Possibilities of generative models
1 like•514 views
This document provides specifications for the NVIDIA Titan RTX graphics card, including: - CUDA Cores: 4608 - RT Cores: 72 - Tensor Cores: 576 - Memory Size: 24 GB or 48 GB GDDR6 - Memory Bandwidth: Up to 672 GB/s - Display Support: 3x DP, 1x HDMI, 1x VirtualLink - Power: 280W, 2x 8-pin PCIe connectors
![17
IMAGE BASED DL IS EASY
Object detection Semantic Segmentation
Figures copyright Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun,
2015. [Faster R-CNN]
Figures copyright Preferred Networks Inc., 2016.](https://image.slidesharecdn.com/nvbmva-191127145115/85/Possibilities-of-generative-models-17-320.jpg)
Possibilities of generative models
- 1. ALISON B LOWNDES AI DevRel | EMEA @alisonblowndes November 2019
- 3. 3
- 4. 4 GPU Architecture Turing CUDA Cores 4608 RT Cores 72 Tensor Cores 576 Memory Size 24 GB GDDR6 48 GB GDDR6 with NVLINK Memory BW Up to 672 GB/s NVLink 2-way, 100 GB/s Display Support 3x DP + 1x HDMI + 1x VirtualLink Board Power (TDP) 280W Power Connectors 2x 8-pin PCle TITAN RTX SPECIFICATIONS
- 5. 5 SELECTING THE RIGHT GPU SOLUTION
- 6. 6 NVIDIA CUDA-X AI ECOSYSTEM FRAMEWORKS CLOUD DEPLOYMENT Workstation CloudServer DA GRAPH DLTRAINML DLINFERENCE Amazon SageMaker Serving Amazon SageMaker Neo Google Cloud ML CUDA-X AI CUDA AzureMachineLearning
- 7. 7
- 8. 8 NVIDIA TOPS MLPERF EDGE SOC BENCHMARKS 0.0x 0.5x 1.0x 1.5x MobileNet-v1 ResNet-50 v. 1.5 SSD MobileNet-v1 SSD ResNet-34 Qualcomm SDM855 Intel i3-1005G1 NVIDIA Xavier MLPerf v0.5 Inference Closed; Retrieved from www.mlperf.org 6 November 2019. Single stream performance derived from reported MLPerf latencies. GNMT omitted due to no submissions among edge and mobile form factor SOCs in v0.5. MLPerf name and logo are trademarks. See www.mlperf.org for more information. Per-AcceleratorPerformance SINGLE-STREAM SCENARIO X = No result submittedX Best Inference Performance Among Commercially Available Edge And Mobile SoCs X X 0 0.5 1 1.5 MobileNet-v1 ResNet-50 v. 1.5 SSD MobileNet-v1 SSD ResNet-34 Qualcomm SDM855 Intel i3-1005G1 NVIDIA Xavier Per-AcceleratorPerformance MULTI-STREAM SCENARIO X X X X X X X X https://github.com/mlperf/inference_results_v0.5/tree/master/open/NVIDIA www.mlperf.org
- 10. 10
- 12. 12
- 13. 13 CuLe: CUDA Learning Env: https://github.com/NVlabs/cule GPU Accelerated Atari Emulation for RL
- 14. NETWORK COMPLEXITY IS EXPLODING
- 15. NVIDIA RESEARCH CNN Image Inpainting Progressive GANNoise-to-Noise Denoising RTX NVSwitch CuDNN
- 17. 17 IMAGE BASED DL IS EASY Object detection Semantic Segmentation Figures copyright Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun, 2015. [Faster R-CNN] Figures copyright Preferred Networks Inc., 2016.
- 18. 18 Numerous applications 3D DL IS EXCITING Simulation Medical imaging Autonomous driving Manipulation Robotics Augmentedreality * This slide is best viewed in "slide show" mode.
