"AI Is Moving to the Edge—What’s the Impact on the Semiconductor Industry?," a Presentation from Yole Développement
•
2 likes•906 views
For the full video of this presentation, please visit: https://www.embedded-vision.com/platinum-members/embedded-vision-alliance/embedded-vision-training/videos/pages/may-2019-embedded-vision-summit-tschudi For more information about embedded vision, please visit: http://www.embedded-vision.com Yohann Tschudi, Technology and Market Analyst at Yole Développement, presents the "AI Is Moving to the Edge—What’s the Impact on the Semiconductor Industry?" tutorial at the May 2019 Embedded Vision Summit. Artificial intelligence is proliferating into numerous edge applications and disrupting numerous industries. Clearly this represents a huge opportunity for technology suppliers. But it can be difficult to discern exactly what form this opportunity will take. For example, will edge devices perform AI computation locally, or in the cloud? Will edge devices use separate chips for AI, or will AI processing engines be incorporated into the main processor SoCs already used in these devices? In this talk, Tschudi answers these questions by presenting and explaining his firm's market data and forecasts for AI processors in mobile phones, drones, smart home devices and personal robots. He explains why there is a strong trend towards executing AI computation at the edge, and quantifies the opportunity for separate processor chips and on-chip accelerators that address visual and audio AI tasks.
Recommended
Recommended
More Related Content
What's hot
What's hot (20)
Similar to "AI Is Moving to the Edge—What’s the Impact on the Semiconductor Industry?," a Presentation from Yole Développement
Similar to "AI Is Moving to the Edge—What’s the Impact on the Semiconductor Industry?," a Presentation from Yole Développement (20)
More from Edge AI and Vision Alliance
More from Edge AI and Vision Alliance (20)
Recently uploaded
Recently uploaded (20)
"AI Is Moving to the Edge—What’s the Impact on the Semiconductor Industry?," a Presentation from Yole Développement
- 1. © 2019 Yole Développement AI Is Moving to the Edge – What’s the Impact on the Semiconductor Industry? Yohann Tschudi Yole Développement May 2019
- 2. © 2019 Yole Développement On the road to augmented intelligence 2 Introduction Robot 2030 PC Mobile Smart Imaging 3D graphics cards 2006 DNN with GPU 2000 2010 Siri 2014 Alexa 2017 First SoC with NPU 1990 2020 Mobile Surveillance IP camera Audio Smart hearable Smartwatch integrating AI HW for audio AI Smartphone Virtual personal assistant Smart camera Next-gen assistant Smartphone integrating AI Smartwatch Companion robot Olfactometry Smell identification Human augmentation >2040 Motion sensing Holographic interaction Augmented human Robot home Personal Computer
- 3. © 2019 Yole Développement Edge computing for consumer applications 3 A 3D look at the market: • Stand-alone chip vs. embedded unit in system-on-chips • Imaging and audio • Systems: smartphones, drones, robotic homes Systems Audio Embedded unit in system-on-chip Stand-alone chip Imaging In imaging & audio there is clear a need for AI, and the technology is ready (or about to be) Other domains are not taken in account: • Technology not ready (olfactometry, taste recognition) • AI unnecessary for achieving the desired goal (motion sensing) Introduction
- 4. © 2019 Yole Développement Definition of the different trends 4 Introduction All computing parts are distinct and can even be mounted on a dedicated board • Graphic card for GPU • Sound card for DSP All computing parts are embedded and interconnected in a single chip • SoC, application processor (SoC for smartphones) Specialization A dedicated hardware architecture is needed to process/compute a specific type of algorithm. This architecture can be embedded in an SoC or integrated as a stand-alone chip • NPU in SoC • Sound processor ARCHITECTURE LOCATION Cloud Cloud computing - computation is done on the server’s computing hardware Edge Edge computing - computation is done on the device’s computing hardware
- 5. © 2019 Yole Développement Imaging – Dedicated hardware for AI 5 Introduction Vision processing unit (VPU) Unit embedded within the system’s application processor (AP) Vision processor (VP) Specialized stand-alone chip AP enables integrating various functions, including imaging, into a single chip Imaging embedded in AP is already in smartphone applications Main consumer applications are biometry, control (drones), AR/VR, and photography New VP applications are for smart homes mostly (surveillance, IoT?) Understand the environment around the device This unit type was originally created to compute computer vision algorithms (Mobileye eyeQ)
