QUALCOMM Incorporated · NASDAQ:QCOM
QUALCOMM — Investor Day (April 2019)
139 slides from QUALCOMM's investor day presentation, Apr 9, 2019
Slide 1: Advancing the AI future
April 9, 2019 San Francisco, CA @cristianoamon Advancing the AI future Cristiano Amon President Qualcomm Incorporated
Slide 2: A world where everyone and everything is intelligently connected
A world where virtually everyone and everything is intelligently connected
Slide 3: We pioneered the technologies powering this future
Connectivity Processing Qualcomm Snapdragon is a product of Qualcomm Technologies, Inc and / or its subsidiaries. We pioneered the technologies powering this future Leadership across advanced connectivity, processing, and systems design
Slide 4: Pioneering on-device intelligence
Automotive Mobile Drones Computing Extended Audio reality (XR) Robotics IP Cameras Qualcomm® AI Engine Pioneering on-device intelligence Qualcomm AI Engine is a product of Qualcomm Technologies Inc. and/or its subsidiaries.
Slide 5: Consistent AI R&D investment is the foundation for product leadership
Qualcomm Artificial Intelligence Research Consistent AI R&D 1st Gen AI 2nd Gen AI 3rd Gen AI Qualcomm® (Snapdragon 820) (Snapdragon 835) (Snapdragon 845) QCS400 (First audio SoC) investment is the foundation 4th Gen AI (Snapdragon 855) Qualcomm® Snapdragon 660 Qualcomm® for product leadership Neural
Slide 6: Collaborating with leading cloud players and ISVs
Collaborating with leading Cloud Software cloud players and ISVs
Slide 7: The age of 5G and AI
Efficiency Immediacy New Local value experiences The age of 5G and AI Billions of connected things Privacy Autonomy Massive amounts of data Powering new services at the wireless edge Security Personalization
Slide 8: Driving significant economic impact
$12.3 Trillion In goods and services enabled by 5G in 2035 Source: The 5G Economy, an independent study from IHS Markit, Penn Schoen Berland and Berkeley Research Group, commissioned by Qualcomm. $3.9 Trillion In business value by 2022 Source: Gartner, Inc., Gartner Says Global Artificial Intelligen
Slide 9: 5G rollout happening faster than 4G
5G rollout happening faster than 4G 4 Operators 20+ Operators launched announced 3 20+ OEMs OEMs launched announced Source: IHS Report Jan ’19, Qualcomm Technologies data Year 1 announcements underscore tremendous momentum with 5G
Slide 10: On-device intelligence is quickly gaining momentum
On-device intelligence is quickly gaining momentum Key segments are expected to see full AI attach rates by 2025 10% AI attach rate 100% AI attach rate 2018 2025 PCs / Smart Mobile Automotive XR Tablets speakers Source: Tractica, 2019
Slide 11: Driving distributed intelligence
Public network Driving distributed intelligence Cloud Edge cloud On-device Bringing the cloud closer to devices at the edge Private networks
Slide 12: AI will drive transformation across industries
AI will drive transformation across industries
Slide 13: Shaping the future of transportation
Shaping the future of transportation Personalized driver settings Driver awareness monitoring Greater autonomous capabilities
Slide 14: Fueling a new era of cloud gaming
Fueling a new era of cloud gaming The cloud is becoming the new console Gaming Personal Smartphones Tablets consoles computers
Slide 15: Boundless mobile XR experiences
Boundless mobile XR experiences
Slide 16: Powering the factory of the future
Powering the factory of the future
Slide 17: The wireless edge realizes the full potential of 5G and AI
The wireless edge realizes the full potential of 5G and AI Inventing technology at scale to realize the promise of AI on trillions of connected devices
Slide 18: Uniquely positioned to power the intelligently connected future
X50 Qualcomm Ventures AI Fund Foundational 5G + AI Systems Advanced Ecosystem R&D technology design silicon investment leadership expertise Uniquely positioned to power the intelligently connected future
Slide 19: Bringing leading performance per watt to the cloud
April 9, 2019 @qualcomm San Francisco, CA Bringing leading performance per watt to the cloud Keith Kressin SVP, Product Management Qualcomm Technologies, Inc.
