Artificial Intelligence Computing for Automotive Webcast for Xilinx Adapt: Automotive: Anywhere January 12Th, 2021

From Technologies to Markets

Artificial Intelligence Computing for Automotive Webcast for Xilinx Adapt: Automotive: Anywhere January 12th, 2021

• Scope

• From CPU to accelerators to platforms

• Levels of autonomy

• Forecasts

• Trends

• Ecosystem

• Conclusions

Not included in the report

Robotic cars Autonomous driving Data center computing Cloud computing Performance

Understand the impact of Artificial Centralized computing Intelligence on the computing Level 5 Level 4 hardware for automotive Advanced Driver- Assistance Systems Driver environment Infotainment Level 3 (ADAS) Multimedia computing Level 2

Edge computing Computing close to Gesture Speech sensor recognition recognition

The focus for the semiconductor industry is shifting from to General Applications + Neural Networks General workloads + Deep learning workloads • Integer operations + Floating operations • Tend to be sequential in nature + Tend to be parallel in nature

Parallelization is key Scalar Engines Platforms and Accelerators explaining why Few powerful cores that tackle computing Hundred of specialized cores working in this is so tasks sequentially parallel popular Only allocate a portion of transistors for Most transistors are devoted to floating floating point operations point operations

Source: Deep Learning: An Artificial Intelligence Revolution by ARK Investment

Artificial Intelligence Computing for Automotive | Webcast for Xilinx Adapt: Automotive: Anywhere | www.yole.fr | ©2020 4 GOING FURTHER WITH DEDICATED PLATFORMS AND ACCELERATORS From Applications processors to GPUs to accelerators to platforms

Scalar Processing • Processes an operation per instruction • CPUs run at clock speeds in the GHz range • Might take a long time to execute large matrix operations via a sequence of scalar operations

Vector Processing Graph • Same operation performed concurrently across a large number of processing at data elements at the same time, the heart of • Single Instruction Multiple Data (SIMD) neural • GPUs are effectively vector processors networks Matrix Processing • Runs many computational processes (vertices) • Calculates the effects these vertices on other points with which they interact via lines (i.e. edges) • Overall processing works on many vertices and points simultaneously • Low precision needed • Names: accelerators, neural engine, tensor processing unit (TPU), neural network processor (NNP), intelligence processing unit (IPU), vision processing unit (VPU), AI Processing Unit (AIPU)…

Artificial Intelligence Computing for Automotive | Webcast for Xilinx Adapt: Automotive: Anywhere | www.yole.fr | ©2020 5 AUTONOMOUS VEHICLES - THE DISRUPTION CASE Two distinctive paths for autonomous vehicles

Technology x Market Penetration Disruption ?

Improvement Autonomous of cars as we vehicles know Automated driving

Industrialization Electric car 2020 should see matures the first phase commercial implementation Robotic cars Below expectation of autonomous “cars” fulfilling needs vehicles in a new plane of consumption

Yole Développement Electronics New use cases © August 2015 Invades cars

1880 80 years 1960 40 years 2000 20 years 2020 10 years 2030 5 years 2035

Acceleration : The speed of technology change doubles every technology shift

ADAS vehicles Robotic vehicles

Artificial Intelligence Computing for Automotive | Webcast for Xilinx Adapt: Automotive: Anywhere | www.yole.fr | ©2020 6 AUTONOMOUS VEHICLES - THE ROBOTIC DISRUPTION CASE Two distinctive paths for autonomous vehicle

Where? Any speed Anywhere ADAS vehicle ? Autonomous driving Medium Levels became speed Is L3 relevant? marketing Anywhere New Exist? Tech giants definitions but entrants Startups we consider that 2020 they do not represent the reality. Limited distance The reality is whether it is Historical autonomous, players 2020 whether it is not Designated areas

Designated Robotic vehicle places Low 2015 Taxi Shuttle/bus 2015 speed Level 1-2 Level 2+ Level 2++ Level 3? Level 4-5 How? 2015 Personal car 2022 Robotic Mobility-as-a-Service Artificial Intelligence Computing for Automotive | Webcast for Xilinx Adapt: Automotive: Anywhere | www.yole.fr | ©2020 7 AUTOMOTIVE MARKET TREND –YOLE MARCH 2020