- 6. © 2019 Yole Développement Smartphone 6
- 7. © 2019 Yole Développement AI-dedicated unit • Processing AI applications with a dedicated unit is faster and consumes less energy • No dependence on internet connection: run applications anywhere, anytime • Improvement of privacy: not all data used for inferences is sent to the cloud • Possible to use personal data to improve applications using AI algorithms • Less latency for critical applications like authentication 7 Market trends – Smartphone Smartphones hardware must manage: AI’s huge requirements • High computational need • Real-time • Always-on • Huge neural network Mobile environment constraints • Thermal efficiency • Low consumption for long battery life • Memory limitations And…
- 8. © 2019 Yole Développement Apple Application Processor evolution 8 Market trends – Smartphone Annotated A4 die photo (source: MuAnalysis) Annotated A6 die photo (source: Chipworks) Annotated A12 die photo (source: TechInsights) Adding more and more elements inside the same chip, and introducing specialized computing units 2010 2012 2018
- 9. © 2019 Yole Développement AI for smartphone – Supply chain 9 Market trends – Smartphone
- 10. © 2019 Yole Développement 721 928 995 1075 1110 1140 1096 1086 1121 1146 1222 233 307 376 406 470 513 602 653 673 693 682 63 167 262 407 591 825 1115 0 200 400 600 800 1000 1200 1400 1600 1800 2000 2013-2023 AP volume shipments by manufacturer type and AP with AI volume shipments Fabless OEM AP with AI OEMs or fabless? 10 Market trends – Smartphone • Because AI will be a key technology, we expect fabless to ship less AP in the next three years - particularly Mediatek, which seem to be a bit behind in the AI race; • Huawei is gaining share and building its own chip through HiSilicon. This move explains the slight tilt in market share, from fabless to OEMs.
- 11. © 2019 Yole Développement 2017 & 2018e market share, AI application processors, volume shipments 11 Market trends – Smartphone • Huawei smartphones equipped with a Kirin 970 or 980 (late-2018) are gaining market share because the technology is equivalent to Apple’s, but much more accessible price-wise; • More players are proposing an AI application processor solution in 2018, explaining the YoY of 165%. Apple 59 94% Huawei 3 5% Qualcomm 1 1% 2017 Apple 115.3 69% Huawei 25.0 15% Samsung 7.6 5% Qualcomm 12.2 7% Mediatek 7.0 4% 2018e 167 Munits 63 Munits
- 12. © 2019 Yole Développement 2017 2018e 2019e 2020e 2021e 2022e 2023e Volume 63 167 262 407 591 825 1115 NPU Penetration rate 4% 10% 15% 22% 30% 39% 50% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0 200 400 600 800 1000 1200 Volume(Munits) 2017 - 2023 AI application processor - volume shipment (in Munits) and penetration rate AI application processors – 5y forecast 12 Market trends – Smartphone • Because AI will be a must-have for multiple reasons (marketing, authentication, privacy, etc.), dedicated hardware is needed • We expect a 46% CAGR for this type of hardware over the five next years Clear risk for some competitors to be kicked out of the AP market by not integrating AI, following Apple & Huawei → One way to catch-up is to focus on audio AI, which is not ready yet, and being the first to provide it
- 13. © 2019 Yole Développement Drone 13
- 14. © 2019 Yole Développement Segmentation of the market Market trends – Drone
- 15. © 2019 Yole Développement Consumer drones landscape 15 Market trends – Drone
- 16. © 2019 Yole Développement SoC hardware for drones players integrating AI 16 Market trends – Drone
- 17. © 2019 Yole Développement Supply chain 17 Market trends – Drone AI hardware AI software OEM Image sensor Possibility to add personalized software something on top
- 18. © 2019 Yole Développement Why use AI in drone applications ? 18 Market trends – Drone • AI is key to improving autonomous capabilities and human comprehension • Three core AI-based technologies can be used to improve drone capabilities: • With object detection and obstacle recognition, the drone identifies obstacles, avoids them, and anticipates user movements • With gesture recognition, the drone can understand the user without needing a smartphone interface • With user authentication and tracking (face authentication), the drone identifies its user, films him, and follows him • The global trend is to reduce the smartphone’s role as the interface between user and drone Follow user Obstacle avoidance Palm control
