Downloaded 166 times
Uploaded byNVIDIA
Opening Keynote at GTC 2015: Leaps in Visual Computing
The document discusses advancements in visual computing presented by Jen-Hsun Huang at GTC 2015, emphasizing the growth of GPU computing, deep learning applications, and NVIDIA's new products including the Titan X GPU and the Drive PX system for self-driving cars. It highlights significant growth metrics, such as a 10x increase in CUDA downloads and the adoption of deep learning techniques in various fields like medical research and autonomous driving. Additionally, the document outlines NVIDIA's roadmap for the future of deep learning and GPU technology.
More Related Content
Similar to Opening Keynote at GTC 2015: Leaps in Visual Computing
Opening Keynote at GTC 2015: Leaps in Visual Computing
- 1. LEAPS IN VISUALCOMPUTING JEN-HSUN HUANG, CO-FOUNDER & CEO | GTC 2015
- 2. FOUR ANNOUNCEMENTS A NewGPU and Deep Learning A Very Fast Box and Deep Learning Roadmap Reveal and Deep Learning Self-Driving Cars and Deep Learning
- 3. AMAZING YEAR INVISUAL COMPUTING © 2015 Industrial Light & Magic. All Rights Reserved.
- 4. 10X GROWTH INGPU COMPUTING 2008 150,000 CUDA Downloads 4,000 Academic Papers 60 Universities Teaching 77 Supercomputing Teraflops 6,000 Tesla GPUs 27 CUDA Apps
- 5. 2008 150,000 CUDA Downloads 4,000 Academic Papers 60 UniversitiesTeaching 77 Supercomputing Teraflops 6,000 Tesla GPUs 27 CUDA Apps 2015 3 Million CUDA Downloads 10X GROWTH IN GPU COMPUTING
- 6. 2015 3 Million CUDA Downloads 10XGROWTH IN GPU COMPUTING 319 CUDA Apps 2008 150,000 CUDA Downloads 4,000 Academic Papers 60 Universities Teaching 77 Supercomputing Teraflops 6,000 Tesla GPUs 27 CUDA Apps
- 7. 2015 3 Million CUDA Downloads 800 UniversitiesTeaching 10X GROWTH IN GPU COMPUTING 319 CUDA Apps 2008 150,000 CUDA Downloads 4,000 Academic Papers 60 Universities Teaching 77 Supercomputing Teraflops 6,000 Tesla GPUs 27 CUDA Apps
- 8. 2015 3 Million CUDA Downloads 800 UniversitiesTeaching 10X GROWTH IN GPU COMPUTING 319 CUDA Apps 2008 150,000 CUDA Downloads 4,000 Academic Papers 60 Universities Teaching 77 Supercomputing Teraflops 6,000 Tesla GPUs 27 CUDA Apps 60,000 Academic Papers
- 9. 2015 3 Million CUDA Downloads 800 UniversitiesTeaching 10X GROWTH IN GPU COMPUTING 319 CUDA Apps 2008 150,000 CUDA Downloads 4,000 Academic Papers 60 Universities Teaching 77 Supercomputing Teraflops 6,000 Tesla GPUs 27 CUDA Apps 60,000 Academic Papers 450,000 Tesla GPUs
- 10. 2015 3 Million CUDA Downloads 60,000 AcademicPapers 800 Universities Teaching 54,000 Supercomputing Teraflops 10X GROWTH IN GPU COMPUTING 450,000 Tesla GPUs 319 CUDA Apps 2008 150,000 CUDA Downloads 4,000 Academic Papers 60 Universities Teaching 77 Supercomputing Teraflops 6,000 Tesla GPUs 27 CUDA Apps
- 12. 8 Billion Transistors 3,072CUDA Cores 7 TFLOPS SP / 0.2 TFLOPS DP 12GB Memory TITAN X THE WORLD’S FASTEST GPU
- 14. 0 1 2 3 4 5 6 7 TITAN X FORDEEP LEARNING Training AlexNet Days 16-core Xeon CPU TITAN TITAN Black cuDNN TITAN X cuDNN ~ 43 …
- 15. 8 Billion Transistors 3,072CUDA Cores 7 TFLOPS SP / 0.2 TFLOPS DP 12GB Memory TITAN X THE WORLD’S FASTEST GPU $999
- 16. FOUR ANNOUNCEMENTS A NewGPU and Deep Learning A Very Fast Box and Deep Learning Roadmap Reveal and Deep Learning Self-Driving Cars and Deep Learning
- 17. A SHORT HISTORYOF DEEP LEARNING Convolutional Neural Networks for Handwritten Digital Recognition LECUN, BOTTOU, BENGIO, HAFFNER, 1998 ImageNet Classification with NVIDIA GPUs KRIZHEVSKY, HINTON, ET AL., 2012 1995 2000 2005 2010 2015 Accuracy % 2010 201420122011 2013 74% 84% DNN CV 72%
- 18. “Delving Deep intoRectifiers: Surpassing Human-Level Performance on ImageNet Classification” — Microsoft: 4.94%, Feb. 6, 2015 “Deep Image: Scaling up Image Recognition” — Baidu: 5.98%, Jan. 13, 2015 “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariant Shift” — Google: 4.82%, Feb. 11, 2015 IMAGENET CHALLENGE Accuracy % 2010 201420122011 2013 74% 84% DNN CV 72%
- 19.
