Intel Corporation (INTC) Earnings Call Transcript & Summary

All right. Good day, everybody. My name is Ross Seymore. I'm the semiconductor analyst here at Deutsche Bank in the United States, and we're very pleased that you can join us for the 2020 Autotech Conference. For the next presentation, we're absolutely thrilled to have Professor Amnon Shashua from Mobileye. He's the Senior VP at Intel. He's also the President and CEO of Mobileye. Amnon is going to go through a number of slides first, maybe 20, 25 minutes of slides, and then we're going to go into Q&A after that, that I'll be moderating. If you have any questions, there's a Q&A box on your presentation screen. You can put a question in there, and we'll try to get to you. So with that, let me pass it over to Amnon to go through his presentation. Thanks, Amnon.

Thank you, Ross. Hello, everyone. I'll -- our first few slides kind of capturing our strategy from driving assist up to autonomous driving. And what is kind of unique to the way we look at this industry and how it evolves from driving assist to autonomous driving. So we'll start with that we have 3 pillars in our business. First is driving assist. This is the main bread and butter of today's revenue, going from Level 1, Level 2. We're building Level 2+ systems, we're working with Tier 1 suppliers and carmakers. We have technology that we announced back in 2015, where cars with our front-facing technology, front-facing camera technology, send specific data to the cloud. And there, we use all this data aggregated and build high-definition maps and other insights about the road structures, and we built a business around it. And the third is the full stack self-driving system. Now the Holy Grail is consumer autonomous vehicles, but we see robotaxis as a step along the way. So those are the 3 pillars. Just kind of a one slide that captures where is Mobileye. We sell our silicon with all the software integrated into it. To date, more than 65 million chips have been sold. It means more than 65 million cars growth with our technology. We have 48 running production programs. This year, year-to-date, we had 29 design wins. And if you look at the 5-star rated vehicles since 2018, we have 70% of them. The bar below shows all sorts of milestones starting from 2007 until today of industry-first launches now from the first traffic sign recognition, and the first pedestrian lane detection, the first camera-only ACC, the first camera-only FCW, the test auto pilot back in 2015 up to the SuperVision that I'm going to explain a bit in detail next year. So this is kind of Mobileye at a glance. In terms of the -- what we sell, we have kind of a food chain of different positions. The first is the silicon, silicon with our software. Second one is silicon plus the PCB, where it's kind of a subsystem. Then we're talking about the full stack self-driving system. It's the entire hardware powering a self-driving vehicle. Integrated into a vehicle, which gives us a self-driving vehicle, we have partnerships with some carmakers to allow us to integrate our SDS into their vehicle. And the further right is a customer-facing Mobility-as-a-Service. That is the reason that we acquired Moovit a few months ago. So now I'll spend 3 slides on what is our approach. And our approach is somewhat different from the garden-variety approach of looking at driving assist and autonomous driving. So we look at it as the camera subsystem is really the anchor and all the rest is added for a higher mean time between failure. Stating it differently, we see driving assist and autonomous driving living in the same space. The difference between them is not -- it's not capability, it is mean time between failure. The mean time between failure that you need for a Level 4 system is much, much higher than the mean time between failure you need for a Level 2 system where the driver is responsible. But still, a Level 2 system can have the same performance envelope as a Level 4 system, but with a lower mean time between failure. So if you look at the 2 blocks, driving assist on the left-hand side. We have algorithmic redundancies in order to increase the mean time between failure of the function. And when you look at autonomous driving, we need in addition also that redundancies due to different modalities like radars and LiDARs. So the way we look at it is we build 2 separate subsystems. One is a camera-only subsystem. So it's a full end-to-end autonomous driving capability. The second one is no cameras, just radars and LiDARs subsystem, again, end-to-end capability. Now the camera-only subsystem is then productized into the ADAS space. And that product is called what we call a SuperVision. Now what it will serve in the ADAS space is both for safety and comfort. In terms of safety, by extending the envelope of AEB, autonomous emergency braking, to surround sensing, we can provide a much, much higher level of collision avoidance. And then we can also provide comfort function like a Level 2+ system in which the car has hands-free capability just like a Level 4 system, but with a lower mean time between failure. This is why you