Slide 20: Uniquely positioned to make the intelligent wireless edge a reality
Client AI Bringing Artificial Intelligence to the Client Edge 5G + AI Leader in wireless 10+ years of AI Broad industry leadership edge development research collaboration Uniquely positioned to make the intelligent wireless edge a reality Qualcomm Snapdragon is a product of Qualcomm Technologies, In
Slide 21: Extensive mobile heritage
Process node leadership Design Extensive expertise mobile heritage Low power consumption Scale Signal processing expertise
Slide 22: Bringing power efficiency to all mobile clients
Bringing power efficiency to all mobile clients
Slide 23: Process node leadership
7nm 10nm SD855 SD835 14nm/12nm SD820 2015 2016 2018 Process node leadership
Slide 24: Scale
Scale
Slide 25: Low power signal processing across all key user experiences
Camera AI Video Gesture Audio XR Low power signal processing across all key user experiences
Slide 26: An architecture shift in AI cloud inferencing
An architecture shift in AI cloud inferencing AI Accelerators FPGA >10X Improvement >10X Improvement
Slide 27: Market ramp of AI in datacenters
By 2025, market ramp 10x Growth of AI in 18 datacenters 16 14 Revenue (Billions) 12 $17 10 8 6 4 Billion 2 0 2018 2019 2020 2021 2022 2023 2024 2025 Source: Tractica
Slide 28: Need to lower latency and increase cloud AI processing performance
Signal Input Processing Client Edge Cloud Edge Cloud Technology Technology Imaging & video Voice Audio Sensors Need to lower latency and increase cloud AI processing performance
Slide 29: 5G and more powerful edge cloud processing will transform user experiences
Signal Input Processing Client Edge Connectivity Edge Cloud Cloud Technology Technology Imaging & video Voice Audio Sensors 5G and more powerful Edge Cloud processing will transform user experiences
Slide 30: AI at Facebook
April 9, 2019 @qualcomm San Francisco, CA AI at Facebook Joe Spisak Product Manager Facebook AI
Slide 31: Full Stack Approach
Full Stack Approach DATA MODELS LIBRARIES FRAMEWORKS COMPILERS AND OPTIMIZERS HARDWARE
Slide 32: AI Powered Use Cases
AI Powered Use Cases
Slide 33: Enhancing Existing Products
Enhancing Existing Products SOCIAL R E C O M M E N D AT I O N S
Slide 34: Enhancing Existing Products
Enhancing Existing Products SOCIAL MACHINE R E C O M M E N D AT I O N S T R A N S L AT I O N S
Slide 35: Enhancing Existing Products
Enhancing Existing Products SOCIAL MACHINE ACCESSIBILITY R E C O M M E N D AT I O N S T R A N S L AT I O N S IMAGE C O N TA I N S Diana and Hanna on green bicycles Riding bikes in Santo Domingo Biking near a red building
Slide 36: Powering New Experiences
Powering New Experiences BOTS & G E N E R AT E D AR VR A S S I S TA N T S CONTENT EFFECTS HARDWARE
Slide 37: Datacenter AI Trends
Datacenter AI Trends
Slide 38: 5.95B+ Translations Per Day
O# 1) Like © comment @ Share S.} Abélia Cocher Gardez les étagéres loin du lit. Abélia Cocher and 23 others 8 Comments 5.95B+ Translations Per Day 1h Like Reply See Translation Write a comment... 2)
Slide 39: NLP In Action
NLP In Action 35M+ Blood Donors
Slide 40: $1B+ Raised For Charitable Giving Powered By AI
$1B+ Raised For Charitable Giving Powered By AI
Slide 41: 200+ Trillion Inferences Per Day
200+ Trillion Inferences Per Day
Slide 42: Data-Center Power Consumption
