EPSRC CDT Conference
0 likes311 views
The document outlines advancements in AI and GPU technology led by NVIDIA, including the introduction of the A100 GPUs and multi-instance GPU capabilities aimed at enhancing performance and efficiency in AI applications. It highlights the deployment of AI supercomputers worldwide to support national AI initiatives and various sectors like healthcare and energy management. Additionally, it discusses NVIDIA's AI ecosystem tools and resources for developers and researchers, emphasizing climate adaptation and resilience efforts through AI-driven simulations and applications.
EPSRC CDT Conference
- 1. ALISON B LOWNDES Senior Scientist, Global AI @alisonblowndes August 2021
- 3. 3 NVIDIA A100 80GB Supercharging The World’s Highest Performing AI Supercomputing GPU 80GB HBM2e For largest datasets and models 2TB/s + World’s highest memory bandwidth to feed the world’s fastest GPU Multi-Instance GPU 3rd Gen NVLink 3rd Gen Tensor Core https://devblogs.nvidia.com/nvidia-ampere-architecture-in-depth/
- 4. 4 NEW MULTI-INSTANCE GPU (MIG) Optimize GPU Utilization, Expand Access to More Users with Guaranteed Quality of Service nvidia.com/en-us/technologies/multi-instance-gpu/ Up To 7 GPU Instances In a Single A100: Dedicated SM, Memory, L2 cache, Bandwidth for hardware QoS & isolation Simultaneous Workload Execution With Guaranteed Quality Of Service: All MIG instances run in parallel with predictable throughput & latency Right Sized GPU Allocation: Different sized MIG instances based on target workloads Flexibility to run any type of workload on a MIG instance Diverse Deployment Environments: Supported with Bare metal, Docker, Kubernetes, Virtualized Env. Amber GPU Mem GPU GPU Mem GPU GPU Mem GPU GPU Mem GPU GPU Mem GPU GPU Mem GPU GPU Mem GPU https://blogs.nvidia.com/blog/2020/05/14/multi-instance-gpus/
- 5. 5 NVIDIA SELENE Now Featuring NVIDIA DGX A100 640GB #1 in Green 500 | #5 Top500 | #1 MLPerf | #1 Industrial System 4,480 A100 GPUs 560 DGX A100 640GB systems 850 Mellanox 200G HDR switches 14PB of high-performance storage 2.8 EFLOPS of AI peak performance 63 PFLOPS HPL @ 24GF/W 1 DGX A100 replaces 150 CPU servers and saves 300 tons of CO2 per DGX per year!
- 6. 6 NVIDIA DGX SUPERPOD READY MADE NATIONAL AI INFRASTRUCTURE Thailand CMKL Univ installs DGX A100 Pod to advance National AI Program sponsored by MEHESI Luxembourg EuroHPC peta- scale AI cluster, MeluXina, built with ATOS + NVIDIA Vietnam VinAI deploys DGX A100 Pod to support AI researchers and engineers UAE Ministry of AI deploys DGX Pod with 30 petaflops to advance national AI programs Egypt 1st national AI super- computer as AI testbed built with Dell and NVIDIA Slovenia EuroHPCpeta-scale AI supercomputer,Vega, with ATOS and NVIDIA Sweden WASP installs #1 AI supercomputer at Linköping Univ with ATOS and NVIDIA India C-DAC installs largest national AI supercomputer with ATOX and NVIDIA Italy #1 AI supercomputer, LEONARDO,to launch with 14,000 NVIDIA Ampere GPUs UK #1 AI supercomputerin Cambridge built by NVIDIA for life science R&D US fastestAI supercomputer for academiaat University of Florida built with NVIDIA Canada Shared Services installs DGX Pod to support AI adoption across Canadian gov’t agencies Czech Republic’s largest AI supercomputer by NVIDIA& HPE at TechnicalUniv Ostrava
