Ambarella, Inc. (AMBA) Earnings Call Transcript & Summary

Well, here we are. It's wonderful to be able to resume Ambarella's CES tradition after a virtual event last year. Ambarella's first Capital Markets Day was in March of 2018, so more than 3 years ago. And we tried to have this twice at our office in Parma, Italy in the last year and were thwarted by the pandemic both times. So we're really excited, as you could tell. We've taken a lot of precautions to protect your health, to protect our employees' health and appreciate everything you've done to persevere and to be here. Let me just read our forward-looking statements. Please allow me to do that. Today's presentation and discussions will contain forward-looking statements regarding our strategy, plans and objectives for future operations, technology trends, future product introductions, projected financial targets and the size of markets addressed by our solutions and the growth rates of those markets, our ability to achieve design wins, our ability to retain and expand our customer and partner relationships, among other things. These statements are subject to a variety of risks, uncertainties and assumptions. Should any of these risks materialize or assumptions prove to be incorrect, actual results could differ materially from these forward-looking statements. We're under no obligation to update these forward-looking statements. These risks, uncertainties and assumptions as well as other information on potential risk factors that could affect our business are more fully described in the documents we filed with the SEC. So again, thank you for persevering to be here for our second Annual Capital Markets Day. I greatly appreciate it. There's been a significant global and cross-functional effort to make this a very rich fundamental research experience for you. And so we hope that you'll be pleased. We'll have 6 presentations from executives at the company. We have more than 30 demos, both indoors and outdoors. And again, I just hope this will be a good positive experience for you and a safe one. So let's just go through the logistics of the day. The floor map here shows -- we're in the Tropicana room right now, CMD, highlighted in yellow. As many of you have seen, we have demos in the Flamingo room. Please also note our Oculii demos are in the Stardust room. So given the timing of the acquisition, we had already laid out our CES plan, and that's why Oculii is out in Stardust. Please make sure you ask your host to see them. And then, of course, outside, we have several different vehicle demos, an ADAS car, 2 EVA cars with our latest L4 stack and a car with Oculii's radar installed on it. The demos will be open today until 5, maybe 5:30 p.m. And if you have a reservation, please meet out in the parking lot. You don't need to wait indoors for it. Please go out to the tents. And if you don't have a reservation, we might be filled up for today. But if you want to try to do it later in the week, the front desk where you got your name badge will be able to schedule that for you on Wednesday, Thursday or Friday. And turning to the agenda again, 6 presenters today. We're going to run through them all together, no breaks. One of our engineers, Lars, told me it's like a Las Vegas timeshare and that you can't leave the room until you buy. So -- but you're welcome, when we do start Q&A, some of you have appointments for the car. Some of you might want to go see demos, some of you might want to stay for Q&A. This whole event will be recorded. So if you do miss Q&A, you can catch it online later. Questions from online participants, and we thank you very much for joining with us today. We hope this is a very fruitful experience for you and over the next month, we'll be adding videos online that you can access to see some of what you're missing today. I will be accepting questions from the online participants and reading them, mixing them in with questions from the room. I won't go through the bios. Again, this will be available in a document after our presentation is over today. However, I would like to introduce 2 executives who will be available for Q&A. They will be available at the management presentation and just so you know who they are, Chan Lee. Chan, if you could raise your hand. There's -- Chan has been very busy recently with a product you'll be learning about shortly; and John Young, our VP of Finance. So with that, I'm going to turn it over to Dr. Fermi Wang. As you all know, Fermi is Co-Founder, President and CEO of Ambarella. Fermi?

