ADAS: The Near-Term Opportunity for Components
SEE LAST PAGE OF THIS REPORT Paul Sagawa / Tejas Raut Dessai
FOR IMPORTANT DISCLOSURES 203.901.1633/.901.1634
psagawa@ / email@example.com
March 25, 2019
ADAS: The Near-Term Opportunity for Components
We believe demand for Advanced Driver Assistance (ADAS) components will exceed near-near term expectations, but peak within a decade, as autonomous robo-taxi services effect the private vehicle market. In the next several years, ADAS functions enhancing safety (e.g. collision avoidance, automatic breaking, enhanced visibility, driver alertness monitors, etc.) and convenience (e.g. “autopilot”, “traffic jam assist”, automatic parking, etc.) will be an increasingly important differentiator for car makers, as smartphone-linked infotainment systems become more generic. The key components for ADAS –processors (GPUs, FPGAs, and ASICs), sensors (e.g. cameras, RADAR, ultrasonic, etc.), radios (5G, vehicle-to-vehicle, etc.) and even memory- will see upside as the capabilities and penetration of ADAS functions expands. However, the incremental ADAS approach faces major obstacles on the way to full autonomy – accessing hyperdetailed digital maps, adapting sensor-fusion solutions to include LiDAR, addressing human/computer control transitions, navigating government approval and oversight, establishing liability/insurance norms, etc. In this context, we expect robo-taxi services to be the primary agent of fully autonomous travel and that adoption will begin to have noticeable effect on US new car demand before 2030.
- ADAS should outgrow near-term expectations. The $4B global market for ADAS components is an early manifestation on what we believe will be a multi-decade paradigm shift to self-driving vehicle technology. ADAS use sensor data to augment human driving, offloading certain functions, operating the vehicle under certain conditions, and overlying safety backstops to limit the likelihood and severity of accidents. These systems will be an increasing point of differentiation for automakers, as consumer reliance on smartphones and the cloud commoditizes infotainment. As such, we believe market growth will likely outpace market expectations for 16% annual growth in the next 7 years.
- Processors, sensors and radios. ADAS use various sensors, such as cameras, RADAR, LiDAR, ultrasonic, and others, to generate data on the environment around and inside the car. In real time, AI infused programs analyze the data on high-powered processors to provide guidance for the car and driver, and with time, to other vehicles on the road. This functionality is bundled into specific features, some for driver convenience (like TSLA’s “Autopilot”, Audi’s “Traffic Jam Assist”, and self-parking systems) and others for safety (e.g. blind-spot warnings, automatic hazard breaking, driver alertness indicators, etc.). The primary demand is for specialized processors (i.e. GPUs, FPGAs, ASICs), sensors (cameras, RADAR, LiDAR, etc.), and, in time, radios for vehicle-to-vehicle communications (5G, V2V standard, etc.). Memory will benefit, but automotive will remain a small piece of the total market.
- Full autonomy for private vehicles will take many years. GOOGL’s Waymo is already offering commercial self-driving robo-cab services, but private vehicles have major obstacles to surmount. First, robo-cabs rely on EXTREMELY detailed digital maps limited to geofenced service areas. These are very expensive to create, tightly limited, and proprietary. Second, LiDAR is essential for full autonomy in all conditions, but car makers have rejected it in private cars for its cost and bulk. Adding it later, once it is cheaper and smaller, will require expensive and time-consuming retraining of the whole system. Third, robo-cabs are designed to be entirely driverless while private cars must engineer for hot transfers of control between human and AI drivers, adding a major point of complexity and risk. Moreover, robo-cab leaders have more AI talent, more experience and much more testing data than private vehicle focused initiatives and are further along in addressing local approvals and regulatory oversight.