would need a driver in the loop. So this is how we see the interplay between driving assist and autonomous driving. Where driving assist, at the ultimate point, is a productization of a subsystem of an autonomous car. When we look at the safety comfort paradigm or trade-offs, today, in terms of safety, driving assist provides AEB, autonomous emergency braking, and lane-keeping assist. And this translates into comfort functions like ACC, adaptive cruise control, and lane centric. What we see evolving into the future, if you have surround computer vision, you have 360-degree sensing of high-resolution cameras and the compute that comes with it, you can provide a high level of preventive maneuvers, what we call that Vision Zero. And in terms of the comfort, we can provide a hands-free driving experience, just like a Level 4 but with a lower MTBF because the driver is still in the loop. We need 2 additional building blocks to execute our strategy. One is the ability to drive everywhere, what we call geographic scalability. For that, we have our crowd-sourced technology for building high-definition maps. And this allows a car with our technology to drive everywhere and not only in geo-fenced areas. And the second one is standardizing the human judgment of what it means to drive carefully. This is a formal model of decision-making that we developed 3 years ago, which we call Responsibility Sensitive Safety, or RSS in short. So those are additional 2 building blocks in order to execute the strategy. So now productizing the camera-only subsystem, we call that SuperVision. It is based on 2 EyeQ5s. So EyeQ5 is our latest generation chip that is going into volume production 2021. The chip has been sampled already a year ago, and it's in development systems, but it's going to be in volume production next year. So 2 such chips. We're talking about 11 cameras, 8 megapixels each surrounding the car, 360 degree. We are talking about our high-definition maps. We crowdsourced high-definition as an integrated part of the system. Our RSS safety model is integrated into it. And also our driving policy. So driving policy is all the decision-making. So it's not only the perception, understanding the environment, but also making the decision and planning and control of the car. And this enabling on one hand a preventive maneuver for safety based on RSS. And then on the comfort function, it's hands-free from highway driving, highway -- point-to-point in terms of navigation, rural roads, urban roads. It's basically a Level 4 capability but at a lower mean time between failure, which necessitates a driver behind the steering wheel. And of course, it comes with over-the-air updates and also auto parking. So this is the productization of our camera on the subsystem. And just to show you a clip. What you see here is a drone view of this system. So it's powered only by cameras. And the car, circled in blue, is our AV car. And you see it's driving in really dense traffic. This is the heart of Jerusalem. And it needs to make complicated decisions, merging into traffic and there's obstacles. It needs to give way, take way. Pedestrians -- complicated encounters with pedestrians, unpredicted turns. So we are kind of speeding this just as you see the kind of the range of scenarios that this car is handling. And it's just based on camera perception. So think of something like this as an assist feature, this kind of capability as an assist feature in a driving assist system. And if you add then a layer of radars and LiDARs for redundancy, you get the Level 4 vehicle. Okay. And we announced about a month ago that the first OEM which the SuperVision is going to go into volume production is going to be Geely in China, and it's going to be launched around Q3 next year. So a year from now, this system is going to be productized. Now in terms of the technology for map building, as I mentioned before, it's based on crowdsource. So every vehicle with a front-facing camera having our technology underlying it, having our EyeQ chip there, packs certain key data, which is detection of landmarks, detection of lanes and then the tolls and reflectors. Sparse data for the scenes stacked into about 10-kilobytes per kilometer, and then sent to the cloud. In the cloud, we have algorithms that we have been working for years to refine and perfect, in which automatically high-definition maps are being built. And you see this picture in the center of what this means. It's very, very detailed map with all the driving -- drivable paths, all the stationary structures on the road, centimeter-level accuracy. And it's all built automatically. There is no manual intervention. Really, it's all automatically. And on the right-hand side, you can see how the system then uses this map and online sensing in order to localize itself inside the map at a centimeter-level accuracy. So we have today, 6 major global carmakers sending data. We're talking about today, about 7.5 billion kilometers of roads that have been collected globally around the world. 