Thank you, Louis. And again, thank you very much for coming here today and attending our Capital Market Day. When Les and I cofounded this company in 2004, the underlying premise of the company was that the digital video was such a unique data type that it requires a unique silicon architecture to address all those difficult problems. In the first 12 years, we focused on human viewing applications with our high-performance, low-power video processors and focusing on market like consumer camera as well as security camera. And in the last 5 years, we augmented our SoC portfolio with a deep learning-based AI processor, which enable machines to perceive the environment and to make intelligent decisions. And also enable many levels of automation, which is used by many different industries. And today, I'm excited to announce that we're expanding our technology processing performance and getting to a new market with the introduction of CV3 and the press release will be available soon, and -- which CV3 will be the first domain controller that Ambarella introduced, and Les will spend a lot of time to talk about the details of that. And this diagram basically shows what I just said. We start with video processor, moving to computer vision, moving to domain controller. And with this progress, in addition to human vision applications, now we can start to address a lot of different applications, which require not just perception, but fusion planning and execution. And all this new market is available to us. And before we talk about the size of those markets, I want to be clear about -- in this AI processing hierarchy where Ambarella is planning to serve. In this pyramid, the top 2 layers where the AI processing has taken place in servers, in data centers or cloud or in enterprise ages, in servers, lab are plugged in to the wall and most of the time, in air-conditional rooms. Obviously, those 2 layers are not our target market. Instead, we focus on this foundation layer, which we call it IoT endpoints, where to execute AI processing requires a very fundamentally different and optimized silicon architecture. And this architecture need to address issues like latency, privacy, security, bandwidth or any other real-time applications at a very limited power envelope. And this is where we're going to focus at. And as you can see in this IoT endpoint market, there are many different applications. One of the largest one is AI security camera that Ambarella has been dominant in this market. And another example of this market is this automotive ADAS market, where Ambarella just started taking market shares in the last year. And there are many other emerging applications in this layer and which will become mature for us in the near future. So how large are those markets combined? In this background, on the right-hand side, we're showing you the serviceable market revenue for Ambarella in the next several years. And the darker bars represent the market for automotive and the lighter bars is for AIoT markets. And if you compare -- of course, that Chris will cover all these numbers in detail later. But if you compare this diagram with the diagram we generated 5 years ago, you will notice that the key driver 5 years ago was those consumer discretionary products like drones, like sports camera. But today, in addition, that we are addressing those megatrends like security, safety or higher and higher levels of automation and eventually go to robotics. And if you look at financially, there are 3 major triggers -- drivers, 3 major drivers that driving this large and growing market opportunity for us. And all those 3 drivers are triggered by our computer vision technologies. First, the computer vision really helped us to have a brand-new product cycle in our existing IoT markets. For example, security. Second, the computer vision help us reach into new markets that were not available to us. For example, the Level 2, 3, 4 autonomous driving. And third, I think the -- we talk about this like computer vision SoC ASPs, twice of a video processor on a like-for-like basis. So with that, we believe the quality and the diversification of our revenue has never been higher before. And until recently, if you look at our SoC strategy, our SoC portfolio has been mainly focused on video perception, which is showing at this bottom left corners in the pyramid. And that we have talked about before that, the performance requirement for video perception is much, much higher than any other sensor or other processing requirement in this pyramid, which give us a great opportunity and a great position to continue to integrate more functions within the pyramid into a product, therefore, capture more values. So moving forward, our SoC strategy is try to expand horizontally and vertically within this pyramid. The horizontal expansion means in addition to just a video perception, we are going to provide complete perception level to our customers by integrating other sensor modalities. Now the acquisition of Oculii is the first step, that with the acquisition, we got the radar technology in-house. And moving forward, we will continue to integrate and to interface with more sensor modality in the perception level, for example, ultrasonic. And also, we plan to integrate vertically, meaning in addition to the perception layer, we want to address fusion, planning and even the controlling of the whole software stack. And with the introduction of CV3, we will have enough hardware support and hardware performance to complete this vertical and horizontal integration. But also, I want to make it clear that this SoC strategy is designed not only for automotive, but also for all the other IoT devices, IoT markets, including robotics. On the software side, as you might know that a