2018 Intel AI Developer Conference Keynote
2 likes203 views
The document discusses advancements in technology from operant conditioning in 1938 to the development of deep learning and machine learning tools by major companies like Intel and Amazon. It outlines various hardware and software optimizations, machine learning applications, and performance testing conducted primarily on Intel's processors, emphasizing benchmarks and configurations used. Furthermore, it mentions the importance of collaboration, forward-looking statements, and the need to review performance tests for accurate evaluations.
2018 Intel AI Developer Conference Keynote
- 3. WELCOME TO
- 7. BEST TOOLS TO BUILD TOWARDS
- 9. Tools
- 11. Convolution Matrixmultiplication BatchNorm poolingnormalization activation HELPS REALIZE THE INCREDIBLE BENEFITS OF DIRECT OPTIMIZATION Intel/mkl-dnn OPEN SOURCE
- 12. OPTIMIZING TENSORFLOW Other names and brands may be claimed as the property of others
- 13. OPEN SOURCE COMPILER ENABLING FLEXIBILITY TO RUN MODELS ACROSS A VARIETY OF FRAMEWORKS AND HARDWARE NervanaSystems/ngraph
- 14. Other names and brands may be claimed as the property of others
- 15. ENABLING DEEP LEARNING TO TAKE ADVANTAGE OF SCALABLE SPARK AND HADOOP CLUSTERS Other names and brands may be claimed as the property of others intel-analytics/BigDL
- 16. USING DATA SCIENCE TO Other names and brands may be claimed as the property of others
- 17. USING DATA SCIENCE TO
- 18. EmergencyResponse Financialservices MachineVision Cities/transportation AutonomousVehicles ResponsiveRetail Manufacturing Publicsector
- 19. VISUAL INFERENCING AND NEURAL NETWORK OPTIMIZATION DEPLOY COMPUTER VISION AND DEEP LEARNING CAPABILITIES TO THE EDGE HighPerformance,highEfficiencyfortheedge Writeonce+scaletoDiverseAccelerators BroadFrameworksupport Other names and brands may be claimed as the property of others VPU = Vision Processing Unit (Movidius)
- 20. Tools
- 21. Hardware
- 23. Other names and brands may be claimed as the property of others Source: https://research.fb.com/publications/applied-machine-learning-at-facebook-a-datacenter-infrastructure-perspective/
- 24. Other names and brands may be claimed as the property of others
- 25. Other names and brands may be claimed as the property of others AScientificCollaborationbetween IntelandNovartis
- 26. 224x224x3 ImageNet Other names and brands may be claimed as the property of others Software and workloads used in performance tests may have been optimized for performance only on Intel® microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit www.intel.com/benchmarks.
- 27. 1024x1280x3 26xlarger Other names and brands may be claimed as the property of others Software and workloads used in performance tests may have been optimized for performance only on Intel® microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit www.intel.com/benchmarks. 224x224x3 ImageNet
- 28. 26xlargerMultipleobjects Other names and brands may be claimed as the property of others Software and workloads used in performance tests may have been optimized for performance only on Intel® microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit www.intel.com/benchmarks. 1024x1280x3
- 29. Other names and brands may be claimed as the property of others § Configuration: CPU: Xeon 6148 @ 2.4GHz, Hyper-threading: Enabled. NIC: Intel® Omni-Path Host Fabric Interface, TensorFlow: v1.7.0, Horovod: 0.12.1, OpenMPI: 3.0.0. OS: CentOS 7.3, OpenMPU 23.0.0, Python 2.7.5 Time to Train to converge to 99% accuracy in model Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit http://www.intel.com/performance. MultiscaleConvolutionNeuralNetwork Intel® MKL/MKL-DNN, clDNN, DAAL OptimizedLibraries Intel®Omni-PathArchitecture ScalingofTimetoTrainIntel® Omni-Path Architecture, Horovod and TensorFlow® Speedupcomparedtobaseline1.0 measuredintimetotrainin1nodes 1 Node 2 Nodes 4 Nodes 8 Nodes 1 Node 2 Nodes 4 Nodes 8 Nodes TOTALMEMORYUSED192GB DDR4 PER INTEL® 2S XEON® 6148 PROCESSOR 128.6GB 257.2GB 514.4GB 64.3GB
- 31. Other names and brands may be claimed as the property of others
- 32. ZIVA VFX Authoring Tools Shipped as a Maya Plugin DISTRIBUTED TO SERVER FARM FOR SHOT RENDERS Intel Paradiso IntelBLAS IntelLapack ZIVA FEM PHYSICS SOLVER IntelMKL Bones/ Muscles Fascia FatandSkin ZIVA CHARACTERS SIMULATED IN PASSES Charactertransfer/ automation Volumetriccapture augmentation Real-TimeTraining Embeddedplayersin anysoftware Distributedgraphs NativeNodegraphui withrobustapi A.I & M.L TECHNOLOGIES DISTRIBUTED FUNCTIONALITY AND EMBEDDED PLAYERS ZIVA INTEGRATION FRAMEWORK CLOUD SERVICES TOOLS ENGINES Other names and brands may be claimed as the property of others
- 37. •Deploy DNN and Computer Vision at the Edge •Native FP16 and Fixed Point 8 bit support •4 TOPS with 1 TOPS of DNN Compute at 1W
- 42. Chip X Lake Crest Theoretical Reality TOPs 0 Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit http://www.intel.com/performance Chip X GEMM based on DeepBench training data for A(5124, 9124), B(9124,2560) matrix size GEMM operations performing DeepSpeech using FP16+ mixed precision at 27.43 TOPs. Source: Lake Crest, Based on Intel measurements on limited distribution SDV, General Matrix-Matrix Multiplication; A(1536, 2048), B(2048, 1536).