8 million of them covered daily. And we have, on a weekly basis, we're updating 50,000 kilometers of map data. This is -- the volume is building up. We project that in 2024, we'll be collecting 1 billion kilometers of roads every day and 1 million kilometers of map will be updated every day. So this is an example of -- Munich, as an example. We have test cars in Munich since July. Munich is mapped completely automatically. We have been covering thousands of hours of autonomous vehicle testing since July. Carmakers are also testing the car to understand the promise for driving assist or as a subsystem for a Level 4 design that would come later. The next building block is RSS. This is the formal safety layer. It's all about how do you protect yourself against lapse of judgment in decision making? Because when you look at humans, the critical source of error that leads to an accident with humans is lapse of judgment. And we don't want a robotic engine to have any lapse of judgment. It's a computer. There is no reason to have a lapse of judgment. But we need to define what are good judgments. What are careful judgments. And the way we handle this is through a formal theory. It basically has 2 components. On one hand, it specifies in a parametric way, what are the assumptions that people make when merging into traffic. And then putting parameters on those assumptions. Second is once you have those assumptions set, the theory takes the worst-case analysis. So rather than predicting human behavior, predicting what another road user would do, we take the worst case. And we also assume that the right-of-way is given and not taken. This gives us a clear definition of what it means to be in a dangerous situation. How to get out of a dangerous situation. And since the driver is a computer, the driver would never put himself into a dangerous situation because a formal definition is provided by this RSS methodology. RSS has been -- since 2017, when we published the paper in a transparent way with all the details, we have been working to evangelize it and standardize it because we think this is something that is important for the entire industry, not just for Mobileye. There is nothing in the theory that promotes Mobileye's technology. It's a theory of safety that is useful for everyone to adopt. We have, for example, Intel is chairing an IEEE workshop exactly on this topic. And this work group consists of 25 leading industry players, basically all the actors in the AV space are part of this group and it's chaired by Intel. And it's working along similar lines of the RSS methodology by looking at a formal definition of decision-making. We have in China, back in March this year, RSS -- the Chinese version of RSS was standardized. And we have all sorts of partnerships with industry players around our RSS. And we're working with regulatory bodies around the world in order to use RSS as a starting point for a regulatory discussion for providing regulatory certainty for autonomous driving. Lastly, the Mobility-as-a-Service effort that we have it is based on the following observation. That if you look at the world today it's -- the mobility world today, there are public transit operators like the MTA in New York, Keolis, Transdev, RATP in France and so forth are tens of thousands of PTOs around the world. And there are transportation network companies like Uber and Lyft, [ that do when they perform ] ride-hailing. Both of those mobility providers, they have all the incentives in going to autonomous. Autonomous PTOs and autonomous TNCs. And this provides a rich set of business models starting from providing an autonomous car. We call this vehicle as a service. Providing rides through a fleet, we call this ride as a service. And providing an end-to-end, including the customer-facing application, which is the Mobility-as-a-Service. And we are building all the layers to provide both the VaaS, the RaaS and the MaaS. And those layers are done through a cooperation with Moovit Company that we acquired a few months ago. And they're helping us to build the tele ops, the back end, the mobility intelligence, the front-facing applications around Mobility-as-a-Service. And then once we have everything vertically integrated, we can then have the flexibility to break it down to different business models and partnerships where we can license a part of the stack like the mobility intelligence or the customer facing or we can do it ourselves. And we have also partnerships and deployment plans. For example, in Israel. We have a joint venture with Volkswagen to deploy Mobility-as-a-Service. In Israel, starting from 2022, 100 vehicles in Tel Aviv. We are working closely with the Israeli government to promote regulatory certainty for the pilot of 100 vehicles. We have engagements in France with partnerships with the PTOs. We recently signed a deployment in Dubai. In Japan with the WILLER Group. In South Korea, Daegu City. So we'll gradually expand. But '22, '23, this looks like the first trial for the testing the technology of driverless cars by removing the driver from the wheel. I think this ends the 20-minute kind of bird's eye view of Mobileye, our approach, our strategy, what we are doing different than others. Thank you.