- The private vehicle market faces big long-term challenges. We believe self-driving robo-cabs can reduce the cost of ride hailing services by 65% with time – eliminating driver costs, increasing utilization, gaining efficiencies for network costs, etc. – delivering safer and more satisfying rides for consumers in all but the most sparsely populated markets. 63% of US car trips are within 5 miles of home, and the average US worker spends 2.5 hours per week commuting. US ride hailing services have added over 6B miles driven in the past 6 years in just 9 top cities, and this will accelerate as robo-cab service becomes ever more available. Currently, US households own an average of 1.97 cars, and, with exception of the 2008-09 recession, the installed base has been growing for many decades. We believe that we are nearing an inflection point.
- The trajectory for ADAS will crest within a decade. The growth for ADAS components will be strong over the next several years, as automakers add more complex systems to an increasing proportion of their vehicles. However, this will inevitably slow as the functionality becomes more widespread. Meanwhile, we believe ride hailing, particularly autonomous robo-cab service, will have meaningful effect on US consumer new car purchases before 2030. Combined with saturation and price pressure, we believe this will coincide with an inflection point for ADAS components.
- Upside for NVDA, XLNX, QCOM, ON, STM. For specialty processor manufacturers NVDA, XLNX, QCOM, AMD and others, automotive represents less than 10% of their total sales. While we expect growth from the segment to top 20% CAGR over the next 5 years, strong datacenter demand will keep Auto to a low-teens share of revenue. Sensor vendors, like camera-chip makers TNX, ON and IIVI, RADAR suppliers Continental, Delphi, Autoliv and others, are more levered to the automotive opportunity. As noted, the radio opportunity is longer term, but should benefit QCOM, AVGO, and a few others. We do not believe that the opportunity for memory is enough to change the overall trajectory for that market.
There are two distinct branches of autonomous driving development. The first, which we will call “robo-cabs”, gets the headlines. GOOGL’s Waymo is well in the lead with its commercial mini-launch in Phoenix, while GM’s Cruise Automation, The Ford-backed initiative with Argo.AI, Uber’s star-crossed efforts, and a handful of startups race from behind to deliver fleets of driverless vehicles available to shuttle customers on demand. These vehicles are being designed from the foundation to operate without any need for a human driver, relying on extremely detailed digital maps of a strictly defined service area. With maps for a new territory and some local training to pick up on the idiosyncrasies of the geography and regional driving norms, service can be rolled out market by market.
The second branch focuses on privately owned cars. Robo-cabs rely on those proprietary maps and LiDAR sensors that add thousands of dollars of cost and an obtrusive bump atop the car. In contrast, car makers don’t restrict their sales to narrowly defined geographies and they don’t have access to those special maps. Their customers are sensitive to the added cost of LiDAR and find the unseemly bulge distinctly unattractive, so the cars for sale don’t use it. Even though Elon Musk keeps trying to claim it, full autonomy on this track is out of the question, so the auto industry is taking it one step at a time. Cameras and RADAR don’t give as full a picture without LiDAR, but they are enough to support useful functions that improve safety, such as lane departure warnings, automatic hazard braking, and driver alertness monitors. These Advanced Driver Assistance Systems (ADAS) also include convenience features, such as automatic parking, Audi’s “traffic jam assist”, and TSLA’s infamous “Autopilot”.
With consumers now expecting the automobile infotainment system to synch up with their smartphones and be done with it, car makers are turning to ADAS as key differentiators. TSLA’s image as the tech leader was burnished by the early buzz on Autopilot (although the bad pub from highly publicized crashes had to have hurt). GM’s Cadillac, Mercedes, Audi, Nissan and others are currently running TV ads with ADAS front and center. Expect this trend to keep picking up steam. We think current 15-20% annual growth market estimates for the specialized processors, sensors and radio components needed for ADAS are conservative for the next few years – penetration will be higher, and the system complexity will increase.
However, by 2030, we believe that enough households will be using robo-cab services for daily commutes and local errands that new car sales in the US and other forward markets will be in decline. While the fully autonomous fleets will certainly have high end automation solutions with processors, sensors, radios and memory, the very high utilization of these vehicles means that the volumes are very unlikely to make up for a shrinking private vehicle market. Thus, we expect the ADAS component market to crest within the decade.