- 20.
- 21.
- 22. Detecting Mitosis inBreast Cancer Cells — IDSIA Predicting the Toxicity of New Drugs — Johannes Kepler University Understanding Gene Mutation to Prevent Disease — University of Toronto DEEP LEARNING REVOLUTIONIZING MEDICAL RESEARCH
- 23. “Automated Image Captioningwith ConvNets and Recurrent Nets” —Andrej Karpathy, Fei-Fei Li
- 24. DIGITS DEEP GPU TRAININGSYSTEM FOR DATA SCIENTISTS Design DNNs Visualize activations Manage multiple trainings USER INTERFACE Visualize Layers Configure DNN Process Data GPUGPU HW CloudGPU ClusterMulti-GPU Theano Torch Monitor Progress Caffe cuDNN, cuBLAS CUDA
- 25. Monitor Progress DIGITS Configure DNNProcessData Visualize Layers Test Image
- 26. DIGITS DEVBOX World’s fastestGPU Max GPU out of a plug Multi-GPU training & inference
- 27. “I’ve never seenAlexNet run this fast…TitanX is a monster, Crazy Fast” DIGITS DEVBOX — EARLY RESULTS “DIGITS makes it way easier to design the best network for the job” 0x 1x 2x 3x 4x 1 2 4 Multi-GPU scaling on Torch AlexNet VGG — Simon Osindero A.I. Architech — Soumith Chintala Research Engineer
- 28. DIGITS DEVBOX Available May2015 $15,000
- 29. FOUR ANNOUNCEMENTS A NewGPU and Deep Learning A Very Fast Box and Deep Learning Roadmap Reveal and Deep Learning Self-Driving Cars and Deep Learning
- 30. SGEMM/W 2012 20142008 20102016 48 36 12 0 24 60 2018 72 Tesla Fermi Kepler Maxwell Pascal Mixed Precision 3D Memory NVLink Volta GPU ROADMAP Pascal 2x SGEMM/W
- 31. FrameBufferCapacity(GB) 2012 20142008 20102016 40 30 10 0 20 50 2018 60 Tesla Fermi Kepler Maxwell Pascal Mixed Precision 3D Memory NVLink Volta GPU ROADMAP Pascal 2.7x Memory Capacity
- 32. HGEMM/W 2012 20142008 20102016 96 72 24 0 48 120 2018 144 Tesla Fermi Kepler Maxwell Pascal Mixed Precision 3D Memory NVLink Volta GPU ROADMAP Pascal 4x Mixed Precision
- 33. STREAMGB/s 2012 20142008 20102016 600 450 150 0 300 750 2018 900 Tesla Fermi Kepler Maxwell Pascal Mixed Precision 3D Memory NVLink Volta GPU ROADMAP Pascal 3x Bandwidth
- 34. PASCAL 10X MAXWELL CONVOLUTIONFULLY CONNECTED FULLY CONNECTED CONVOLUTION (compute) (bandwidth) (bandwidth) (compute) WEIGHT UPDATE (interconnect) 4x (FP16) 6x 6x 4x 10x Mixed Precision 3D Memory NVLINK forward backward Mixed Precision3D Memory 5x 2x * Very rough estimates
- 35. FOUR ANNOUNCEMENTS A NewGPU and Deep Learning A Very Fast Box and Deep Learning Roadmap Reveal and Deep Learning Self-Driving Cars and Deep Learning
