jetson : ai at the edge
Post on 20-Mar-2017
139 Views
Preview:
TRANSCRIPT
Serge Palaric, VP Sales & Marketing EMEA - Embedded
JETSON : AI AT THE EDGE
2
NVIDIA: THE AI COMPUTING COMPANY
GPU Computing Visual Computing Artificial Intelligence
3
AMAZING ACHIEVEMENTS IN AI
Play Go Play Doom Learn Paint Style Synthesize Voice
Write Captions Learn Motor Skills Learn to Walk Drive
4
NVIDIA AI COMPUTING ECOSYSTEM
AI-powered Consumer Services
AI-as-a-Service AI for Enterprise >1,500 AI Startups
iQIYI JD.comGoogleFlickr
Amazon FacebookeBayBaidu
ShazamQihoo 360 Skype Sogou
PeriscopePinterestNetflixMicrosoft
TwitterTencent Yandex Yelp
AI for Auto
5
GPU DEEP LEARNING IS A NEW COMPUTING MODEL
DGX-1
Training Datacenter Inference
Tesla
Inference at the Edge
Jetson
6
WHY AI AT THE EDGE MATTERSLatencyBandwidth Availability
1 billion cameras WW (2020)
30B Inference/Second
30 images per second
200ms latency
“Billions of intelligent devices will take advantage of DNNs to provide personalization and localization as GPUs become faster and faster over the next several years.” — Tractica
50% of world at less than 8mbps
Only 73% 3G/4G availability WW
7
JETSON: THE PLATFORM FOR AI AT THE EDGE
Industrial InspectionSearch and Rescue
Package DeliveryFactory AutomationEnterprise Collaboration Public Safety
Personal Assist
Service Robotics
Portable Medical Academia and Research
8
AI Smart Industries
MiningLogisticsIntelligent FactoryIntelligent WarehouseSmart Operation….
Smart CameraNVRAppliancesServers
Service RobotsUniversities
LogisticAgricultureInspectionSecurityEmergency
PlanesRadarsTanksTrainsVision system….
CamerasScannersDiagnostic….
9
AI: THE NEW INDUSTRIAL REVOLUTION
Intelligent FactoryPick and placeComplex/custom tasksVisual inspectionTask consolidationDynamic reconfigurationCollaborative roboticsEfficiency optimizationFactory simulation
Smart OperationsInfrastructure inspectionPredictive maintenancePhysical security
LogisticsAutopilot/self-driving trucksRobot/drone delivery and support
Intelligent WarehouseInventory managementBin pickingPallet movement
MiningEquipment automationOperational safety
10
SMART FACTORY EXAMPLE
AOI • Autonomous Optical Inspection
Operational efficiency• Energy efficiency• Improved Uptime• Predictive Maintenance• Make one of many• Make many of one• Picking-placing• Screwing/fastening/rivetingMan-Machine Co-existence• Collaborative Human-RobotTraceability
Challenge• Inspection/Quality
assurance • DL-Picking and placing• Cobot applications
( Deep Learning)• Big Data analytics• Real time analysis of
data from sensors
Solution
11
AI FOR INDUSTRIAL AND COMMERCIAL UAVS
LogisticsWarehouse automationPackage delivery
InspectionWind turbinesBridgesOil rigsPipelinesHigh-voltage power linesCell towers
Precision AgriculturePlantingSpraying
SecurityEnterprise securityAd hoc security systems
Emergency ResponseFirst ResponderSearch and Rescue
12
ProblemAerial inspection is- Imprecise: often needs
multiple flights- Time consuming: manual
review of footage- Dangerous: drone
crashes into subject or operator
SolutionAutomate the process- Vision-enabled
navigation- On-board verification- On-board fault
classification
13
AI CITY — 1B CAMERAS BY 2020
~1 billion cameras worldwide by 2020
Þ 30 billion inferences/sec
Tesla P40: 2,500 inferences/sec @ 720P
Þ AI City needs ~10M P40 servers
DATA: 1B cameras, IHS “Video Surveillance Intelligence Service, Aug. 2016”