- 19. 19 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
- 20. AI SEES IN TEXTURES, NOT SHAPES https://arxiv.org/pdf/1811.12231.pdf University of Tübingen & University of Edinburgh
- 21. 21 What’s really going on? My Python Program DNN Framework CPUs GPUs System Memory Network Drives PCI Express
- 22. 22 It’s Like tuning an orchestra GPU CPU GPU CPU System Memory SSD A SSD B Network PCI Express
- 23. 23 NVIDIA Nsight Systems • Balance your workload across multiple CPUs and GPUs • Locate idle CPU and GPU time • Locate redundant synchronizations • Locate optimization opportunities • Improve application’s performance System Wide Profiling Tool developer.nvidia.com/tools-overview
- 25. 25 A DAY IN THE LIFE OF A WEIGHT Size of Error https://arxiv.org/pdf/1708.03888.pdf
- 28. 28 EXAMPLE: AUTOENCODERS UNSUPERVISED feature learning Sparse Representation Training Data Reconstruction Encoder Decoder Minimize Reconstruction Error - InputLayer HiddenLayer HiddenLayer HiddenLayer BottleneckLayer HiddenLayer HiddenLayer HiddenLayer OutputLayer
- 30. 30 STACKED CAPSULE AUTOENCODERS Toronto, Google, ORI: https://arxiv.org/pdf/1906.06818.pdf
- 34. www.Robust.ai Gray Marcus, Rodney Brooks, Steven Pinker et al
- 35. 35 POET
- 36. 36
- 37. REINFORCEMENT LEARNING “from our own mistakes”
- 39. Website: https://web.stanford.edu/~yukez/ https://ai2thor.allenai.org Paper: https://arxiv.org/pdf/1807.03480.pdf Stanford: adversarial learning
- 40. 40 Proxyless Neural Architecture Search https://arxiv.org/pdf/1812.00332.pdf https://github.com/MIT-HAN-LAB/ProxylessNAS
- 41. 41
- 42. 42NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE. ISAAC Isaac Robot Engine – Modular robot framework | Isaac Sim - Virtual roboticslaboratory Isaac Gym – Reinforcement learning simulator | Isaac Robot Apps – Kaya, Carter and Link Available at developer.nvidia.com/isaac-sdk CARTER (Xavier)KAYA (Nano) LINK (Multi Xavier) JETSONNANO ISAAC OPEN TOOLBOX Sensor and Actuator Drivers Core Libraries GEMS Reference DNN Tools CUDA-X Isaac Robot Engine JETSONTX2 JETSONAGX XAVIER Isaac Sim Isaac Gym
- 44. 44 Simulating Reality Design Robotics Autonomous Vehicles Media and Entertainment
- 46. 46 SAFE AV VALIDATION – THE CHALLENGES Highly Complex System Large Computers, DNNs, Sensors Real-Life Scenario Coverage Account for Rare & Unpredictable Cases Continuous Reaction Loop Vehicle & World are Dependent
- 48. 48
- 49. 49 COMPLEX RENDERING PIPELINE CONCEPT | MODELING | TEXTURE | RIGGING | ANIMATION | LIGHTING | RENDER
- 50. 50
- 51. 51 FIRE VIDEO
- 52. 52
- 53. 53 REVOLUTIONIZE GRAPHICS WITH DEEP LEARNING
- 54. 54 CONDITIONAL GANS Generates output consistent with the training set Generator (Regressor) Discriminator (Classifier) Generated Target • If the output is under-constrained your output will look fuzzy. • It won’t capture the true peaks and valleys • We can overcome this problem using a conditional GAN • Generates output matching the training set Loss fcn
- 55. 55 The High Altitude Water Cherenkov (HAWC) Observatory A cosmogenic gamma ray observatory, examining some of the most energetic light in the universe Located on Pico de Orizaba, Mexico High duty cycle, high statics, high energy physics experiment Daniel Ho, Gefen Kohavi, Michael Gussert
- 56. 56 PixelCNN Images sampled from a PixelCNN model trained on PMT charge data show realistic features. - smooth gaussian distribution for dense events - good distribution of event sparsity - varying angles / direction of hits Generated Images
- 58. 58 Detection Planning Acceleration Assimilation Enhancement Parametrization AugmentationPrediction Use Inpainting to Repair Damaged GOES-17 Observations
- 60. 60 A new vocoder for speech synthesis built on a flow based generative model Fast, completely parallel inference procedure 150X real-time on one GPU mel-spectrogram audio samples WaveGlowTacoTron Hello, world! Waveglow synthesized speech http://nv-adlr.github.io/WaveGlow
- 61. SOTA NLP Techniques Transformer: A confluence of all three SOTA NLP 61 ● Encoder + Decoder Structure ● Attention mechanism ● Self-Attention within each encoder & decoder No more recurrent structure! Let’s witness the power of GPU parallelization!