Still, that leaves time for companies like NVDA, XLNX, QCOM, ON, STM, and IIVI to make hay while the sun shines. Expect the automotive segment for these and other companies to surprise to the upside, alongside likely concurrent strength in hyperscale datacenter investment and 5G build-outs. Just don’t expect it to last forever.
Exh 1: SSR Summary of Two Roads to Self-Driving System Development
Two Paths Diverged in a Wood
There are two distinct approaches to autonomous driving (Exhibit 1). The Disruptors are the ones in the news. Google’s Waymo subsidiary began working on a fully autonomous car more than a decade ago, recruiting the winners of DARPA sponsored self-driving competitions and building on its established strength in AI and simulation. From the start, the goal was a vehicle that could safely and efficiently take passengers as directed without a human driver at all. Millions of road miles and billions of simulation miles later, Waymo cars have operated without drivers in Phoenix Arizona, and even though the limited commercial launch underway has emergency safety drivers back on board, its clear that Google can pull them as soon as it wants to scale the business. GM’s Cruise Automation business is next in line, hoping to launch its commercial service in San Francisco before year end (but probably not until at least 2020, given known technical issues). Ford’s project with Argo.AI is beginning to get notice for its progress, Baidu is trying to build a robo-cab ecosystem around its open-sourced self-driving software and proprietary maps in China, and a handful of well-funded start-ups are in extreme catch up mode with feverish development and testing underway.
These initiatives have several elements in common (Exhibit 2). First, they rely on extremely detailed digital maps that note precise locations, dimensions, and relevance of all objects – signs, bumps, unmarked pedestrian crossings, building entrances, ramps, overhanging trees, blind driveways, rough pavement, construction sites, frequent double parking, loading zones, blah, blah, blah. If it is noted in the map, the self-driving system doesn’t have to notice and interpret it on the fly, simplifying the task of driving. Second, all the robo-cab developers use LiDAR as part of their sensor suite. LiDAR systems send laser signals out in all directions, sensing the reflections returned to the vehicle to determine the size, location, shape and trajectory of objects in the environment. LiDAR is MUCH more precise than image sensing cameras (or even RADAR) and not limited by the available light. The downside is that LiDAR costs several thousand dollars and, until very recently, required a turret mounted to the roof of the vehicle. Third, robo-cab developers have focused on addressing urban and suburban driving conditions – where the demand for ride-hailing is perceived highest. Finally, because the plan is to be fully autonomous, the robo-cab developers don’t spend a lot of resources working on transitioning control between the AI and a human driver.
Exh 2: Primary Elements of Complete Self-Driving Technology
The Incrementalists path is increasingly well traveled as well. Unlike the robo-taxi pioneers, most automakers are primarily interested in selling more cars to consumers. A self-driving capability could be a boon to their customers, offloading burdensome tasks and strengthening safety while still allowing consumers the joy of driving and personal car ownership. Full autonomy requires a 360-degree 3D real time map marking the precise location, speed and trajectory of all objects in the environment, a predictive engine that interprets the map and delivers an optimized approach for conducting the vehicle, and autonomous control over the car to execute split-second decisions to safely complete the trip under all driving conditions. The incremental approach is willing to compromise well short of that, using incomplete and imprecise mapping, limited predictive scope and control under tightly circumscribed conditions to offer narrow convenience and safety features for drivers. Some of these are now commonplace – lane departure warnings, blind-spot warnings, parking proximity indicators. Adaptive Cruise Control (ACC) – intended to keep the car in its lane at appropriate distance from other vehicles, stopping the car if necessary – is offered by most luxury car brands, some with better reviews than others, and seems poised to move into mass
Exh 3: Incremental Automation approach taken by Traditional Automakers faces exponential AI, 3D-Mapping and Data challenges
market vehicles (Exhibit 3). The best reviewed ACC solutions – generally Tesla’s Autopilot, Cadillac’s SuperCruise, Audi’s Traffic Jam Assist, Volvo’s Pilot Assist and Mercedes’s prosaically named Adaptive Cruise Control – are front and center in vehicle advertising. Future versions may handle lane changes or learn repetitive routes.