Data-Center Power Consumption Data-center power consumption is doubling every year 4x 3x 2x 1x 0 Y1 Q3 Y1 Q4 Y2 Q1 Y2 Q2 Y2Q3 Y2 Q4 Y3 Q1 Y3 Q2 Deep Learning Inference in Facebook Data Centers: Characterization, Performance Optimizations and Hardware Implications – Nov 2016, https://www.researchgate
Slide 43: Key Attributes For Next Generation Inference
Key Attributes For Next Generation Inference Reliability Latency Power Efficiency
Slide 44: Modular server, 8-socket server
Modular server 8-socket server OCP accelerator module 8-accelerator baseboard 8-accelerator platform
Slide 45: Research to Production
Research to Production
Slide 46: Research to Production
Research to Production PROTOTYPING DEPLOYING
Slide 47: PyTorch
PyTorch A machine learning framework with an emphasis on: DYNAMIC HARDWARE NEURAL ACCELERATED NETWORKS INFERENCE EAGER & DISTRIBUTED SIMPLICITY GRAPH TRAINING OVER BASED COMPLEXITY EXECUTION
Slide 48: PyTorch
PyTorch 2.8X #2 INCREASE IN GITHUB FASTEST GROWING OPEN CONTRIBUTORS SOURCE PROJECT
Slide 49: Glow
Glow INPUT COMPILER CODE-GEN TARGETS ONNX X86 CPU PYTORCH GLOW CORE ASIC1 C++ ASIC2 PROFILE DATA
Slide 50: Full Stack Approach
Full Stack Approach DATA MODELS LIBRARIES FRAMEWORKS PYTORCH 1.0 COMPILERS AND OPTIMIZERS GLOW, … HARDWARE BIG BASIN, TIOGA PASS, …
Slide 51: Software
Software
Slide 52: Full software stack
Full Features software Newsfeeds Advertising Personalized Search XR Gaming videos stack GLOW Performance Compiler Compiler Monitoring Tools Profilers Debuggers Card Tuning Quantizers Runtimes Frameworks Cognitive Toolkit
Slide 53: AI at Microsoft
April 9, 2019 @qualcomm San Francisco, CA AI at Microsoft Venky Veeraraghavan (@venkyv) Partner Group Program Manager Microsoft Corp
Slide 54: AI on a massive global network
AI on a massive Data global network Your data + Microsoft data Cloud Power of Azure Models © Microsoft Corporation Breakthrough advancements
Slide 55: Cloud to edge AI devices
Sensors + control Sensors to interactive On-premises infrastructure Global scale processing MCU based IoT devices Edge devices Edge appliances Edge cloud Hyperscale cloud Inferencing Inferencing Inferencing Inferencing Inferencing Inferencing Training Training Training Training Training Training Dat
Slide 56: Slide 56
Slide 57: Missing equipment, flammable equipment near heat source
Missing equipment Flammable equipment near heat source No hat
Slide 58: No hat
No hat
Slide 59: ONNX - open specification for ML models
ONNX - open specification for ML models Create Deploy Frameworks Azure Azure Machine Learning services Ubuntu VM Windows Server 2019 VM ONNX model Specification Devices Services Edge Cloud & Appliances Azure Custom Vision Service Edge & IoT Devices ONNX Runtime is open source
Slide 60: Windows ML
Windows ML • Runs on Qualcomm Snapdragon SDM850 and the new Snapdragon 8CX Always Connected PCs • Windows ML API allows developers to easily integrate pre- trained ML models into their applications • Built on top of DirectML, a low- level API in the DirectX family Thundersoft AI Developer Platform
Slide 61: Intelligent Edge AI Devices
Intelligent Edge AI Devices Vision AI Developer Kit Hololens 2 Based on the Qualcomm 603 Based on the Qualcomm 845
Slide 62: Responsible AI
Responsible AI Better product serving the Broader Population Responsibility and Social Impact Legal and Policy Competitive Advantage and Brand
Slide 63: In closing...