Perfect. Thank you so much, Amnon. That was a great overview. So let me hop into a number of questions. First, why don't we start at the high level. And you talked about L1, L2+, the design win traction that Mobileye has. I think the 70%-plus market share for the 5-star vehicles. I want to talk a little bit about the evolution of the technology. What do you see is the timing and the biggest hurdles of going from L1/2, where we are today, to full consumer autonomous vehicles? When do you think that happens? And what do you think are the biggest hurdles that need to be overcome to accomplish that goal?

I'd say that the biggest hurdle is accuracy. So we define accuracy as mean time between failure. How many hours of driving between a failure of the system that will lead to an accident. So in order to kind of internalize the challenge, when you look at products today on the market, they are divided into two. One family is very, very sophisticated products like a smartphone or a PC. Now very, very high-tech in terms of its software, in terms of silicon, but low accuracy. Low accuracy, meaning that the product can fail and it can fail often, and it's okay. The other family of products are simple products, or simple in terms of the technology, but very, very high accuracy. An airplane could be an example. It's simple, but the accuracy, the mean time between failure, is almost infinite, right? Airplanes are not supposed to fall from the sky. It's very, very rare. When you think about autonomous driving, you have to have both. On one hand, it's very, very sophisticated, and I'm talking about the most sophisticated compute, the most sophisticated algorithms, the most sophisticated AI, both in terms of perception and decision making. A huge challenge is from the point of view of sophistication. On the other hand, it has to be very, very accurate. The mean time between failure should approach also infinite. And this causes a challenge because there aren't many, or I think this is the only example of a product that is both very, very sophisticated and very, very high accuracy. And this is really the challenge. And the leap that you need to make is how do you build a system both from the kinds of guarantee that you make in terms of mean time between failure and how do you go and validate it. And the way we approached it is we started looking at the human statistics. Now there are about 3 trillion miles driven here in the U.S., about 6 million crashers a year. Divide 1 by the other, you have a mean time between a crash of 500,000 miles. Say you drive 10 miles per hour on average, 50,000 hours. So every 150,000 hours you have a crash with humans. So say you build a robotic car with this kind of statistics, which is human-level statistics. It means that if you deploy 1,000 vehicles, you have a crash every 50 hours. Or if you deploy 10,000 vehicles, you'll have a crash every 5 hours, and then this is unbearable from a business standpoint. So you need to be much, much better than human statistics. And we are targeting 2 orders to 3 orders of magnitude better. And this is really -- this kind of accuracy is a huge challenge. And this is what led us to this redundancy approach, where rather than taking all the sensors together into a low-level fusion, mixing camera input, LiDAR input, radar input into one compute engine and outputting a sensing state, we are building separate subsystems, such that each subsystem on its own can do everything. It can do an end-to-end driving experience. Such as then we can then multiply the probabilities, such as you say, if you want to reach 10 to the power of 8 hours in the mean time between failure, approximately, you can take the square root of it if you have 2 separate and independent subsystems. So this is one of the biggest challenge. How do you meet the mean time between failure guarantees, and how do you go and validate it? Second big challenge for consumer AV is scalability, geographic scalability. A robotaxi can be limited into a geo-fenced area. A consumer car should drive everywhere. And for that, we built this crowdsourced REM technology for building high-definition maps so that we can have the necessary maps that we need in order to power consumer AV, at first, it's going to be a Level 2+, but later consumer AV everywhere, just using crowdsourcing. And the third big challenge, which is enormous, is the regulation challenge. You need regulatory certainty in order to put such a machine on the road because accidents will happen. So how do you handle this? Who is to blame? What are the liabilities? Are they criminal liabilities? Financial liabilities? How do you get a license from the state to drive without a driver behind the steering wheel? All sorts of laws needs to change in order to support it. This is a huge challenge. So those 3 are the big major challenges: accuracy, geographic scalability and regulatory certainty. Those -- all those 3 need to come together in order to make autonomous driving a reality.

So if you dive into the first of those on the technology side of things, you mentioned a bit of having the separate subsystems as opposed to a sensor fusion box. Talk a little bit about the advantage of that approach, both from a technology redundancy capability point of view, but also the economic side of things. A number of smaller subsystems rather than one larger one. Is there an economic trade-off you're making for the eventual cost to the consumer or to your OEM customer?

Okay. So from a safety perspective or redundancy perspective, I -- we talked about it, that the redundant systems, it helps you reach the mean time between failure. It's like I have an iOS device and an Android device in my pocket, and I'm asking what is the probability that both of them crash at the same time. It's kind of the product of the 2 probabilities because those are independent systems. From an economical standpoint, it's a huge win because we can take the camera subsystem, which is cameras are low-cost sensor devices. Silicon is low-cost device. So we can take the camera subsystem and then influence the evolution of driving assist going from today's driving assist into something that has the feeling of a Level 4 system but requires a driver to be in the loop. Just like the GM Super Cruise that have a driving monitoring system. As long as the driver is alert and his eyes are facing forward, you have the ability to have your hands off of the steering wheel for a certain limited period of time. So you can think of taking something like this but extending the envelope of performance up to a Level 4 system, drive in an urban setting, drive in rural setting, drive in arterial roads, navigate from point to point. All in a low-cost system because it's only cameras and compute. Along the way, you gather data from millions of cars to help you refine that subsystem as part of an autonomous car that has another subsystem, which is the radar and LiDARs. So this idea of redundancy, where you have a separate camera system and a separate radar-LiDAR system has a huge economical impact. Another impact it has is now, when you look at the camera subsystem as the anchor point and you ask yourself, what are the -- what is the radar-LiDAR setup that I need that would be sufficient for the MTBF that I want to, that I want to reach? Of course, you can over engineer it and say, okay, I need 360-degree LiDAR sensors, 360-degree radar sensors. But now you have more flexibility. For example, the radars are evolving into imaging radars. We are building such an imaging radar, a software-defined imaging radar at Intel targeting 2023. So imaging radars are still low cost, especially compared to a LiDAR sensor. And it could give you almost what the camera provides in terms of redundancy. And maybe you'll not need a LiDAR at all. Or maybe you'll need a LiDAR, just front facing. So front facing, you'll have a LiDAR, you'll have an imaging radar and you'll have the camera setup. It provides you flexibility because you are thinking of the other sensors as redundancies. And then you say, okay, what is this sufficient envelope of redundancy that I need in order to reach the MTBF that I agreed with the regulatory body of the country that I'm deploying? Or from a business perspective, what is the MTP that I need in order to have a sufficient business envelope from a liability standpoint? So it's -- and that also has an economical impact because if we want to reach a consumer AV, consumer autonomous car, it has to have the MTBF of a Level 4 or Level 5, but you need to reach cost levels that are relevant for consumer. So the more you can -- the more you are efficient in terms of the sensory setup, the better you are from a cost perspective. Later on, when you want to deploy at the consumer level, say, in a 2025 and beyond that time frame.