Pay No Attention to Elon Musk
Tesla has been promising fully automated driving for a couple of years, with CEO Musk periodically taking to Twitter to pronounce new target delivery dates – always just around the corner. However, scientists who have been working in the field for years have a nearly unanimous skepticism that Tesla’s technical approach can possibly achieve full autonomy at all, much within the few months touted by Musk.
There several big obstacles between Adaptive Cruise Control and full autonomy. First, because LiDAR sensors have cost thousands of dollars and required an unsightly bulge atop the vehicle, automakers chose not to use it. While undoubtedly the right choice, given consumer preferences for lower prices and cars without giant turrets on the roof, the lack of LiDAR in the sensor suite severely compromises the ability of ACC systems to accurately identify the location and speed of objects in the environment. (RADAR has very
Exh 4: Disruptor systems (LiDAR + Radar + Optical) are the most comprehensive
low resolution and can be troublingly inaccurate, particular in marking “soft” objects, like pedestrians) Elon Musk argues that much faster processing will allow on-board AIs to infer distance from the inherently 2D image data provided by cameras, but such firepower is FAR from realistic today, and since camera sensors require ambient light, doesn’t solve the problem for nighttime driving. As such, EVERY company pursuing the full autonomy robo-cab path is using LiDAR (Exhibit 4).
Recently, LiDAR units have come down in price dramatically – from $70K to less than $5K – and the form factor has been streamlined to eliminate the need for the rooftop turret (Exhibit 5). However, it is not so easy to simply add LiDAR to the ADAS systems sold with new cars. Self-driving features rely on an internally generated real-time 3D picture of the environment surrounding the car. That picture is created by an AI that fuses the inputs of multiple sensors – for ADAS, cameras, RADAR, and ultrasonic proximity sensors. Adding LiDAR to the mix requires taking giant steps backward to re-engineer the way the AI creates the 3D map. The training data must include the LiDAR input, so all previous data is likely unusable. We note that Tesla likes to tout its “millions of miles” of driving data. This dataset is very shallow – mostly just telemetry (location, speed, etc.) with forward facing video collected for “unusual” circumstances – and it is NOT enough to train an adequate 3D real-time mapping system for full autonomy.
Exh 5: Highlight of LiDAR System Types and Target Costs
Second, creating the real-time 3D map is considerably easier when all the static elements have been previously identified. Would-be robo-cab operators invest heavily in maps with MUCH more detail than is available in the digital maps available on Google Maps or Waze. They note bumps in the road. They know the content of every sign. They warn of locations where pedestrians often jaywalk. They mark the height of every curb and the location of every building entrance. They are aware of overhanging trees and blind driveways. Building these maps is a huge investment and the companies making it focus on the specific markets where they intend to launch service and keep their data jealously proprietary. None of it is, or will be, available to private vehicles without substantial licensing fees, nor is likely to cover territory outside of closely defined service areas for a long time.
Third, unlike robo-cabs, vehicles intended for private ownership must facilitate the transfer of control between a human driver and the car’s self-driving system. This is a very difficult engineering challenge with enormous consequences for failure. Already, fatalities have resulted from drivers misusing ACC systems and failing to take control of the car as necessary. As car makers look to expand the situations under which a driver may cede control to a self-driving system, the handoffs must be intuitive and certain while the mechanism for the car to compel the driver to retake control must eliminate the risk of catastrophe, bringing the vehicle to a safe stop off of the road if the driver is unable to comply. It also needs to accommodate drivers who decide to take the wheel on the fly, perhaps with inadequate attention to possible hazards. Car makers, regulators and insurers must all reach agreement on permissions and liability, a process that could be messy.
Driver’s Little Helpers
It will be a decade or more before automakers will be able to offer a car that can drive its new owner home from the dealer lot. However, this doesn’t mean that the incremental ADAS features that will be available won’t be valued by car buyers. In a crowded market, vendors compete on multiple fronts – how the car looks, how it performs on the road, safety, amenities, etc. A decade ago, infotainment – systems to play music, facilitate hands-free telephone calls, provide navigation, control the in-car environment, etc. – was front and center in car ads. Today, not so much, as the ubiquity of smartphones, cloud services and of connectivity standards have consumers asking for carmakers to step aside in favor of their personal devices.