In closing... Microsoft provides a comprehensive platform for Machine Learning that leverage Qualcomm silicon • Hyperscale end-to-end Cloud AI platform • Consistent & Open Edge AI platform for IOT and Windows devices • Foundation for Responsible AI © Microsoft Corporation
Slide 64: Performance
Performance
Slide 65: Built from the ground up for AI: Cloud AI 100
Built from the Peak AI performance ground lcomm ® up AI 100 Cloud >50x 3x 1x Snapdragon Snapdragon Qualcomm 845 855 Cloud AI 100
Slide 66: Qualcomm Cloud AI 100
Qualcomm Cloud AI 100 ® • Built on 7nm • >350 TOPS Peak AI Performance • Sampling 2nd half of 2019
Slide 67: Mobile AI
April 9, 2019 @qualcomm San Francisco, CA Mobile AI Ziad Asghar VP, Product Management Qualcomm Technologies, Inc. (QTI)
Slide 68: Mobile: the most pervasive AI platform
Mobile – THE most pervasive AI platform >1 Billion AI capable devices enabled with QTI technology
Slide 69: Consistent AI R&D investment is the foundation for product leadership
Qualcomm Artificial Intelligence Research Consistent AI R&D 1st Gen AI 2nd Gen AI 3rd Gen AI Qualcomm® (Qualcomm® (Snapdragon 835) (Snapdragon 845) QCS400 Snapdragon™ 820 (First audio SoC) Mobile Platform) investment is the foundation 4th Gen AI Qualcomm® (Snapdragon 855) Cloud AI 100 Qualcomm® Snap
Slide 70: Power and thermal efficiency are essential for on-device AI
The challenge Constrained mobile of AI workloads environment Very compute intensive Must be thermally efficient for sleek, ultra-light designs Large, complicated neural network models Requires long battery life for all-day use Complex concurrencies Storage / Memory Always-on / Real-time Power and th
Slide 71: Adreno 640, Hexagon 690, Kryo 485
Adreno 640 50% More ALUs* FP32 & FP16 Hexagon 690 New Tensor Accelerator Optimized scalar • QTI designed Voice Assistant • Dedicated to AI • Multidimensional math and INT16, INT8 & Mixed integrated nonlinear functions 4x Vector eXtensions* Kryo 485 New dot product instructions FP32 & INT8 *Compared
Slide 72: 4th Gen AI Engine
Qualcomm® AI Engine Mobile Apps NN Frameworks Cognitive Toolkit Runtime Software Frameworks 4th Gen Tensorflow Lite Google NN API Neural Processing SDK AI Engine Libraries Qualcomm® Math Libraries Open CL Hexagon NN Cores Hexagon Kryo CPU Adreno GPU Scalar Vector Tensor Qualcomm Math Libraries and Q
Slide 73: Snapdragon 855 AI performance
AI performance (common networks) Competitor A 855 Snapdragon 855 AI 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 Note: Average over MobileNet, Inceptionv3, Resnet34 on commercial devices. performance AI performance (public benchmarks) Competitor A 855 0.0 0.5 1.0 1.5 2.0 2.5 Note: Average over ETH, NeuroScope, A
Slide 74: Snapdragon AI power advantage
AI performance/power Snapdragon Competitor A 855 AI power advantage 855 0.0 0.5 1.0 1.5 2.0 Note: Power measured running AIMark.
Slide 75: Innovating in new tiers
Announcement Camera AI Gaming Innovating in new tiers
Slide 76: Snapdragon 665 Mobile Platform
Snapdragon 665 Mobile Platform
Slide 77: Snapdragon 665 Mobile Platform
Snapdragon 665 Mobile Platform Stunning pictures at every angle • Up to 2X faster AI* • Triple camera • 48MP * Compared to Snapdragon 660
Slide 78: Qualcomm AI Engine + Camera
Qualcomm AI Engine + Camera
Slide 79: Snapdragon 730 Mobile Platform
Snapdragon 730 Mobile Platform
Slide 80: Snapdragon 730 Mobile Platform
Snapdragon 730 Mobile Platform Awe-inspiring video capabilities • 1st CV-ISP in 7 Series • 1st Tensor Accelerator in 7 Series • 1st True HDR in 7 Series
Slide 81: Firsts in the 7 series
Firsts in the 7series 4th Gen 1 ST Qualcomm 1ST CV-ISP AI Engine Qualcomm Tensor 1ST Spectra 1 ST Accelerator 3-series Qualcomm ® apt-X™ Kryo 4- 1 ST 1ST series Adaptive Audio Wi-Fi 6 8nm 1ST ready 1ST process * Compared to Snapdragon 710 Qualcomm aptX is product of Qualcomm Technologies, Inc. and/o