- 62. 62 MEGATRON (August 2019) https://devblogs.nvidia.com/training- bert-with-gpus/ https://github.com/nvidia/megatron-lm BERT-Large trained in 53 minutes on 92x DGX-2H systems with model parallelism 2 millisecond inference with TRT/T4
- 64. 64 BIOBERT https://github.com/naver/biobert- pretrained https://github.com/dmis- lab/biobert Korea University, Naver Corp BioBERT (Bidirectional Encoder Representations from Transformers for Biomedical Text Mining) https://arxiv.org/pdf/1901.08746.pdf “NLP’s ImageNet moment” Reddit, Jan 2019
- 65. 65 MONTE CARLO SIMULATION repeated random sampling to solve problems that might be deterministic in principle @python4finance
- 66. 66 Algorithmic Trading using Deep Autoencoder based Statistical Arbitrage NVIDIA Deep Learning Institute
- 67. National Academy of Sciences, Siyu He et al, Flatiron Institute https://arxiv.org/pdf/1811.06533.pdf https://news.developer.nvidia.com/researchers-develop-the-first-deep-learning-based-3d-simulation-of-the-universe/
- 68. AI Playground: GANimal • https://www.nvidia.com/ en-us/research/ai- playground/ • http://nvidia-research- mingyuliu.com/ganimal/i ndex.html
- 70. 70 ONE ARCHITECTURE FOR DATA SCIENTISTS Simulation Data Analytics Visualization
- 71. 71 GET STARTED WITH NGC Deploy containers: ngc.nvidia.com Learn more about NGC offering: nvidia.com/ngc Technical information: developer.nvidia.com Explore the NGC Registry for DL, ML & HPC
- 72. 72 Pretrained Models All models are trained on google openimages public dataset Available to download on ngc.nvidia.com
- 74. RAPIDS RAPIDS GPU Accelerated End-to-End Data Science RAPIDS is a set of open source libraries for GPU accelerating data preparation and machine learning. OSS website: rapids.ai GPU Memory Data Preparation VisualizationModel Training cuGraph Graph Analytics cuML Machine Learning cuDF Data Preparation
- 75. 75 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++APIsto 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)
- 76. 76 NVIDIA TensorRT From Every Framework, Optimized For Each Target Platform TESLA V100 DRIVE AGX TESLA T4 JETSON Xavier NX NVIDIA DLA TensorRT
- 77. TF-TRT = TF + TRT https://docs.nvidia.com/deeplearning/dgx/tf-trt-user-guide/index.html
- 78. 78NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE. NVIDIA TRANSFER LEARNING TOOLKIT FEATURES Model pruning reduces size of the model resulting in faster inference Faster Inference with Model Pruning GPU-accelerated high performance models trained on large scale datasets. Efficient Pre-trained Models Re-training models, adding custom data for multi GPU training using an easy to use tool Training with Multiple GPUs Packaged in a container easily accessible from NVIDIA GPU Cloud website. All code dependencies are managed automatically Containerization Abstraction from having deep knowledge of frameworks, simple intuitive interface to the features Abstraction Models exported using TLT are easily consumable for inference with Deep Stream SDK Integration
- 79. 7979 AUTOMATIC MIXED PRECISION Insert ~two lines of code to use Automatic Mixed-Precision and get up to a 3X speedup Support for TensorFlow, PyTorch and MXNet Easy to Use and Great Performance Automatic mixed precision applies two techniques to maximize performance while preserving accuracy: 1) Optimizing per operation precision by casting to FP16 or FP32 2) Dynamic loss scaling to properly handle gradient accumulation NEW
- 80. AUTOMATIC MIXED PRECISION ● Speed-up: 1.5x - 3x ● Memory footprint reduction: increase batch size up to 2x, more capability ● No accuracy drop Speedup Your Network Across Frameworks With Just Two Lines of Code Tensor Cores NVIDIA AMP Frameworks Models CNN, RNN, GAN, RL, NCF…
- 81. 81 CUTLASS 2.0 https://github.com/NVIDIA/cutlass optimal HMMA, IMMA, and BMMA kernels compile with clang supporting CUDA Toolkits 9.2 + kernels targeting all WMMA configurations documentation and SDK examples
