Slide 1: NVIDIA
©& NVIDIA
Slide 2: The Expanding Universe of HPC
THE EXPANDING UNIVERSE OF HPC JENSEN HUANG | SC19 2
Slide 3: Safe Harbor
SAFE HARBOR Forward-Looking Statements Except for the historical information contained herein, certain matters in this presentation including, but not limited to, statements as to: our strategies, growth, position, opportunities, and continued expansion; the performance, benefits and impact of our p
Slide 4: At the Intersection of Graphics, Simulation, AI
AT THE INTERSECTION OF GRAPHICS, SIMULATION, AI 4
Slide 5: 5
5
Slide 6: Computing for the Da Vincis of Our Time
COMPUTING FOR THE DA VINCIS OF OUR TIME CLIMATE GENOMICS ABCI LBNL | NVIDIA ORNL NUCLEAR WASTE CANCER DETECTION SUMMIT REMEDIATION ORNL | Stony Brook U. LBNL | PNNL Brown U. NVIDIA 42 NEW TOP 500 SYSTEMS FIRST AI SUPERCOMPUTERS FIRST EXASCALE SCIENCE 6
Slide 7: Full Stack Speed-Up
IndeX 2.1 OptiX 7.0 RAPIDS 0.10 CG RAPIDS AI CLARA METRO DRIVE ISAAC AERIAL Spark XGBoost cuTENSOR 1.0 cuSOLVER 10.3 cuBLAS 10.2 cuDNN 7.6 CUDA-X TensorRT 6.0 DALI 0.15 NCCL 2.5 CUDA 10.2 CUDA FULL STACK SPEED-UP 3x in 2 Years Amber Chroma GROMACS 2017 27 Hours GTC LAMMPS MILC 20 Hours NAMD 2018 QE
Slide 8: The Expanding Universe of HPC
THE EXPANDING UNIVERSE OF HPC AI Arm SIMULATION EDGE ANALYTICS Data Edge NETWORK Analytics EXTREME IO Cloud Extreme IO 8
Slide 9: Incredible Advances in AI
INCREDIBLE ADVANCES IN AI WRITING IMAGE GENERATION DIALOG 3D POSE CLASSIFICATION DENOISING TRANSLATION SEGMENTATION SUMMARIZATION ALEXNET OBJECT RECOGNITION BERT CNN TRANSFORMER Q&A CLASSIFICATION 2012 2019 9
Slide 10: GPU Computing Powers AI Advances
GPU COMPUTING POWERS AI ADVANCES Super Moore’s Law — From 600 to 2 Hours in 5 Years Two Distinct Eras of AI Training K80 SERVER 600 Hours Doubling 3.4 Months DGX 2 Hours Doubling 2 Years Time to Train (ResNet-50) #1 MLPERF — AI TRAINING + AI INFERENCE HPC COMPUTING CHALLENGE 10
Slide 11: NVIDIA AI End-to-End Platform
NVIDIA AI END-TO-END PLATFORM TRAINING CLOUD EDGE AI AUTONOMOUS MACHINES DGX HGX EGX AGX 11
Slide 12: AI for Science
AI FOR SCIENCE Design Space Exploration Inverse Problems DATA DATA ICF + MERLIN — Fusion LIGO — Gravitational Waves Real-time Steering Fast Approximation Faster Prediction Real-time Steering EXPERIMENTATION NEURAL ESTIMATION SIMULATION ANI + MD – Chemistry ITER – Fusion Energy 12
Slide 13: 13
13
Slide 14: Streaming AI
STREAMING AI LSST: 20 TB/day ECMWF: 287 TB/day 100x 1x Data Collected Data Transfer SKA: 16 TB/sec SOFTWARE-DEFINED STREAMING AI BUILD MODELS SENSORS PROCESSING 14
Slide 15: NVIDIA EGX Edge Supercomputing Platform
NVIDIA EGX EDGE SUPERCOMPUTING PLATFORM NGC Third-Party ISVs METROPOLIS DEEPSTREAM DECODE IMAGE DNN GRAPHICS ENCODE PROCESSING CUDA-X NVIDIA EGX STACK Kubernetes Networking Storage Security Powered by NVIDIA CUDA Tensor Core GPU | Secured Boot Root of Trust Cryptographic Acceleration for IPsec and T
Slide 16: Putting AI to Work