Slide 82: Always-on, fast, more intuitive and secure
Smart cropping Payments Augmented Reality Voice Assistant On- device AI Always-on, fast, more intuitive and secure
Slide 83: Snapdragon 730G Mobile Platform
Snapdragon 730G Mobile Platform
Slide 84: Snapdragon 730G Mobile Platform
Snapdragon 730G Mobile Platform Exceptional gaming performance and more • Select Snapdragon elite gaming features • 15% faster graphics rendering than Snapdragon 730
Slide 85: AI support on QTI SoCs
= AI Engine AI support on QTI SoCs = AI Capable Snapdragon 8 Series SD820 SD835 SD845 SD855 Snapdragon 7 Series SD710 SD730 SD730G SD660 SD665 Snapdragon 6 Series SD630/36 SD670 Snapdragon 4 Series SD427 SD429 SD439 SD 450 Compute 835 850 8cx IOT QCS8053 QCS8009 QCS603 QCS605 XR XR1 Auto 820A 3rd Ge
Slide 86: Key AI use cases
Style Transfer Speaker Single & Dual Camera Voice Landmark Detection Recognition Bokeh Biometrics Gesture Recognition Video Style Transfer OIS Ultrasound based Artistic Style Transfer Audio Context virtual sensors Noise Detection Super HDR Suppression Text Super Nightshot Object Location Spatial Aud
Slide 87: Google Lens Dense Text Copy
Google Lens Dense Text Copy CNN CNN + QLSTM Android’s NN API Hexagon CPU DPU Qualcomm AI Engine Qualcomm NNAPI is a product of Qualcomm Technologies, Inc. and/or its subsidiaries.
Slide 88: XR gaming
XR gaming AI improves gaming experience on the device Display Mechanics XR World Visual style transfer AI opponent People/object detection AI game assistant Scene segmentation Control Pose estimation Intent detection Hands tracking Hearing Neural Processing SDK Eye tracking Audio No howling Body mot
Slide 89: AI "On" and "In" the device
AI “On” and “In” the device True personal assistance Modem Superior photography VR /AR Natural user interfaces Power Enhanced connectivity Enhanced security IC Design A new development paradigm where things repeatedly improve
Slide 90: AI software ecosystem
Camera Audio / Translation Gesture Automotive AI software ecosystem Augmented reality
Slide 91: Frameworks, OS, Ecosystem
Frameworks OS Ecosystem Features Devices Noise Suppression Super Resolution Night Shot Face Recognition Speech Recognition Qualcomm Object AI Engine Detection Video Segmentation Cognitive Bokeh Toolkit
Slide 92: AI for IoT
April 9, 2019 @qualcomm San Francisco, CA AI for IoT Sahil Bansal Sr. Director, Product Management Qualcomm Technologies, Inc.
Slide 93: Emerging growth and opportunities for AI in IoT
Emerging growth and opportunities for AI in IoT
Slide 94: On-device AI and IoT
On-device AI and IoT AI Engine Bluetooth Camera processing Bluetooth Mesh Audio processing Wi-Fi Connectivity Computing Sensor processing 5G CPU 4G/LTE GPU DSP NFC Building on Powerline our leadership Video processing in mobile GNSS/ Display Location innovation processing Power management Secure boo
Slide 95: On-device computing is paramount
On-device computing is paramount Process data closest to the source, complement the cloud Privacy Perceive Hear, see, monitor, and observe Reliability Reason Learn, Immediacy Connectivity infer context, and anticipate Training 5G, 4G LTE, Efficiency Execution/inference NB IoT, Wi-Fi, BLE, 15.4 Act I
Slide 96: To scale, intelligence must be distributed to the wireless edge
Realize low latency Privacy New experiences Immediacy Edge cloud Customization To scale, Efficiency On-device intelligence must be distributed to the wireless edge, on the device Extract local value Reliability Billions of connected things Massive amount of data Security and privacy Content closer t
Slide 97: Combination of 5G and compute is essential for IoT
Combination of 5G and compute is essential for IoT Compute Security Connectivity Powerful processing On-device CPU / DSP / GPU, computer vision, audio, sensing,… 5G Artificial intelligence Efficient machine learning, on-device intelligence Gigabit LTE 5G NR LTE IoT Low power Edge services/cloud Mesh