- 7. 7 First and only workstation with 4-way NVIDIA A100 GPUs, NVLink, and MIG Four A100 Tensor Core GPUs, 320 GB total HBM2E Multi-Instance GPU (MIG) for up to 28 GPU instances in a single DGX Station A100 3rd generation NVLink 200 GB/s bi-directional bandwidth between any GPU pair, almost 3x compared to PCIe Gen4 New maintenance-free refrigerant cooling system DGX STATION A100 320G Workgroup Appliance for the Age of AI CPU and Memory 64-core AMD® EPYC® CPU, PCIe Gen4 512 GB system memory Internal Storage 1.92 TB NVME M.2 SSD for OS 7.68TB NVME U.2 SSD for data cache Connectivity 2x 10GbE (RJ45) 4x Mini DisplayPort for display out Remote management 1GbE LAN port (RJ45)
- 8. 8 NVIDIA CUDA-X AI ECOSYSTEM
- 9. 9
- 11. 11
- 12. 12
- 13. Lunar surface imagery enhancement 13 Fig. 1: NASA’s Lunar Reconnaissance Orbiter Fig. 2: Apollo XVI landing site
- 14. Team Photo from Bootcamp Introduce your team and faculty - connect with your audience. TIME: No more than 30 seconds Researchers Team Advisors: Paula Harder Jose I. Delgado- Centeno Team leads Ben Moseley Valentin Bickel Siddha Ganju Miguel A. Olivarez- Mendez Freddie Kalaitzis Muhammed Razzak Yarin Gal Chedy Raissi
- 15. 15
- 16. 16 fdlausnz.org
- 17. Next Level of AI GPGPU in Space Applications Aitech’s S-A1760 Venus™: most powerful and smallest space AI GPGPU in small form factor (SFF). Suitable for the next gen of short duration spaceflight, NEO and LEO. 17
- 18. 18
- 19. 19 19 Main Computational Challenges Edge computing for smart grid management: • Better management of data for better adequacy of supply and demand => this will accelerate integration of renewables therefore limiting need for storage • Finding the shortest path in a grid is computationally expensive for a neural network • Real-time monitoring and adjustments of electricity flows • Massive usage of IoT devices and sensors that gather and transmit data back and forth in the system ❖ AI-based approaches for power grid stability ❖ Solutions like digital twins for logistics ❖ Smart metering ❖ Exploring Artificial Neural Networks (ANN) algorithms for network planning, electric load forecasting ❖ Supply and demand forecasting with weather forecasting as input ❖ Help developing supply chain digital twins as detailed simulation models which use real-time data and snapshots to forecast supply chain dynamics Smart Grids Assessment using intelligent applications at the EDGE
- 20. 20 SIMNET CLIMATE Rapid design optimization for alternative-energy solutions https://windinspire.jhu.edu/wp-content/uploads/2016/12/Large-Eddy-simulation.jpg
- 21. 21 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE. SIMNET: AI BASED SIMULATION http://developer.nvidia.com/simnet AI based Techniques Traditional Simulations Physical Prototyping Past Present Future Traditional numerical solvers work on one problem at a time making design process time consuming, do not address real-time simulations, data assimilation, inverse problems Data driven NN require data, are oblivious to physics laws, suffer from interpolation/extrapolation errors and are not generalizable Physical Prototyping is iterative, time consuming, costly and not optimized for material and characteristics
- 22. 22 SIMNET 22 Product Architecture Visualization Geometry & Point Clouds PINNs based Solver Framework HW DGX POD DGX GPU CSV Tensor Board VTK Paraview Boundary Conditions Monitor Inference Data Validation Data Training Domain Monitor Domain Inference Domain Validation Domain
- 23. 23 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE. AI ENABLING NEXT GENERATION SIMULATION Inverse & Data Assimilation Problems Improved Physics & Predictions