2014-2050 Light vehicle sales breakdown forecast by level of autonomy (in M units)

ADAS Level 0 ADAS Level 1-2 ADAS Level 2+ ADAS Level 2++ ADAS Level 3-4 ADAS Level 5 Robotic cars produced (Munits)

140 7

120 6 6 2 3 4 8 10 12 15 1 1 19 0 23 25 Covid-19 0 0 11 28 0 5 7 9 14 17 19 21 31 0 20 4 24 26 34 100 impact 10 2 28 5 0 10 34 0 9 12 14 17 19 0 0 6 8 21 23 26 37 0 10 50 28 5 0 0 13 20 3 31 40 0 13 14 34 80 9 11 14 36 39 4 15 16 18 20 22 24 39 42 27 29 44 10 31 42 21 34 453-4 45 18 19 22 25 12 37 47 60 26 10 40 47 3 41 45 10 32 2.8 33 43 44 2.6 36 38 2++43 2.4 40 24 40 43 42 2.3 ADAS ADAS light vehicles sales (Munits) 2.2 40 20 40 2+ 42 2.0 41 2 1.9 40 40 1.7 40 40 37 1.6 36 38 1.5 34 56 55 58 53 1-2 36 1.4 34 53 49 33 1.2 31 31 20 39 30 1.0 1.1 27 1 33 35 32 28 0.9 24 30 29 27 0.7 24 20 24 0.6 20 17 20 17 0.5 17 15 14 12 14 0.312 0.4 11 8 10 9 0 0.2 9 8 7 6 5 0.1 6 4 0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.1 4 3 3 2 2 2 2 2 1 1 1 14 03 03 02 02 0

By 2035 cars with autonomy level L3-4-5 will represent 25% of global production the rest will be L1-2-2+

Artificial Intelligence Computing for Automotive | Webcast for Xilinx Adapt: Automotive: Anywhere | www.yole.fr | ©2020 8 ARTIFICIAL INTELLIGENCE IN AUTOMOTIVE 2015-2025e Total AI computing hardware volume forecast

• Artificial Intelligence based 80 on images from cameras for automotive will generate a total volume of 70 67M(unit) chips for computing hardware in 2025 for ADAS vehicles 60 only. • Currently, we consider 50 that for one type of camera, one chip is used. 40 We can find multiple chips

in one car. Fusion of inputs Volumes inMunits from multiple types of 30 cameras is not considered, as fusion is mostly used for inputs of different types of 20 sensors for now. • Ultimately, platform 10 solutions will enable a variety of vision formats 0 CAGR19- 2015 2016 2017 2018 2019 2020e 2021e 2022e 2023e 2024e 2025e and sensor inputs to 25 enable multiple strategies. Total - - 0.09 0.26 0.43 0.38 4.26 10.08 25.27 42.82 67.19 132.4%

Artificial Intelligence Computing for Automotive | Webcast for Xilinx Adapt: Automotive: Anywhere | www.yole.fr | ©2020 9 FROM LEVEL 0 TO LEVEL 2++ ON ONE SIDE, ROBOTIC ON THE OTHER Inclusion of accelerators and multiplication of the number of chips

• Level 0 to Level 2+/2++ are differentiating mostly by improved functionalities such as Automatic Emergency Braking (AEB) and some new functionalities such as Traffic Jam Assist (TJA) or Lane Keeping Assist (LKA) • On the robotic side, full autonomy was first realized in closed area at low speed (<15miles/h) to open designated area and at medium speed (<30 miles/h)

Computing These improvements are realized thanks to the introduction of AI introduces AI and follows algorithms and its related hardware what has been seen in consumer Implementation of AI : following what have been done in consumer applications

Implementation in SoC > as a standalone Multiplication of the number of chip of accelerators > platfoms x2 x4 x8 computing chips

Artificial Intelligence Computing for Automotive | Webcast for Xilinx Adapt: Automotive: Anywhere | www.yole.fr | ©2020 10 FROM SENSOR SUITE TO COMPUTING SUITE FOR AUTONOMY Levels and functionalities AEB L1 AEB L2 AEB L2+ HP AP Level 1 Level 2 Level 2+/2++ Level 3/4 Level 5/Robotic

ACC LKA PA TJA DM

1 5 >10 Technology penetration Radar Fusion Fusion

Forward/Rear Cam. Fusion

Surround Cam.