14 NVIDIA CONFIDENTIAL. DO NOT DISTRIBUTE.
FROM CAMERA TO CLOUD
CAMERA APPLIANCE DATACENTER
Trained Neural Networks
+
ESLA
EDGE CLOUD
TRAINING DATA
JETSON
JETSON
TESLA
TESLA
15
JETSON FOR AI AT THE EDGE
Decode DetectorsClassifiersTrackers
Composite Encode
4K display
4K video
Jetson TX1 AI Pipeline
JETSON TX13 DNNsObject tracking4K30 video decodeVideo compositing4K HDMI outputH.265 video encode
16
Jetson TX1AI Computer on a ModuleAdvanced tech for intelligent machinesUnmatched performance under 10WSmaller than a credit card
17
Develop and deployJetson TX1 Developer Kit & Jetson TX1 Module
18
Jetson TX1 Developer KitJetson TX1Developer Board5MP Camera
19
Jetpack SDKLibrariesDeveloper toolsDesign collateralDeveloper ForumTraining and TutorialsEcosystem
http://developer.nvidia.com/embedded-computing
COMPREHENSIVE DEVELOPER SITE - JEP
20
JETSON TX1
GPU 1 TFLOP/s 256-core Maxwell
CPU 64-bit ARM A57 CPUs
Memory 4 GB LPDDR4 | 25.6 GB/s
Storage 16 GB eMMC
Wifi/BT 802.11 2x2 ac/BT Ready
Networking 1 Gigabit Ethernet
Size 50mm x 87mm
Interface 400 pin board-to-board connector
21
NVIDIA JETPACK 2.3SDK for embedded AI computing
Deep Learning
TensorRTcuDNN
DIGITS Workflow
Computer Vision
VisionWorksOpenCV
Multimedia
ISP SupportCamera Imaging
Video CODEC
Also includes ROS compatibility, OpenGL, advanced developer tools, and much more
CUDACUDA Libs
GPU Compute
22
JETSON TX1: 20X DEEP LEARNING EFFICIENCY
Footnotes:The efficiency was measured using the methodology outlined in the whitepaper: https://www.nvidia.com/content/tegra/embedded-systems/pdf/jetson_tx1_whitepaper.pdfJetson TX1 efficiency is measured at GPU frequency of 691 MHz.Intel Core i7-6700k efficiency was measured for 4 GHz CPU clock.GoogLeNet batch size was limited to 64; that is the maximum supported by Jetpack 2.0. With Jetpack 2.3 and TensorRT, GoogLeNet batch size 128 is also supported for higher perf.FP16 results for Jetson TX1 are comparable to FP32 results for Intel Core i7-6700k as FP16 incurs no classification accuracy loss over FP32. Latest publicly available software versions of IntelCaffe and MKL2017 beta were used.For Jetpack 2.0 and Intel Core i7, non-zero data was used for both weights and input images. For Jetpack 2.3 (TensorRT) real images and weights were used.
1 2 640
5
10
15
20
25Inference (GoogLeNet)
Intel core i7-6700kJetson TX1/Jetpack 2.1 - Nov '15Jetson TX1/Jetpack 2.3 - Sept '16
Batch Size
Imag
es/se
c/wa
tt
23
DEEP LEARNING END-TO-END
Train
Step 1: Train
Optimize Deploy
NVIDIA DGX-1: Train your model with NVIDIA DIGITS Software on DGX-1, this highest performance training solution for DNNs
TensorRT
TensorRT: Dramatically speed up and reduce memory usage for your model on Jetson TX1
Jetson TX1: Deploy your model to your fleet of Jetson TX1-enabled products. Jetson TX1 is the highest performance inference solution under 10W
24
DEEP LEARNING INSTITUTELearn the latest Deep Learning techniques
Online coursesDeep Learning labsNanodegreesNetworking and events
www.nvidia.com/object/deep-learning-institute.html
25
TWO DAYS TO A DEMODevelop a Deep Learning Demo on Jetson TX1 in two days
Prerequisites:• GPU-enabled PC or server• Jetson TX1 Developer Kit• One engineer
developer.nvidia.com/embedded/twodaystoademo
26
PARTNERSHIPLatest Technology
GPU/DL Experts
Global Sales & Marketing Network
Training from Deep Learning Institute
INCEPTION
PROGRAM
CONNECTING WITH AI STARTUPS
nvidia.com/Inception
27
Jetson (corp): www.nvidia.com/object/embedded-systems.htmlJetson (dev): developer.nvidia.com/embedded-computingEmbedded Country selector: www.nvidia.com/embedded
Jetpack (under develop/tools): https://developer.nvidia.com/embedded/develop/tools Jetpack is a one click installer that will perform BSP installation, plus all relevant SDK (CUDA, cuDNN, TensorRT, VisionWorks)
Success Stories: developer.nvidia.com/embedded/learn/success-storiesPartners and Ecosystem: developer.nvidia.com/embedded/communityDeep Learning Institute: www.nvidia.com/object/deep-learning-institute.html Two Days To A Demo: developer.nvidia.com/embedded/twodaystoademoInception Program: www.nvidia.com/inception
RESOURCES
28
END TO END SOLUTIONONE ARCHITECTURE FOR ALL
HYPERSCALE HPC
TESLA
VISUALIZATION
QUADRO
RESEARCHERS/ EARLY ADOPTERS
DGX-1
EDGE COMPUTING
JETSON
THANK YOU
top related