PUTTING AI TO WORK Clara Metropolis DRIVE Isaac Omniverse Aerial VERTICAL INDUSTRY FRAMEWORKS WORLD’S LARGEST DELIVERY SERVICE ADOPTS NVIDIA AI 16
Slide 17: NVIDIA EGX
NVIDIA EGX Edge Supercomputing Platform 17
Slide 18: Supercomputing Is Hard — Cloud HPC Is Expensive
SUPER COMPUTING IS HARD — CLOUD HPC IS EXPENSIVE SUPERCOMPUTING CLOUD CPU Instance 48 Hours, $152 Amber, Chroma, GROMACS, GTC, LAMMPS MILC, NAMD, QE, SPECFEM3D, TensorFlow, VASP Benchmark Application: Amber [PME-Cellulose_NVE], Chroma [szscl21_24_128], GROMACS [ADH Dodec: Dev Prototype], GTC [moi#pr
Slide 19: Supercomputing Is Hard — GPU Cloud 1/7th Cost of CPU Cloud
SUPER COMPUTING IS HARD — GPU CLOUD 1/7TH COST OF CPU CLOUD SUPERCOMPUTING CLOUD 48x Faster, 1/7th the Cost CPU Instance 48 Hours, $152 1x GPU Instance Amber, Chroma, GROMACS, GTC, LAMMPS MILC, NAMD, QE, SPECFEM3D, TensorFlow, VASP 8x GPU Instance Benchmark Application: Amber [PME-Cellulose_NVE], Ch
Slide 20: The Largest Cloud Simulation in History
THE LARGEST CLOUD SIMULATION IN HISTORY Events Processed 50K NVIDIA GPUs IN THE CLOUD Per GPU Type 350 PF OF SIMULATION FOR 2 HOURS K80 PRODUCED 5% OF ANNUAL SIMULATION DATA M60 AWS, MICROSOFT AZURE, GOOGLE CLOUD PLATFORM V100 DISTRIBUTED ACROSS U.S., EUROPE, APAC T4 P40 Frank Wüerthwein, Ph.D. P100
Slide 21: Announcing World's Largest On-Demand Supercomputer
ANNOUNCING WORLD’S LARGEST ON-DEMAND SUPERCOMPUTER Up to 800 V100 GPUs Connected via Mellanox InfiniBand 21
Slide 22: Diverse Arm Architectures
DIVERSE ARM ARCHITECTURES AMPERE COMPUTING eMAG AMAZON GRAVITON Hyperscale and Storage Hyperscale and SmartNIC MARVELL THUNDERX2 FUJITSU A64FX HUAWEI KUNPENG 920 Hyperscale, Storage and HPC Supercomputing Big Data and Edge 22
Slide 23: NVIDIA CUDA on Arm at Oak Ridge National Lab
NVIDIA CUDA ON ARM AT OAK RIDGE NATIONAL LAB Benchmark Application [Dataset]: GROMACS [ADH Dodec- Dev prototype], LAMMPS [LJ 2.5], MILC [Apex Small], NAMD [apoa1_npt_cuda], Quantum Espresso [AUSURF112-jR], Relion [Plasmodium Ribosome], SPECFEM3D [four_material_simple_model], TensorFlow [ResNet50: Ba
Slide 24: Announcing NVIDIA HPC for Arm
ANNOUNCING NVIDIA HPC FOR ARM GPU GPU GPU GPU NIC PCIe Switch PCIe Switch NIC CPU CPU HPC Server Reference Platform | 8 V100 Tensor Core GPUs with NVLink 4 100 Gbps Mellanox InfiniBand| Systems Ranging from Supercomputer, Hyperscale, to Edge CUDA on Arm Beta Available Now 24
Slide 25: Announcing NVIDIA HPC for Arm
ANNOUNCING NVIDIA HPC FOR ARM GPU GPU GPU GPU NIC PCIe Switch PCIe Switch NIC CPU CPU HPC Server Reference Platform | 8 V100 Tensor Core GPUs with NVLink 4 100 Gbps Mellanox InfiniBand| Systems Ranging from Supercomputer, Hyperscale, to Edge CUDA on Arm Beta Available Now 25
Slide 26: Announcing NVIDIA HPC for Arm
ANNOUNCING NVIDIA HPC FOR ARM APPLICATIONS COMET GAMERA INDEX LSMS MILC OPTIX QUANTUM ESPRESSO TENSORFLOW DCA++ GROMACS LAMMPS MATLAB NAMD PARAVIEW RELION VMD PROGRAMMING MODELS SDKS C++ OPENACC ARM ALLINEA STUDIO CUDA-GDB LLVM PERFORCE TOTALVIEW SINGULARITY CUDA PYTHON BRIGHT COMPUTING CUPTI NVCC P