- 43. Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit http://www.intel.com/performance Chip X GEMM based on DeepBench training data for A(5124, 9124), B(9124,2560) matrix size GEMM operations performing DeepSpeech using FP16+ mixed precision at 27.43 TOPs. Source: Lake Crest: Based on Intel measurements on limited distribution SDV 1 General Matrix-Matrix Multiplication; A(1536, 2048), B(2048, 1536) 2 Two chip vs. single chip GEMM performance; A(6144, 2048), B(2048, 1536) 3 Full Chip MRB-CHIP MRB data movement using send/recv, Tensor size = (1, 32), average across 50K iterations MULTI-CHIP COMMUNICATION3 Power<210W 2.4Tb/s OFF CHIP BANDWIDTH <790ns LATENCY 96.4% GEMM OPERATION UTILIZATION1 A(1536, 2048) B(2048, 1536) MULTI-CHIP SCALING2 96.2%A(6144, 2048) B(2048, 1536) Chip X Lake Crest Theoretical Reality TOPs 0
- 44. Chip X Lake Crest Spring Crest (Estimate) Theoretical Reality TOPs 0 Software and workloads used in performance tests may have been optimized for performance only on Intel microprocessors. Performance tests, such as SYSmark and MobileMark, are measured using specific computer systems, components, software, operations and functions. Any change to any of those factors may cause the results to vary. You should consult other information and performance tests to assist you in fully evaluating your contemplated purchases, including the performance of that product when combined with other products. For more complete information visit http://www.intel.com/performance Chip X GEMM based on DeepBench training data for A(5124, 9124), B(9124,2560) matrix size GEMM operations performing DeepSpeech using FP16+ mixed precision at 27.43 TOPs. Source: Lake Crest - Based on Intel measurements on limited distribution SDV Source: Spring Crest - Intel measurements on simulated product
- 45. firstcommercialNNP Intel®Nervana™ NNPL-1000in 2019 3-4x training performance of first generation Lake Crest product PURPOSE BUILT DESIGN OPTIMIZED ACROSS MEMORY BANDWIDTH, UTILIZATION, AND POWER
- 46. Hardware
- 47. Community
- 48. PARTNERSHIPS AND
- 49. Other names and brands may be claimed as the property of others
- 50. Other names and brands may be claimed as the property of others
- 51. Machine Learning at Amazon: a long heritage Person alized recommen d ation s Inventin g entirely n ew c u stomer exp erien c es F u lfillment au tomation / inventor y man agement Cargo Voic e d riven interac tion s Other names and brands may be claimed as the property of others
- 52. Fully managed hosting with auto- scaling One-click deployment Pre-built notebooks for common problems Built-in, high performance algorithms One-click training Hyperparameter optimization BUILD TRAIN DEPLOY A m a z o n S a g e M a k e r
- 53. A W S M a c h i n e L e a r n i n g S t a c k FRAMEWORKS AND INTERFACES APPLICATION SERVICES PLATFORM SERVICES KERAS Frameworks Interfaces P O L L YR E K O G N I T I O N L E XR E K O G N I T I O N V I D E O T R A N S C R I B E T R A N S L A T E C O M P R E H E N D AMAZON SAGEMAKER INFRASTRUCTURE EC2 GPUs EC2 CPUs IoT Edge AWS DEEPLENS
- 54. Other names and brands may be claimed as the property of others
- 55. Other names and brands may be claimed as the property of others
- 58. GOAL IS TO SELECT A SET OF ML PROBLEMS, EACH DEFINED BY A DATASET AND QUALITY TARGET, THEN MEASURE THE WALL CLOCK TIME TO TRAIN A MODEL FOR EACH PROBLEM. For more information, see: https://mlperf.org/ Other names and brands may be claimed as the property of others
- 59. AICHALLENGE.INTEL.COM *Idea implementation at the Olympic Games subject to approval
- 61. Today’s presentation contains forward-looking statements. All statements made that are not historical facts are subject to a number of risks and uncertainties, and actual results may differ materially. Please refer to our most recent earnings release, Form 10-Q and 10-K filing available on our website for more information on the risk factors that could cause actual results to differ.