- 24. 24 24 EXPANDING NGC NEW CONTAINERS FOR A100 & ARM Now NGC-READY SYSTEMS FOR A100 Starting Q3 NGC Private Registry NGC Container Environment Modules Higher HPC app performance w/ NVTAGS NEW FEATURES Now Multi-arch support for x86, Arm and Power Learn More – ngc.nvidia.com | NGC Private Registry | NVTAGS | NGC Container Environment Modules HPC Simulation & Visualization AI Frameworks (A100) Chroma AutoDock 4 VMD ** * Available week of June 22 ** Available starting with v20.06 * * *
- 26. 26 EARTH-SYSTEM MODELS Long-wave and short-wave radiation Cloud macro and micro-physics Deep and shallow convection Planetary boundary layer Turbulent mountain stress Gravity wave drag Surface fluxes Aerosols Chemistry Simulating the Earth LARGE SCALE DYNAMICS SMALL SCALE PHYSICS Rotational Fluid Dynamics, Confined to Sphere Compressible Atmosphere Chaotic Internal Variability
- 27. 27 OMNIVERSE EARTH Digital Twins Of The Earth for Climate Adaptation And Resilience ESA EUMESTAT ECMWF NVIDIA
- 28. 28
- 29. 29
- 30. 30 KAOLIN - A Pytorch library for 3D DL - Supports a wide range of 3D data representations - Convenient dataloading/preprocessing/conversions - Large collection of 3D neural nets to choose from - Optimized implementations - Omniverse-Kit integration for easy rendering, interactive visualization, and much more. https://gitlab- master.nvidia.com/Toronto_ DL_Lab/kaolin
- 31. 31 GANVERSE
- 32. DiSECt BEST STUDENT PAPER, RSS 2021
- 35. 35 NVIDIA TOOLS EXTENSION LIBRARY (NVTX) ● Nsight 2021.2 release ● NVTX is a platform agnostic, tools agnostic API ● Allows developers to annotate(mark) source code, events, code ranges etc ● NVIDIA optimized Tensorflow, PyTorch, MXnet have NVTX annotations built in Import Library Insert Python Annotations Use Any NVIDIA Profiling Tool *Nsight Systems, Nsight Compute and Deep Learning Profiler make use of NVTX markers https://docs.nvidia.com/cuda/profiler-users-guide/index.html#nvtx
- 36. 36 Processes and threads CUDA and OpenGL API trace Multi-GPU Kernel and memory transfer activities cuDNN and cuBLAS trace Thread/core migration Thread state
- 37. 37 NVIDIA DATA LOADING LIBRARY (DALI) Fast Data Processing Library for Accelerating Deep Learning DALI in DL Training Workflow Currently supports: • ResNet50 (Image Classification), SSD (Object Detection)n • Input Formats – JPEG, LMDB, RecordIO, TFRecord, COCO, H.264, HVEC • Python/C++ APIs to define, build & run an input pipeline Full input pipeline acceleration including data loading and augmentation Drop-in integration with direct plugins to DL frameworks and open source bindings Portable workflows through multiple input formats and configurable graphs Flexible through configurable graphs and custom operators Over 1000 GitHub stars | Top 50 ML Projects (out of 22,000 in 2018)
- 38. 38 DALI RESOURCES Official Documentation (Quick Start, Developer Guides) https://docs.nvidia.com/deeplearning/sdk/index.html#data-loading GitHub Documentation https://docs.nvidia.com/deeplearning/sdk/dali-developer-guide/docs/index.html DALI Samples & Tutorial https://docs.nvidia.com/deeplearning/sdk/dali-developer-guide/docs/examples/index.html DALI Blog https://devblogs.nvidia.com/fast-ai-data-preprocessing-with-nvidia-dali/
- 39. 39 TENSORS: MULTI-DIMENSIONAL STRUCTURE Spatio-temporal data, (f)MRI, Deep Net Features, LIDAR data, …
- 40. 40 4D CONVNET OVER SPACE AND TIME 3D space + time as a single entity (Minkowski space) 40 Slides by Chris Choy, NVIDIA
- 41. 41 SPATIAL SPARSITY IN 3D 3D PERCEPTION WITH SPARSE TENSORS BY CHRISTOPHER CHOY Dai et al., ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes, CVPR’17
- 42. 42 MINKOWSKI ENGINE Discriminative Networks Benjamin Graham, Sparse 3D convolutional neural networks, BMVC’15 Chris Choy, JunYoung Gwak, Silvio Savarese, 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks, CVPR’19 github.com/NVIDIA/MinkowskiEngine