ADAS is filling that differentiation gap. TV ads tout ADAS safety features like automatic hazard breaking, blind-spot monitoring, lane departure warnings and collision avoidance. Convenience systems – automatic parking, adaptive cruise control that can drive under closely defined situations (e.g. keeping up to speed in one lane on the highway or following safely during stop-and-go traffic jams), heads up AR navigation displays, etc. – can automate mundane tasks or provide richer information for decisions.
Tesla was early to hype its ACC system, dubbed “Autopilot”. CEO Musk touted Autopilot as on the verge of full self-driving capabilities, spurring a few overly confident customers to crash their Teslas while sleeping or watching videos. More cautious carmakers have implemented similar tech but included systems to assess driver alertness, sensing if hands are off the wheel or eyes off the road for more than a moment. These monitors are another ADAS feature. Eventually, these systems will get a lot better, and the conditions under which they might be safely used will expand – perhaps future cars can be trained to operate on fixed routes, like a repetitive commute to work, or at slow speeds under emergencies (i.e. “drive me to a hospital” with hazard flashers on).
The Component Opportunity
Currently, about 76% of passenger cars sold in the US and 32% world-wide include some level of ADAS features. This is up by 6550bp over the past 5 years, and we expect the penetration to continue to grow at a strong pace. We also see that the semiconductor content of a passenger car has grown 45% from $310 to $450 over the same timeframe (Exhibit 6). In this context we expect the opportunity for ADAS components to grow somewhat faster than the rough 16% consensus CAGR over the next 7 years.
According to McKinsey, more than 35% of ADAS semiconductor spending will be on processors (Exhibit 7). Some of this will be on simple microcontrollers, but the most growth will come from more powerful GPUs, FPGAs and ASICs used to process the AI inference models needed to generate the real time 3D map and to make decisions from the data. Nvidia put up 19% YoY growth in its automotive segment in 4QFY19 (Exhibit 8). Xilinx projects an 18% CAGR from its automotive segment through 2023. We believe the pace will pick up for both and that current projections will prove conservative. Qualcomm and AMD are also positioned with appropriate products, and Intel’s purchase of camera-based ACC subsystem vendor MobileEye could position it to take a bigger role.
Exh 6: Average Semiconductor Component Revenue Per Vehicle, 2005 – 2025E
Sensors are the second main category for ADAS components, making up nearly 30% of the opportunity. The growth will come from optical chips needed for cameras and LiDAR (Exhibit 9). Today, this is mostly photo imaging (only Audi’s “Traffic Jam Assist” ACC for its most expensive model uses LiDAR) and RADAR. While RADAR components are almost universally employed for proximity alert applications, such as automatic emergency breaking, its importance to future systems is likely to diminish as cost and bulk reductions drive adoption of LiDAR for ADAS. The top automotive image sensor suppliers are ON Semiconductors, OmniVision, STM Micro, Sony and Samsung, with total sales to the sector at ~$500M and growing at a 15.7% pace. The LiDAR component market – composed of laser diodes and receivers – is diffuse with many small private suppliers vying to take leadership. At the subsystem level, privately held Velodyne has the most traction, with Valeo having scored a coup supplying a low-profile mechanical LiDAR system for Audi’s ACC solution – the only private vehicle equipped with the technology. Google’s Waymo recently announced that its internally developed leading edge LiDAR solution would be made available for non-automotive use cases (robots and other automation) (Exhibit 10).
Exh 7: Share of Total ADAS Semiconductor by Category in Passenger Cars, 2025E
Exh 8: NVDA and XLNX Automotive Segment Revenue, 1FQ15 – 4FQ18
Exh 9: ADAS Systems Revenue Distribution by Functional Type, 2015 – 2025E
Exh 10: Top Automotive Semiconductor and Sensor Vendors by Category
Radio frequency chips will get more complex to support autonomous driving use cases and the rise of new communications standards (e.g. 5G, Vehicle-to-Vehicle, etc.). Given the volumes relative to the annual sale of personal devices, we see automotive having a small impact on the addressable market. The same is true for an even larger extent for memory (Exhibit 11, 12).