Slide 98: Cloud services, local network, on-device
Cloud services Local network On-device Cloud analytics and virtualized Local gateway and/or local Various degree of core network functions core network functions on-devices capabilities Sensor fusion Sensors Security 15.4 5G Analytics Gateway W-HART Wi - Fi BLE Computer vision Wired Machines 5G Sens
Slide 99: AI for IoT across home, industrial/enterprise, and smart cities
Autonomous manufacturing Smart security for home Smart displays Smarter and robotics and enterprise and speakers agriculture More efficient use Home hubs and Sustainable cities Digitized logistics of energy and utilities smart appliances and infrastructure and retail IoT AI for IoT across the home,
Slide 100: IoT devices are already available
Security & monitoring Home security camera IoT devices are already Autonomous navigation for home vacuums Voice recognition and identification available
Slide 101: Power and thermal efficiency
The challenge Constrained of AI device workloads environment Compute Thermally efficient intensive for sleek designs Large, complicated models Requires long battery life for Complex all-day use concurrencies Real-time Storage/memory bandwidth limitations Always-on Power and thermal efficiency Critic
Slide 102: Heterogeneous computing key for on-device intelligence
Heterogeneous computing key Entire SoC is used for on-device intelligence Designed to deliver performance and efficiency improvements Broad portfolio of SoCs addresses different levels of performance and price points High-utilization Graphic not to scale Qualcomm Spectra is a product of Qualcomm Tec
Slide 103: Smart Speaker
Smart Speaker • Voice Recognition and ID • AI-based speech recognition • Far-field voice, beam-forming and echo cancellation • Support for cloud-based voice assistants • Wi-Fi, BLE connectivity Blocks Utilized Graphic not to scale Qualcomm Spectra is a product of Qualcomm Technologies, Inc. and/or i
Slide 104: Smart Display
Smart Display • Face/family recognition and audio sound recognition • HD Display with 4K encode/ decode, video telephony • Face detection object tracking • HW-based acceleration support for popular neural networks Blocks Utilized Graphic not to scale Qualcomm Spectra is a product of Qualcomm Technol
Slide 105: Enterprise IPC Camera
Enterprise IPC Camera • People and object recognition • Dual ISPs with advanced IQ – sHDR and TNR for low light • Upto 4K encode/decode @60fps HEVC video • Support for popular neural networks • Wi-Fi, Ethernet, Cellular (4G, 5G ) Connectivity Blocks Utilized Graphic not to scale Qualcomm Spectra is
Slide 106: Qualcomm Vision Intelligence Platform
Heterogenous compute Superb image with high Advanced connectivity performance and audio camera processor Qualcomm Vision Qualcomm® AI engine Hardware-based Intelligence for on-device AI security Platform Ecosystem Partners Qualcomm Artificial Intelligence Engine is a product of Qualcomm Technologies
Slide 107: Qualcomm Vision Intelligence Platform and Microsoft Azure IoT Edge
Qualcomm Vision Intelligence Platform Microsoft Azure IoT Edge http://www.visionaidevkit.com Run AI models on the edge with Qualcomm AI Engine or utilize the cloud Create, deploy and manage your models in the cloud and the edge with Azure ML and Azure IoT Edge
Slide 108: Qualcomm QCS400 series for smarter audio
Qualcomm ® QCS400 series for smarter audio High-performance Power optimized Tightly integrated Enhanced with AI Qualcomm QCS400, Qualcomm Adreno are products of Qualcomm Technologies Inc., and/or its subsidiaries.
Slide 109: Transforming automotive with AI
April 9, 2019 @qualcomm San Francisco, CA Transforming automotive with AI Nakul Duggal SVP, Product Management – Automotive Qualcomm Technologies, Inc.