- 43. 43 Jean Kossaifi Anima Anandkumar Maja Pantic Yannis Panagakis Jeremy Cohen Julia Gusak Meraj Hashemizadeh Aaron Meurer Yngve Mardal Moe Taylor Lee Patti Marie Roald Caglayan Tuna … Aaron Meyer tensorly.org
- 44. 44 Materials/MDL Physics/VFX AI Path-Tracing USD NVIDIA CORE TECH UNIVERSAL ASSET EXCHANGE AND SHARED VIRTUAL WORLD COLLABORATORS PORTAL
- 45. 45 CUTTING EDGE APPLICATIONS Core Omniverse Apps FOR DESIGNERS, CREATORS, ENGINEERS FOR ROBOTICISTS, SIMULATION SPECIALISTS FOR GEFORCE RTX GAMERS FOR DESIGNERS, CREATORS, ENGINEERS FOR 3D DEEP LEARNING RESEARCHERS FOR GAME DEVELOPERS, ANIMATORS
- 46. 46
- 47. 47 OMNIVERSE MACHINIMA Advanced Simulation Technologies wrnch AI Pose Estimation Use a mobile camera to capture human body motions and automatically apply to 3D character mesh. Omniverse RTX Renderer 1440p @ 30 fps, NVIDIA MDL materials, fully dynamic lighting. Easily toggle between interactive path tracing mode and real time ray tracing mode. NVIDIA PhysX 5, Flow, and Blast Exclusive access to NVIDIA PhysX 5 advanced physics simulation tool kit, plus Flow and Blast for easily implemented realistic fire, smoke and explosions.
- 48. 48 OMNIVERSE AUDIO2FACE › Powered by NVIDIA AI › Instant automatic facial animation with realistic, believable motion › Switch between voices, genders, and languages › Use dialogue track, or singing AI-Powered Facial Animation from an Audio Track
- 49. Best in Show Award, Siggraph 2021
- 51. 51
- 52. 52 DEPLOY ON ANY NVIDIA RTX GPU From Laptop, to Data Center. On-premise or in the Cloud. NVIDIA Studio Any RTX Workstation or Laptop EGX Platform NVIDIA Certified Systems with RTX Professional Visualization Quadro RTX 4000 to NVIDIA RTX A6000
- 53. 53 SEE YOU IN OMNIVERSE DEVELOP ON OMNIVERSE DOCUMENTATION TUTORIALS AND WEBINARS DOWNLOAD OPEN BETA EXPLORE OMNIVERSE ENTERPRISE
- 54. BEST PAPER, ICRA 2021 REACTIVE HUMAN-TO-ROBOT HANDOVERS OF ARBITRARY OBJECTS https://sites.google.com/nvidia.com/handovers-of-arbitrary-objects.
- 56. 56 KAYA — A ROBOT FOR MAKERS Low-cost platform to get started with robotics Follow Me App Object Detection DNN NVIDIA Jetson Nano
- 57. MUCH MORE WITH ISAAC SOFTWARE GPU Accelerated Algorithms/DNNs (GEMs) And more… Free Space Segmentation 3D Object Pose Estimation Motion Planning Stereo Depth Stereo Visual Inertial Odometry Super Pixels April Tags 2D Skeleton Pose Estimation DeepStream Integration Planner with Costmaps Multi Lidar Support Navigation (LQR Path Planner) Sensors Robot Platforms Audio
- 58. (1) RL trains expert (2) Student mimics expert Training Environment policy gradient Exper t Weak Augment Exper t Supervise gradient Studen t Strong Augment SECANT SECANT: Self-Expert Cloning for Zero-Shot Generalization of Visual Policies. Fan, Wang, Huang, Yu, Fei-Fei, Zhu, Anandkumar. https://arxiv.org/abs/2106.09678
- 59. 59
- 60. GANcraft learns details that are much finer than a single block
- 62. 62 DRAFT – FOR PARTNER INTERNAL USE ONLY THE NEW ERA OF VISUAL COMPUTING Generative Design Analytics Ray Tracing AR, VR, XR Design Reviews Virtualization Performance Quality AI EVERYWHERE ADVANCED VISUALIZATION INTERACTIVE SIMULATION REMOTE COLLABORATION Image courtesy of Altair Engineering Image courtesy of KPF
- 63. 63 DRAFT – FOR PARTNER INTERNAL USE ONLY
- 64. Pyramid Vision Transformer (PVT) https://arxiv.org/abs/2105.15203 Wang et al., Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions, arXiv21