Slide 27: 27
27
Slide 28: Extreme Compute Needs Extreme IO
EXTREME COMPUTE NEEDS EXTREME IO System Memory System Memory CPU CPU 50 GB/s 50 GB/s PCIe Switch PCIe Switch NIC NIC GPU GPU NODE A NODE B TRADITIONAL RDMA 28
Slide 29: Extreme Compute Needs Extreme IO
EXTREME COMPUTE NEEDS EXTREME IO System Memory System Memory CPU CPU PCIe Switch 100 GB/s PCIe Switch NIC NIC GPU GPU NODE A NODE B GPUDIRECT RDMA 29
Slide 30: Extreme Compute Needs Extreme IO
EXTREME COMPUTE NEEDS EXTREME IO System Memory System Memory System Memory CPU CPU CPU 50 GB/s PCIe Switch 100 GB/s PCIe Switch PCIe Switch NIC NIC NIC GPU GPU GPU NODE A NODE B GPUDIRECT RDMA TRADITIONAL STORAGE 30
Slide 31: Extreme Compute Needs Extreme IO
EXTREME COMPUTE NEEDS EXTREME IO System Memory System Memory System Memory CPU CPU CPU Storage PCIe Switch 100 GB/s PCIe Switch PCIe Switch 100 GB/s NIC NIC NIC GPU GPU GPU NODE A NODE B GPUDIRECT RDMA GPUDIRECT STORAGE 31
Slide 32: Announcing NVIDIA Magnum IO
ANNOUNCING NVIDIA MAGNUM IO Acceleration Libraries for Large-scale HPC and IO High-bandwidth, Low-latency, Massive Storage Access with Lower CPU Utilization System Memory System Memory System Memory CPU CPU CPU Storage PCIe Switch 100 GB/s PCIe Switch PCIe Switch 100 GB/s NIC NIC NIC GPU GPU GPU NOD
Slide 33: NVIDIA RAPIDS Data Science
NVIDIA RAPIDS DATA SCIENCE PYTHON RAPIDS DEEP LEARNING FRAMEWORKS DASK PANDAS SCI-KL / XGBOOST CUDF CUML CUGRAPH CUDNN CUDA APACHE ARROW Open Source | Multi-GPU and Multi-Node | Up to 100x Speed-Up | 150K Downloads in 1 Year Data Load and Processing Times from Hours to Minutes | Used by NERSC, ORNL,
Slide 34: NVIDIA Magnum IO Boosts RAPIDS Data Analytics
NVIDIA MAGNUM IO BOOSTS RAPIDS DATA ANALYTICS Q4 TPC-H Benchmark Work Breakdown: With Repeated Query Repeat Clean Query Up WITHOUT GDS CUDA GPU and CPU Data Warmup Driver Startup Allocation Preload Query Close WITH GDS 0 400,000 800,000 1,200,000 Latency (msec) NEW PANGEO XARRAY ZARR READER 20x ON T
Slide 35: Announcing World's Largest Interactive Volume Visualization
ANNOUNCING WORLD’S LARGEST INTERACTIVE VOLUME VISUALIZATION Simulating Mars Lander with FUN3D | Interactively Visualizing 150 TB; Unstructured Mesh 4 NVIDIA DGX-2 Streaming 400 GB/s | NVIDIA Magnum IO | NVIDIA IndeX 35
Slide 36: Announcing NVIDIA DGX-2 as Supercomputing Analytics Instrument
ANNOUNCING NVIDIA DGX-2 AS SUPERCOMPUTING ANALYTICS INSTRUMENT Powered by NVIDIA Magnum IO TPC-H RECORD 10 TB JOIN EXTREME WEATHER NVIDIA RAPIDS AI INFERENCE NVIDIA TENSOR RT VMD COMPUTATIONAL 3D VOLUME ANALYTICS MICROSCOPE PANGEO XARRAY NVIDIA OPTIX 3D INTERACTIVE VOLUME RENDERING NVIDIA INDEX 16 V
Slide 37: The Expanding Universe of HPC
THE EXPANDING UNIVERSE OF HPC NVIDIA HPC for ARM Azure SIMULATION NVIDIA EGX EDGE ANALYTICS Edge NETWORK Supercomputing Platform NGC EXTREME IO NVIDIA DGX-2 NVIDIA Supercomputing Magnum IO Analytics Instrument 37
Slide 38: NVIDIA
©& NVIDIA