Great. The second thing that you mentioned in there was on the scalability of the mapping side of things. And you have your REM solution that you crowdsourced. Talk a little bit about the benefit of that approach versus those that are trying to do it more from a virtual mapping point of view to scale the miles quickly because they don't have the vehicles on the road to do it. Is one vastly superior to the other? Or are they complementary approaches to get to the same goal?

Well, there are basically 3 approaches. One approach, you don't need a map. Tried just from online sensing to be as good as possible and try to drive without the map. Maybe drive only with a regular navigation map and, say, some semantic information about relevancy of traffic lights, which lane is relevant to which traffic light, and so how many lanes are in the road, so all sorts of kind of high-level semantic information that you can capture from some simple crowd-gathering data. This is one approach. The second approach is you go and build a high-definition map using specialized vehicles. So vehicles with a 360-degree LiDAR with an inertial -- very, very accurate inertial sensor. And then you have then you have teams of people taking that data and doing lots and lots of manual work in order to build a high-definition map. Obviously, that approach is good on the -- it's not scalable. It's good only for a geo-fenced area. Updating the map is tricky, and it's very time-consuming building this map. So it's not scalable. If you want to drive it everywhere, say, in the United States, this is not the way -- it's not a cost-efficient way or economical way of building you the map. So we are left with 2 possibilities. One is building high-definition maps using a crowd-sourced approach, which Mobileye -- [ I talk ] Mobileye is the only actor in this game. The technology is very, very difficult to execute. Building a high-definition map of -- the details that are existing in a high-definition map, using data coming from a 10-kilobyte signature per kilometer, and this is what we are gathering from every year from every vehicle, is very, very challenging. That should be compared to driving without a map. I believe that if you don't have this -- the high-definition map, reaching a high-quality product is going to be very, very difficult. Or reaching a high-quality product where the reasonable mean time between failure is going to be very, very difficult. So I think the only really scalable and sensible approach is what Mobileye is doing is relying on the scale of driving assist. You have millions of cars, tens of millions of cars every year, new cars on the road with a sophisticated driving assist with front-facing camera, you can use that, you can leverage that in order to pick up data, low-bandwidth data and have sophisticated algorithms on the cloud, off-line algorithms, in order to patch everything together and build fully high-definition maps. And that gives you the scale that you need. And then along the way, we can pick up additional information for other verticals. For example, we can do infrastructure survey. We know where the potholes are on the road. We know where are the cracks on the road. We can do asset monitoring. We know which lane mark needs to be repainted or which pole is now bent and needs to be maintained. Information that today is also very expensive to gather when you do infrastructure survey, and now you can really automate it. So it provides -- it builds now another new market of infrastructure survey, which is also a market of billions of dollars.

Thanks for all that detail. The last question I have on this topic is for the third point that you talked about, the regulatory environment. How are you seeing the regulatory environment evolve in different countries? What sort of ubiquitous kind of system agreement are you finding? Because it seems like it has to be agreed upon in a platform level where it's the RSS side of things, there's so many regulatory hurdles, and it seems like each country could be doing it differently. How are you approaching this to have an open platform where people can have kind of a ubiquitous cross-country around the world platform that they agree upon from a regulatory perspective?