Exh 11: Financial Summary of leading ADAS Suppliers across categories
All Good Things Must Come to an End
ADAS component growth is predicated on three factors –increasing penetration of ADAS functionality, increasing semiconductor spending per vehicle as systems grow more complex, and a growing new car market (Exhibit 13, 14). There is risk to all three. First, we are in the step part of the penetration curve – over the next 5 years, the cars sold with any ADAS feature is projected to rise from 40% to over 60% of total and vehicles equipped with ACC is expected to rise from about 12% to more than 25%. This will slow down a bit thereafter. Second, in the coming decade, we expect most ADAS features to become much more competitive, with price pressures as OEMs push back on systems suppliers, and so on down to the component vendors. Moreover, ADAS functions sold today as discrete systems – e.g. blind spot monitors, lane departure warnings, and parking assist – will be integrated to more generalized multi-function systems, reducing costs for OEMs. Deflation will hit top line numbers.
Finally, global automobile unit sales have grown almost uninterrupted for over a hundred years. Analysts remain confident that emerging market demand can continue the streak for a couple more decades. We are more circumspect. After years of strong growth, nearly 10% of Americans have used a ride hailing service within the last month, with both penetration and usage on trajectory to grow strongly into the next decade (Exhibit 15). The all-in cost of personal car ownership – loan/lease payments, insurance, gas, parking, etc. – is spread against very low utilization, as the typical US car travels less than 35 miles per day. A full-time ride-hail driver will drive more than 10 times that distance, while a robo-taxi with no need for breaks might
Exh 12: Global Automotive Semiconductor Market Size Forecast, 2017 – 2025E
Exh 13: Industry Forecast for ADAS specific Semiconductor Sales, 2017 – 2025E
Exh 14: SSR Forecast for ADAS specific Semi by Component Type, 2017 – 2025E
Exh 15: US Ride Hailing Services Sales Forecast and Penetration, 2017 – 2030E
Exh 16: SSR Forecast for long-term ADAS Semiconductor Revenue Outlook, 2017 – 2025E
double that again. In this context, substitution for secondary, or even primary, vehicles will be very tempting for many households.
Once the last technical sticking points – i.e. smooth unprotected left turns, precise pick-up and drop -off points, etc. – are fully resolved, robo-cab operators will race to gain the early mover advantage. We believe fully autonomous robo-cab service will be available in at least half of the top fifty metro markets in the US by 2025, and be nearly ubiquitous by 2030, enabling much lower price points for on-demand transportation and accelerating the substitution effect on new car demand. Moreover, very low-cost electric bike and scooter rentals represent a different form of competition for transportation, particularly in markets with aggressive congestion or emissions regulations for cars. Automobile analysts project steady 3%-ish annual unit growth for light passenger vehicles over the next five years, but beyond that, we see significant risk of an inflection point. If so, there will be an obvious effect on the demand for ADAS components, and taken with the other two factors, the market could turn down in the back half of the coming decade (Exhibit 16, 17).
Exh 17: Characteristic differences between Robocab and ADAS Car Operation
How to Play It
In the next few years, ACC subsystems built around camera image sensors and specialized processors will be in the sweet spot for ADAS. NVDA and XLNX likely have the most leverage for processors, with QCOM an aggressive entrant into the segment. For sensors, ON is the leader in the automotive camera segment, with privately held OmniVision the biggest challenger. Sony and Samsung are the leaders in overall image sensors, but their products are tuned more directly at the smartphone market. With the deployment of 5G and the establishment of standards for vehicle-to-vehicle communications, a bigger market for radio communications chips should also develop. Candidates for the market would include Qualcomm, Broadcom, Skyworks, and MediaTek (Exhibit 18).
Exh 18: SSR Summary for ADAS Component Players