LiDAR

Computer Vision algorithms Deep Learning algorithms Increasing number of neural networks 2012 2016 by Tesla – 2021 for others 2030 2040 Performance and ASP

MCU FPGA Centralized platform Low VP High

Four main business models can be identified: - Car manufacturer : OEMs - Autonomous driving : Software and IP hardware:- they are developing the brains of cars and Autonomous driving all the driverless application Most of the - Car electrification : battery and net value will power train manufacturers:- they come from are turning cars into electric mobility as a powered systems service Car manufacturer Car electrification - Mobility as a service : service providers (robot taxis or shuttles):- they will offer transportation as a service

But the major value flow will go Mobility as a service to the service provider: by changing consumer use they will overtake the others

Artificial Intelligence Computing for Automotive | Webcast for Xilinx Adapt: Automotive: Anywhere | www.yole.fr | ©2020 12 SUMMARY Ecosystem • Because the technologies are different, ecosystems and supply chains for ADAS and robotic cars are different; • In both of these ecosystems, the supply chains are organizing;

• ADAS ecosystems are built around historical automotive OEMs, though with classical supply chains going less and less through Tier 1s.

Hardware for ADAS is led by MobilEye but OEM competition is tough

Tier 1

• Robotic vehicles ecosystems are built around full stack solution partnerships – In-vehicle edge platforms to infrastructure to cloud platforms as model to grow transportation as a service

Provide the silicon and the software stacks to

LEVEL OF INVESTMENT OEMs & AUTONOMY

Develop full solution by themselves

Develop their own autonomy stack by using the silicon provided

Use the full solution provided by computing hardware company ADDED VALUE

• Levels of autonomy should not be considered as incremental as the underlying technologies are evolving thereby requiring intermediate levels, currently Level 2+ and level 2++. • Level 3 is questionable as liability and regulations are not included. Many OEMs have, therefore, decided to skip Level 3 and are working directly on Level 4/5 vehicles. • ADAS vehicles and robotic vehicles are two completely different markets as ecosystems, technologies, ASP, objectives and business models are different. • The entrance of Artificial Intelligence, bringing with it their dedicated hardware, called accelerators, enables increased understanding of the environment around the car and can therefore allow more and more autonomy as the software and the hardware are improving. • For ADAS vehicles, AI has been around since 2016 in Tesla vehicles running on their own platform. However, as accelerators will be embedded in next generation SoCs from historical players and partners from OEMs from 2021,AI will spread quickly. • For robotic vehicles, AI has been in use for a few years and GPUs were used to process these algorithms. As fusion of multiple inputs is already in place, a combination of GPUs and accelerators in the same SoC or as two different chips will be used for vision acceleration. • To get to the performance needed for Level 4/5, vision accelerators may not be enough, vision and sensor fusion platform solutions as well as data fusion from infrastructure and cloud services will gain traction, especially in mobility as a service. AI platforms in domain control solutions will take share from independent ADAS application systems. Data fusion will become more prevalent to further autonomous evolution. • The application is key : focus should not initially be on the performance or power, but on the application/technology, so that the software (stack and algorithm) will make this application possible and the hardware can handle this software perfectly.

Tesla was first in introducing AI and is Matrix Accelerators are All along the value chain, NN software driving the AI hardware trend: moving replacing Vector processing for should be the focus for success. To be computer vision from AI on GPUs, to AI algorithms especially for successful, platform providers will also AI on accelerators. vision systems. engage in holistic AI solutions.