EFA Optimizing price performance of FMs running on accelerated computing iow elie NVIDIA NIM Transforming industries with the power of AWS and NVIDIA Abvances amazon
Slide 26: Google Cloud
Google Cloud Data science & analytics C==3 €22? 4x Al platforms & frameworks =D 4D C= CIID ax» Accelerated computing cI? €r x= DeepMind Gog 7 Gemma
Slide 27: Oracle and NVIDIA
ORACLE GYNVIDIA. Sovereign Al ORACLE Database ORACLE CLOUD Infrastructure Generative Al Enterprise Gen Al ORACLE Cloud Applications ORACLE CLOUD Infrastructure Supercluster Healthcare Al ORACLE CLOUD Infrastructure ORACLE Distributed Cloud
Slide 28: Microsoft and NVIDIA
EE Microsoft @GINVIDIA Powering the most Largest NVIDIA InfiniBand Most deeply integrated cloud sophisticated Al innovators connected supercomputer for NVIDIA DGX Cloud ADEPT Me™ A & ss MicROsOFT Nv FABRIC sx cLoup Inflection GOpenal = A icRosoFT ‘AZURE Microsoft and NVIDIA Announce A e D fe ° Micro
Slide 29: Slide 29
Slide 30: 2012 AlexNet First Contact
2012 ALEXNET “FIRST CONTACT”
Slide 31: Learn and understand everything
TEXT IMAGE LEARN AND UNDERSTAND TEXT EVERYTHING VIDEO IMAGE IMAGE VIDEO 3D TEXT SOUND SPEECH ANIMATION MULTI-MODAL MANIPULATION AMINO ACID PROTEIN BRAINWAVES SPEECH
Slide 32: Slide 32
Slide 33: The Weather Company and NVIDIA
The Weather | > | Company NVIDIA iil Tweo TWCo PROCESSING 40078 OF TWCo GRAF TWco WEATHER DATA PER DAY FORECAST MODELLING WEATHER ENGINE WEATHERVERSE Providing actionable insights to fertetordecabnmanhy Next-generation Al (One ofthe largest collections of forecast modeling Improve social resince to
Slide 34: NVIDIA Healthcare
NVIDIA HEALTHCARE IMAGING & ROBOTICS GENOMICS BIOINFORMATICS DRUG DISCOVERY MONAI | Holoscan Parabricks RAPIDS BioNeMo
Slide 35: Drug discovery workflow with AlphaFold2 and DiffDock
AlphaFold2 Protein Database User DiffDock Gradient-Free Optimization UniDock MolMIM Optimized Molecules
Slide 36: Slide 36
Slide 37: AI model builders
ADEPT _ Al21labs Google Google DeepMind Hugging Face Scohere Inflection MISTRAL j hail ai 6 Recursion. 00 Meta BE Microsoft S@ANvipia , XI
Slide 38: NVIDIA Inference Microservice
NVIDIA INFERENCE MICROSERVICE Pre-Trained AI Models Packaged and Optimized to Run Across CUDA Installed Base
Slide 39: NIM architecture and CUDA installed base
Industry Standard APIs Text, Speech, Image, Video, 3D, Biology TensorRT LLM and Triton cuBLAS , cuDNN, In-Flight Batching, Triton Inference Server Memory Optimization, FP8 Quantization cuDF, CV-CUDA, DALI, NCCL, Post Processing Decoder Optimized Model Single GPU, Multi-GPU, Multi-Node Cloud Native S
Slide 40: Top Open Foundation Models
‘Top Open Foundation Models.
Slide 41: AI application plan architecture
Prompt Event Vector Database Other Agents Structured Data Plan Application Platform (ERP, CRM) Knowledge Graph Application Platform
Slide 42: ChipNeMo Llama 2 70B before fine-tuning
Prompt Event Response NVIDIA ChipNeMo Llama 2 70B Before Fine-Tuning
Slide 43: ChipNeMo Llama 2 70B with fine-tuning
Prompt Event Response NVIDIA ChipNeMo Llama 2 70B With Fine-Tuning
Slide 44: NeMo Microservices
Prompt Event Custom Model in NIM NeMo Microservices Auto Scoring, Human Ratings, LLM-as-a-Judge Fine-Tuning Foundation with P-Tuning, NeMo Evaluator LoRA, SFT Model Alignment with Microservice RLHF, SteerLM Domain, Complexity, Task Classification NeMo Customizer Topical Content, PII, Toxicity Micros
Slide 45: NeMo Retriever
NeMo Retriever Vector Database Text Docs RAPIDS cuVS Embedding Accelerated Indexing Microservices PDF Image Retrieval Microservices RAPIDS cuVS Text-to-SQL Plan Accelerated NN Search Microservice
Slide 46: NeMo Retriever Chat with PDF
NeMo Retriever Vector Database Text Docs RAPIDS cuVS Embedding Accelerated Indexing Microservices PDF Image Chat with PDF Retrieval Microservices RAPIDS cuVS Text-to-SQL Plan Accelerated NN Search Microservice
Slide 47: NVIDIA DHT Plan
Prompt Event Vector Database NVIDIA Other Agents DHT Structured Data Plan Application Platform (ERP, CRM) Knowledge Graph Application Platform
Slide 48: SAP Business AI and NVIDIA
ww” NVIDIA. Relevant Reliable Responsible @ Joule ' A copilot that truly understands your Embedded Al capabilities ‘Supply Chain ‘Spend Management CustomerRelationship Business Technology & Business Network ‘Management Platform, Al Foundation ‘on Business Technology Platform SAP Business Al Ecosyste