- 36. TODAY’S ADAS PLAN ACT CPU WARN FPGA CVASIC SENSE BRAKE
- 37. NEXT-GENERATION ADAS PLAN ACT CPU WARN FPGA CVASIC SENSE BRAKE STEER ACCELERATE
- 38. NVIDIA DRIVE PXSELF-DRIVING CAR COMPUTER PLAN ACT CPU WARN FPGA CV ASIC DNN SENSE BRAKE STEER IMAGENET CHALLENGE Accuracy % 2010 201420122011 2013 74% 84% DNN CV 72% ACCELERATE
- 39. NVIDIA DRIVE PXSELF-DRIVING CAR COMPUTER PLAN ACT CPU WARN FPGA CV ASIC DNN SENSE BRAKE STEER IMAGENET CHALLENGE Accuracy % 2010 201420122011 2013 74% 84% DNN CV 72% ACCELERATE
- 40. NVIDIA DRIVE PXSELF-DRIVING CAR COMPUTER PLAN ACT CPU WARN FPGA CV ASIC DNN SENSE BRAKE STEER IMAGENET CHALLENGE Accuracy % 2010 201420122011 2013 74% 84% DNN CV 72% ACCELERATE
- 41. NVIDIA DRIVE PXSELF-DRIVING CAR COMPUTER PLAN ACT CPU WARN FPGA CV ASIC DNN SENSE BRAKE STEER IMAGENET CHALLENGE Accuracy % 2010 201420122011 2013 74% 84% DNN CV 72% ACCELERATE
- 42. NVIDIA DRIVE PXSELF-DRIVING CAR COMPUTER PLAN ACT CPU WARN FPGA CV ASIC DNN SENSE BRAKE STEER IMAGENET CHALLENGE Accuracy % 2010 201420122011 2013 74% 84% DNN CV 72% ACCELERATE
- 43. NVIDIA DRIVE PXSELF-DRIVING CAR COMPUTER PLAN ACT CPU WARN FPGA CV ASIC DNN SENSE BRAKE STEER IMAGENET CHALLENGE Accuracy % 2010 201420122011 2013 74% 84% DNN CV 72% ACCELERATE
- 44. DNN-based self-driving robot Trainingdata by human driver No hand-coded CV algorithms PROJECT LEADS Urs Muller: Chief Architect, Autonomous Driving, NVIDIA Yann LeCun: Director, AI Research, Facebook PROJECT DAVE — DARPA AUTONOMOUS VEHICLE IMAGENET CHALLENGE Accuracy % 2010 201420122011 2013 74% 84% DNN CV 72%
- 45.
- 46.
- 47.
- 48. TEST DRIVE Partially Trained(52K images)
- 49. TEST DRIVE Fully Trained(225K images)
- 50. 3,000x Faster DAVE AlexNet on DRIVE PX 3.1Million 12 38 Million 630 Million 184 116 Billion Number of Connections Frames / Second Connections / Second
- 51. NVIDIA DRIVE PXSELF-DRIVING CAR COMPUTER PLAN ACT CPU WARN FPGA CV ASIC DNN SENSE BRAKE STEER IMAGENET CHALLENGE Accuracy % 2010 201420122011 2013 74% 84% DNN CV 72% ACCELERATE
- 52. NVIDIA DRIVE™ PX SELF-DRIVINGCAR COMPUTER Available May 2015 $10,000
- 53.
- 54. LEAPS IN VISUALCOMPUTING TITAN X The World’s Fastest GPU DIGITS DevBox GPU Deep Learning Platform Pascal — 10x Maxwell For Deep Learning NVIDIA DRIVE PX Deep Learning Platform for Self-Driving Cars