Slide 110: Strong asset: Market fit
Strong asset: Market fit Qualcomm’s unique assets enable accelerated innovation and end-to-end system integration in automotive 3G/4G LTE RF CPU GPU DSP Multimedia Power Software/HLOS Computer Machine 5G/C-V2X Management Vision Learning/AI Position Wi-Fi / Bluetooth Ethernet PLC Security EV Charging
Slide 111: World's leading automakers build with our solutions
World’s leading automakers build with our solutions #1 In telematics and Bluetooth for automotive Telematics A leader In premium next-gen infotainment design-wins for production vehicles starting 2019-2020 Infotainment 18 Automakers have selected the Qualcomm Snapdragon ® ™ Automotive Infotainment P
Slide 112: AI is transforming automotive and the transportation industry
Intelligent Smart Cockpit Transportation Premium Pedestrian sound detection AI-based Natural language RSU understanding Cloud Autonomous Traffic flow Connectivity Driver optimization Driving monitoring Computer vision Multimedia Integrated Precise surround view Face positioning recognition Audio Hig
Slide 113: Our vision of scaling AI for different tiers
Our vision of scaling AI for Automotive Autonomous Smart different tiers Cockpit Driving Transportation Scaling AI-based cockpit Scaling AI for different Scaling AI for smarter and markets platforms across all vehicle classes autonomy levels infrastructure and cloud services Within the power and the
Slide 114: Efficiently scaling AI for different use cases
Scaling AI for Optimized for different automotive automotive use cases use cases Very compute intensive Thermal efficiency Snapdragon For cost efficiency (e.g. cooling systems) Most advanced heterogenous compute Architecture for AI Large, complicated neural Power efficiency For longer EV battery ran
Slide 115: Transforming in-vehicle experiences with on-board AI
Transforming in-vehicle experiences with on-board AI
Slide 116: Digital Cockpit: scalable heterogeneous compute architecture
Digital Cockpit – Scalable heterogeneous compute architecture Mirror replacement display AR/HUD display Passenger display Center display Reconfigurable cluster display Rear-seat display Mirror replacement vision system Foundational technologies Artificial intelligence Heterogeneous computing Compute
Slide 117: New on-board AI-based cockpit use cases
Natural language Face Detection Edge NLU based understanding Fingerprint Recognition Driver Monitoring Voice Assistance System Voice/speech Object recognition classification Virtual Assistant Camera Perception Voice/noise Scene Edge First AI Agent cancellation understanding AR based Navigation Locat
Slide 118: Snapdragon automotive cockpit platforms, Gen 3
Qualcomm Snapdragon automotive cockpit platforms, Gen 3 Introduced at CES 2019, bringing new levels of computing and intelligence to every tier SA8195P Visual Immersive Smarter Contextual AI intuitive Highly integrated experiences audio navigation safety experiences Artificial intelligence Panoramic
Slide 119: Shaping smart transportation
Shaping smart transportation
Slide 120: Shaping the Connected Freeway
Upcoming construction alert! Edge AI C-V2X (I2V) Localization Shaping the AI-based RSU detects RSU send road world Vehicle determines its Connected Freeway lane reconfiguration due to construction model to vehicle with lane reconfiguration position and distance from road construction
Slide 121: Defining connected urban transport
RSU with Traffic hazard warning RSU with AI-based AI –based camera detects a hazard on AI-based camera the right lane and alert other cars on the camera road; via precise positioning other cars avoid the lane with the hazard Pedestrian alert Traffic light detects a pedestrian crossing the street and
Slide 122: Smarter transportation infrastructure
Smarter transportation infrastructure Requires next level of compute and intelligence for perception and sensor fusion AI-based RSU Localization Road world model Radar software stack Perception, sensor fusion Multiple Compute cameras Heterogenous computing for perception and sensor fusion Communicat
Slide 123: Paving the road to autonomous driving
Paving the road to autonomous driving
Slide 124: Our system approach: autonomy stack
Perceive Plan Act Connect Camera, radar sensors Behavior prediction Actuation control CV2X, localization in maps Behavior planning Drive-by-wire smooth maneuver extended horizon sensors Motion planning Low level sensor fusion Tele-operations Data analytics Smart transportation Simulator and tools Ou
Slide 125: Example of level 3 autonomy stack