- 19. © 2019 Yole Développement 2018e 2019e 2020e 2021e 2022e 2023e Total high end drones with AI hardware 0.9 1.3 1.9 2.7 3.3 3.8 Total high end drones 1.4 1.8 2.4 3.1 3.6 3.9 Total penetration rate 65.2% 70.6% 77.2% 87.1% 92.2% 97.7% 0.0% 20.0% 40.0% 60.0% 80.0% 100.0% 0 1 1 2 2 3 3 4 4 5 VolumeShipment(Munits) High-end drones market forecast and AI penetration rate 19 Market trends – Drone This forecast include $500 ASP & $1200 ASP drones. • Gesture recognition and understanding the environment (object recognition) are key applications that will explain the appearance of AI hardware; • DJI is a driver in the drone market and owns 70%. Other players must follow innovations like 3DR, or die;
- 20. © 2019 Yole Développement Robot home 20
- 21. © 2019 Yole Développement The robot home • Home automation, or ”domotics”, is the discipline of building automation for a home: called a ”smart home” or ”smart house” • However, the term “smart home” is misused: it is more about a connected home than a smart home • Connected to the internet, devices are part of the Internet of Things • We define “robot home” as a smart home where devices are driven by AI technologies • These technologies include deep learning for recognition and natural language processing • For the robot home, latency and mobility are less important. However, privacy is key - and because of the continuous presence of intelligent systems in our home, a transition from cloud to edge computing, system by system, will happen. Market trends – Robot home Imaging Audio
- 22. © 2019 Yole Développement 2016 2017 2018e 2019e 2020e 2021e 2022e 2023e CAGR Home IoT devices - - 0.3 2.1 7.4 19.3 32.5 60.2 189% Personal Robotics - - - - 1.3 2.9 6.2 11.9 110% Total Vision Processor - - 0.3 2.1 8.7 22.2 38.7 72.1 - 10 20 30 40 50 60 70 80 Volume(Munit) 2017-2023 Smart home systems - Vision Processor volume shipment (in Munits) Imaging AI hardware – 5y forecasts 22 Market trends – Robot home • We imagine that half of IoT devices for smart home that have a camera, computing will be based on a vision processor and the other half based on an embedded VPU in a SoC; • For personal robotics though, because it needs a powerful computing hardware similar to drones, we expect that only a vision processor will be viable for the next 5 years.
- 23. © 2019 Yole Développement On the road to robot home 23 Market trends – Robot home Security Companion robots Connectivity Home help Home comfort Robot home
- 24. © 2019 Yole Développement Conclusion 24
- 25. © 2019 Yole Développement • We expect two different behaviors for imaging and audio: while audio AI will probably run on a specialized stand- alone chip, imaging AI will be embedded • All approaches will have a place in many applications. Embedded and stand-alone specialized chips will have almost equivalent revenue by 2023. Imaging and audio AI computing hardware 25 Conclusion Sound Processor 41% Embedded Sound Processing Unit 6% Vision Processor 15% Embedded Vision Processing Unit 38% 2023e Sound Processor 46% Embedded Sound Processing Unit 4% Vision Processor 7% Embedded Vision Processing Unit 43% 2020e Vision Processor 4% Embedded Vision Processing Unit 96% 2017 CAGR17-20 96% CAGR17-20 501% CAGR20-23 31% CAGR20-23 36% $635 $9,722 CAGR17-20 148% $23,305 CAGR20-23 34% $611 $24 $4,581 $5,141 $10,243 $13,062 Total embedded in AP IP revenues/Added-value of the unit Total specialized chip 2019 - 2020 Audio AI in smartphone
- 26. © 2019 Yole Développement AP designers and manufacturers Without edge computing HW design With edge computing HW design Close to sensor IP companies Value chain 26 Conclusion • Today, most of the value coming from data is captured by GAFAM & BATX; • They are even designing their own hardware to get deeper into the value chain; • There is tough competition on the edge, between computing close to sensor and centralized in AP.
- 27. © 2019 Yole Développement Resources 27 Market reports From image processing to deep learning 2019 https://www.generic.org Hardware & software for AI – Consumer focus https://www.i- micronews.com/produit/hardware-and- software-for-ai-2018-consumer-focus/ From image processing to deep learning, introduction to hardware & software https://www.i-micronews.com/produit/from- image-processing-to-deep-learning- introduction-to-hardware-and-software/ Blogs & websites http://image-sensors-world.blogspot.com/ https://www.i-micronews.com/ https://www.embedded-vision.com/ https://www.visiononline.org/ https://www.mac4ever.com/ https://www.lembarque.com/