Slide 49: ServiceNow
servicenow. THE INTELLIGENT WORKFLOW COMPANY Now Assist Generative Al Experi es powered by ServiceNow Platform Al Research & Engineering & ServiceNow Global Development Applied Science Cloud Services Enterprise Foundation Models Domain Specific Al Models Al Embedded Workflows Al Trustworthiness z Pr
Slide 50: Cohesity Gaia
COHESITY | ZA Cohesity Gaia — First to Provide Al-Powered Business Insights from Secondary Data
Slide 51: Snowflake: Data + AI, all on NVIDIA
lt;snowflake = DATA +Al— ALLON NVIDIA SNOWPARK SNOWFLAKE CORTEX CONTAINER SERVICES Serverless Al, LLMs, Search Model Training, Fine Tuning == ENTERPRISE DATA
Slide 52: NetApp Unlocks Exabytes of Data for Gen AI
NetApp Unlocks Exabytes of Data for Secure, Private Gen Al NETAPP CUSTOMER DATA CHATBOTS 5 ON-PREMISES APPLICATIONS Applic NETAPP CUSTOMERS CAN NOW “TALK TO THEIR DATA” 3) Fi NetApp
Slide 53: Dell AI Factory with NVIDIA
Dell Al Factory with NVIDIA INDUSTRY'S FIRST END-TO-END ENTERPRISE Al SOLUTION Model Lifecycle Dell Professional Services Data Validation Deployment Implementation Dell and NVIDIA Infrastructure Broadest Al solutions portfolio from desktop to data center to cloud, all in one place
Slide 54: DGX, OVX and AGX Omniverse stack
DGX AI OVX DRIVE SIM OMNIVERSE IN DIGITAL TWIN OMNIVERSE AGX ROBOTIC SYSTEM
Slide 55: Warehouse digital twin architecture
AMR Robot Digital Twins Sensor Fusion Isaac Sim Metropolis USD APIs Occupancy APIs Warehouse Digital Twin Route Planning cuOpt APIs NVIDIA NVIDIA Omniverse Cloud DGX Cloud Isaac Sim cuOpt Worker Digital Twins Operator UI Omniverse Channel APIs VLM APIs Isaac Sim Metropolis
Slide 56: i % S
i 2 % S
Slide 57: NVIDIA Omniverse Cloud application
Application Omniverse USD APIs Channel APIs Write | Render | Query | Notify NVIDIA Omniverse Cloud
Slide 58: NVIDIA Omniverse Cloud with generative AI
Prompt Application 3D USD 360 HDRI Edify Materials Omniverse USD APIs Channel APIs USD ChatUSD USD DeepSearch NVIDIA USD Omniverse Cloud Data Lake
Slide 59: Siemens and NVIDIA immersive digital twins
Immersive digital twins x © Siemans NVIDIA Omniverse Xcelerator Cloud ———— Industrial Edge Al & Robotics x Siemens NVIDIA, Xeelerator Robotics SIEMENS @3NVIDIA Building the Next Era of Industrial Digitalization Together Generative Al x Siemens NOIA eelerator ALEntorprise Industrial Al-Physics Simula
Slide 60: Slide 60
Slide 61: NVIDIA Omniverse Cloud
Omniverse Channel APIs USD APIs NVIDIA Omniverse Cloud
Slide 62: Omniverse Cloud application example
Team Hyundai GENERAL D dashboard £B) workspaces ‘ORDERS New Asset = My orders D oat PRESETS LD) Templates USER MANAGEMENT Po Members « » so & 2 & g Environments iv) (= Q 3D 360° 2D REMOVE SAVE AS PRESET Vertical Horizon Offset Environment Rotation Breathtaking swiss mountain landscape with a large o
Slide 63: Apple Vision Pro and Omniverse Cloud
Apple Vision Pro NVIDIA Omniverse Cloud
Slide 64: Autonomous vehicle platform: DRIVE Thor
DGX MSFNET ROADNET PLANNET OVX DRIVE SIM IN OMNIVERSE AGX DRIVE AV STACK DRIVE THOR ASIL-D AV COMPUTER & STACK
Slide 65: NVIDIA Robotics Platform Adoption
DGX ISAAC NVIDIA Robotics Platform Adoption PERCEPTOR 1.3M Robotics Developers 100,000 ROS Developers 6,000 Companies Developing on Orin OVX ISAAC DRIVE SIM IN OMNIVERSE AGX ISAAC ROBOTIC STACK JETSON ORIN ROBOTIC COMPUTER & STACK
Slide 66: Robot manipulator stack
SyntheticaDETR FoundationPose DGX POSE FoundationGrasp cuMotion GRASP MANIPULATE OVX ISAAC DRIVE SIM IN OMNIVERSE AGX ISAAC MANIPULATOR STACK JETSON ORIN ROBOTIC COMPUTER & STACK
Slide 67: GR00T Foundation Model
Language Videos GROOT Model Actions Humanoid Robot DGX Demonstrations Observations GR00T FOUNDATION MODEL OVX ISAAC DRIVE LAB SIM IN OMNIVERSE AGX GR00T STACK JETSON THOR ROBOTIC COMPUTER & STACK
Slide 68: Slide 68
Slide 69: A New Industrial Revolution
$100T GENERATIVE AL _——— ACCELERATED ‘COMPUTING. NEW INDUSTRY BLACKWELL PLATFORM A NEW INDUSTRIAL REVOLUTION NIMs NEMO AND NVIDIA AI FOUNDRY OMNIVERSE AND ISAAC ROBOTICS
Slide 70: NVIDIA Blackwell Platform
NVIDIA Blackwell Platform Spectrum X800 Switch Quantum X800 Switch BlueField-3 SuperNIC ConnectX-8 SuperNIC GB200 Superchip Compute Node NVLINK Switch HGX B100
Slide 71: NVIDIA
1 NVIDIA