- 66. 66 NVIDIA Riva GPU-Accelerated SDK for Multimodal Conversational AI Applications Sign up: developer.nvidia.com/riva End-to-End Multimodal Conversational AI Skills Pre-trained SOTA models-100,000 Hours of DGX Retrain with Transfer Learning Toolkit Deploy Services with One Line of Code <300 ms latency | 1/3rd Cost on A100 versus CPU Audio/ text PRETRAINED MODELS TAO NVIDIA GPU CLOUD RETRAIN ASR Noisy Environments Accents & Jargon Spontaneous, Scripted, Phone NLU Question Answering, Contextual Understanding Sentiment TTS Voice Fonts, Emotional Control Inflection & Cadence Dialog Manager I Riva SKILLS TRANSFER LEARNING TOOLKIT Multi-speaker domain specific output INFERENCE Available in Riva 1.0 Beta
- 67. 67 Real-Time Transcription Virtual Assistant Highly accurate domain specific conversational AI bot Riva OPEN BETA USE CASES End-to-end voice enabled AI assistant Chatbot Generate highly accurate real-time transcriptions
- 68. 68 Riva RESOURCES 10 Pre-Trained Models + Notebooks Collection of pre-trained ASR, NLU, TTS models with notebooks to fine-tune with TLT and export to Riva 5 New Riva Developer Blogs Introduction to NVIDIA Riva Tutorials for building real-time apps with Riva, including: Question Answering | Virtual Assistant | Transfer Learning Transcription & Entity Recognition 4 New Sample Applications Ready-to-run sample apps for transcription and entity recognition, Virtual Assistant, Virtual Assistant with Rasa Integration & Speechsquad
- 69. 69 ORIN
- 70. 70 DEEP LEARNING INSTITUTE Training Labs Nanodegrees nvidia.com/DLI TWO DAYS TO A DEMO Create your first demo today developer.nvidia.com/ embedded/twodaystoademo JETSON DEVELOPER KIT AGX Xavier Developer Kit $699 Xavier NX software patch developer.nvidia.com/ buy-jetson GTC Largest event for GPU developers gputechconf.com JETSON - START NOW
- 71. 71
- 72. ENTERING THE AI HEALTHCARE ERA AI Papers in PubMed (Machine Learning or Deep Learning) 4.5x AI Investment Drugs, Cancer, Molecular, Drug Discovery* *Source: https://hai.stanford.edu/research/ai-index-2021 Healthcare Spend $8T | Growing & Aging Population | Chronic Disease | Public Health NLP IMAGING INSTRUMENTS CONVERSATIONAL AI DRUG DISCOVERY GENOMICS
- 73. NLP IMAGING AGX GUARDIAN DISCOVERY PARABRICKS NVIDIA CLARA Computational Platform for the AI Healthcare Era NVIDIA CLARA Private Cloud Edge Datacenter Embedded Device
- 74. 74 DLI UNIVERSITY TRAINING UNIVERSITY AMBASSADOR PROGRAM • Qualified faculty and researchers can get certified to teach DLI workshops to their students at no cost. • Hundreds of universities certified around the world, including: TEACHING KITS • Qualified university educators can download courseware across deep learning, accelerated computing, and robotics. • Kits include lecture materials, GPU cloud resources, access to self-paced DLI courses, and more. Learn more at www.nvidia.com/dli
- 75. 75 APPLIED RESEARCH ACCELERATOR PROGRAM Use case with deployed GPU-accelerated application Development of GPU- accelerated application Basic Research conducted by University Applied Research project(s) Program focus Supports research projects that have the potential to make a real-world impact through deployment into GPU-accelerated applications adopted by commercial and government organizations. Program Benefits Hardware and funding grants Technical guidance and support Grant application support Hands-on training with the NVIDIA Deep Learning Institute Networking and marketing opportunities Robotics and AI for Automation Apply online: https://www.nvidia.com/en-gb/industries/higher-education-research/applied-research-program/
- 76. 76