- 82. 82NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE. ANNOUNCING MAGNUM IO NVIDIA's Multi-GPU, Multi-Node Networking and Storage IO Optimization Stack CUDA CUDA-X Desktop Development Data Center Solutions GPU-Accelerated Cloud Supercomputers Magnum IO Transport Protocol | System Interconnect | Network Topology | Storage Simulation ML/DL Data Analytics Visualization
- 83. 83 NVIDIA DGX SUPERPOD Mellanox EDR 100G InfiniBand Network Mellanox Smart Director Switches In-Network Computing Acceleration Engines Fast and Efficient Storage Access with RDMA Up to 130Tb/s Switching Capacity per Switch Ultra-Low Latency of 300ns Integrated Network Manager Terabit-Speed InfiniBand Networking per Node … Rack 1 Rack 16 Compute Backplane Switch Storage Backplane Switch 64 DGX-2 GPFS 200 Gb/s per node 800 Gb/s per node White paper: https://nvidia.highspot.com/items/5d073ad681171721086b2788
- 84. 84 JETSON XAVIER NX Up to 21 DL TOPS AI Performance 10W | 15W 384 CUDA Cores | 48 Tensor Cores 6 core CPU | 8 GB Memory 45x70mm $399 Xavier Performance. Nano Size. Get started today: - Jetson AGX Xavier Developer Kit + software patch - Documentation on Jetson Download Center - SOM available Q1 2020
- 85. 85 THE JETSON FAMILY for AI at the Edge and Autonomous System designs Same software Listed prices are for 1000u+ | Full specs at developer.nvidia.com/jetson * TX2i: 10-20W 7.5 – 15W* 50mm x 87mm JETSON TX2 series 1.3 TFLOPS (FP16) 5 - 10W 45mm x 70mm JETSON NANO 0.5 TFLOPS (FP16) 10 – 30W 100mm x 87mm JETSON AGX XAVIER series 11 TFLOPS (FP16) 32 TOPS (INT8) 10 - 15W 45mm x 70mm JETSON Xavier NX 6 TFLOPS (FP16) 21 TOPS (INT8) AI at the edge Fully autonomous machines
- 86. 86 AI FRAMEWORKS TARGET VERTICAL INDUSTRIES Smart Cities, Retail, Manufacturing Software-Defined 5G Telco Networks Conversational AI Placeholder image AERIALMETROPOLIS JARVIS
- 87. Fundamentals Accelerated Computing Game Development & Digital Content Finance NVIDIA DEEP LEARNING INSTITUTE Online self-paced labs and instructor-led workshops on deep learning and accelerated computing Take self-paced labs at www.nvidia.co.uk/dlilabs View upcoming workshops and request a workshop onsite at www.nvidia.co.uk/dli Educators can join the University Ambassador Program to teach DLI courses on campus and access resources. Learn more at www.nvidia.com/dli Intelligent Video Analytics Healthcare Robotics Autonomous Vehicles Virtual Reality
- 88. 88 NVIDIA INCEPTION PROGRAM Accelerates AI startups with a boost of GPU tools, tech and deep learning expertise Startup Qualifications Driving advances in the field of AI Business plan Incorporated Web presence Technology DL startup kit* Pascal Titan X Deep Learning Institute (DLI) credit Connect with a DL tech expert DGX-1 ISV discount* Software release notification Live webinar and office hours *By application Marketing Inclusion in NVIDIAmarketing efforts GPU Technology Conference(GTC) discount Emerging Company Summit (ECS) participation+ Marketing kit One-page story template eBook template Inception web badge and banners Social promotion request form Event opportunities list Promotion at industry events GPU ventures+ +By invitation www.nvidia.com/inception
- 90. 90 CONNECT Connect with hundreds of experts from top industry, academic, startup, and government organizations LEARN Gain insight and valuable hands-on training through over 500+ sessions DISCOVER See how GPU technology is creating breakthroughs in deep learning, cybersecurity, data science, healthcare and more INNOVATE Explore disruptive innovations that can transform your work JOIN US AT GTC 2020 | USE VIP CODE NVALOWNDES FOR 25% OFF March 22—26, 2020 | Silicon Valley Don’t miss the premier AI conference. www.nvidia.com/gtc
- 91. 91 March 22 | Full-Day Workshops March 23 - 26 | Conference & Training Get the hands-on experience you need to transform the future of AI, high-performance computingand more with NVIDIA’sDeep Learning Institute (DLI). Register for GTC 2020 to earn certification in full-day workshops, join instructor-led sessions, and start self- paced training. www.nvidia.com/en-us/gtc/sessions/training/ THE LATEST DEEP LEARNING DEVELOPER TOOLS