Example of level 3 autonomy stack Traffic jam co-pilot (urban/highway) Highway co-pilot Applications Automatic emergency Lane change assist Adaptive cruise control Lane keep assist braking Sensor playback Camera Lidar Map Localization perception Perception Observations Technology Safety manager Visu
Slide 126: Applying advanced AI techniques to autonomous driving needs
• Detect and classify objects • Detection for lane in radar signal and get markers, drivable their distance space vehicles, etc. • Using combination of Radar Lidar • Annotation for CNN and LSTM Perception Perception Deep Learning • Detect and classify lane markers, drivable space, • Reinforcement le
Slide 127: Strong asset: Market fit
Strong asset: Market fit Qualcomm’s unique assets enable accelerated innovation and end-to-end system integration in automotive 3G/4G LTE RF CPU GPU DSP Multimedia Power Software/HLOS Computer Machine 5G/C-V2X Management Vision Learning/AI Position Wi-Fi / Bluetooth Ethernet PLC Security EV Charging
Slide 128: Leading research across the AI spectrum
April 9, 2019 @qualcomm_tech San Francisco, CA Leading research across the AI spectrum Rajesh Pankaj SVP, Engineering Qualcomm Technologies, Inc
Slide 129: Advancing research to make AI ubiquitous
Advancing research to make AI ubiquitous IoT Mobile Automotive Cloud Perception Reasoning Action Object detection, speech Scene understanding, language Reinforcement learning recognition, contextual fusion understanding, behavior prediction for decision making Power efficiency Personalization Effici
Slide 130: Leading research and development across the entire spectrum of AI
Leading research and development across the entire spectrum of AI Fundamental Applied research research Machine Deep Neural Graph and Deep learning G-CNN transfer network kernel learning for training learning compression optimization graphics tools Bayesian Deep Neural Compute CV DL Source combinato
Slide 131: Advancing AI research to increase power efficiency
Advancing AI research to increase power efficiency Trained neural network model New Inference input output data Compute operations • Vector and matrix manipulations • CPU, GPU, DSP, Move data between and AI acceleration memory and compute • Move pieces of input data and AI model from memory to compu
Slide 132: AI model optimization research for power efficiency
AI model optimization research for power efficiency Applying AI to optimize AI models through automated techniques Reduced time-to-market Compression Quantization Compilation and engineering cost Learning to prune model while Learning to reduce bit-precision Learning to compile AI models keeping des
Slide 133: Compression of AI model architectures
Tensor decomposition Channel pruning Compression of Decomposing a single layer into Removing channels from the network two or more efficient layers AI model Spatial SVD L2 filter magnitude and Bayesian techniques architectures Hardware aware compression Automated removal of insignificant/redundant e
Slide 134: Quantization for power efficiency
Models typically trained Inference at lower Quantization at high precision precision for power efficiency 32-bit 8-bit Automated reduction in 01010101 01010101 01010101 01010101 01010101 precision of weights and activations while Floating point 3452.3194 Integer 3452 maintaining accuracy Promising r
Slide 135: Compiler research for efficient hardware usage
Compiler Quantized and research for compressed AI model efficient hardware usage Reinforcement learning AI optimization agent 4x for automated HW Update model based on feedback compilation—as there are billions of potential configurations Feedback: Speedup Schedule improvement over execution compila
Slide 136: AI hardware acceleration research
AI hardware Memory Compute Compute CIM acceleration research Compute-in-memory Traditional computer architecture • Computations, like add and multiply, are • Compute and memory are separate and data done in memory has to be shuffled back and forth Example: compute-in- • Good for simple math operatio
Slide 137: Can we apply foundational mathematics of physics to deep learning?
•Can we apply foundational mathematics of physics, like quantum field theory, to deep learning?
Slide 138: We are advancing AI research to make AI power efficient
We are advancing AI research to make AI power efficient We are conducting leading research and development across the entire spectrum of AI We are creating AI platform innovations that are fundamental to scaling AI across the industry
Slide 139: Thank you
Thank you Follow us on: For more information, visit us at: www.qualcomm.com & www.qualcomm.com/blog Nothing in these materials is an offer to sell any of the Qualcomm is a trademark of Qualcomm Incorporated, registered in the United components or devices referenced herein. States and other countries