majority of our engineers are software and algorithm engineers. And the software value that we have been providing was really focused on these foundational layers that we show in the bottom of the box. And we capture those software value by incorporating it into our SoC price in the past. But with the introduction of CV3 that we need to, and we will develop this complete software stack, so that we can demo the function and performance of CV3 to our customers with that complete software stack that give us new opportunities to capture more software values by offering modules in the upper level of the stack. But we need to make clear that we do not plan to offer a black box total solution to our customer. In fact, our customer has made clear that they don't need it. Well, instead, what we are trying to do is module -- provide a module-based software and we will allow our customers to pick and choose the modules that they need, and we will help them to integrate those modules into their software. For example, we already licensed some of our perception module, like blind spot detection to some of our e-mail customers. And moving forward, we are planning to start licensing more software module like video processing, stereo processing or radar processing technology to our key customer in the automotive space. Combining this strategy with our open platform, we believe we will offer more software value to our customers. And in order to deliver this new SoC and software strategy, we have been investing very heavily on R&D and hope to use this slide to give an idea how much we have invested in the last several years. In fiscal year '16, Ambarella's management came to a conclusion that our traditional consumer product line will not give us enough revenue growth, that we can come down. And therefore, we need to transition our target market to industrial and automotive. And at the same time, we also realize that computer vision technology is one of the key technology will help us to penetrate those market. That's where you start seeing in the last 5 years, we start invest heavily on R&D, particularly in the computer vision R&D. And you can see that it rise quickly, while our overall revenue decreasing because of the weak performance of our consumer product lines. And in fiscal year '21, our total R&D was 60% of our revenue. And unfortunately, that was also a year that our computer vision revenue becomes material. That's definitely the inflection point that we are hoping and that we're glad we reached that in fiscal year '21. And if you look at this year, which our fiscal year will end in the end of this month, and we already provide a forecast that we're going to have a record high revenue this year. And also with the growth of revenue, our total R&D will be roughly 50% of the revenue. And moving forward, we believe that the computer vision revenue will continue to rise and which will help us to achieve a positive operating leverage. A lot of people ask us how we can compete with those large competitors, considering we have only 800 people in Ambarella. Of course, that -- the $600 million R&D expense I just showed in the previous slide helped a little bit. But the most important reason behind it is our technology, our efficiency, our open platform and scalability and flexibilities. And I hope you have time today or later this week to visit our demos in the next room, which will -- hopefully will give you the proof that we do have technology we need to compete. And the reason we can have those technology is really because of people, because of our domain knowledge and our algorithm-first approach. If you follow our engineering execution in the last several years, you should have noticed that we have followed the most advanced processing node very closely. For example, CV3 is our second 5-nanometer chips. Even with this aggressive movement in the process node, we continue to help our customer to build new products efficiently and quickly. I think that's a proof that we have one of the best engineering team out there. And in terms of domain knowledge, I think this founding team, we all have more than 25 years of video experiences or radar or autonomous driving. And on top of that, Ambarella probably is the only pure-play digital video company in the last 17 years. And throughout that what our domain knowledge has, again and again, built the most efficient architecture for this digital video applications. The last, but probably the most important factor is our algorithm-first approach and because that's an engineering topic, I'll leave that to Les to talk about. And I think that's definitely the most important reason that we can continue to develop advanced technology. So of course, that -- I think to answer the question whether we can compete, I think I would like to argue that not only we can compete with those large companies and also winning the market. And I think that confidence is not a wishful thinking, but are supported by the real data. Here, I'll show you 2 sets of different data. On the left hand side, the first green bars right here is our CV Wave 1 revenue, which comes from enterprise security camera. The second green bar right here is our Wave 2 CV revenue, which is from home security camera. And the third layers of the green bar is our Wave 3 revenue which comes from automotive revenue. And the yellow part is our video processor revenues. And as you can see, in fiscal year 2022, our CV revenue will grow to 25%, which we have been talking about for a while. And moving forward, in the fiscal year '23, we believe this number will reach 45%. That just show you we continue to build our CV revenue momentum. On the right-hand side, a different set of data which probably -- but show -- will help you to come to a similar conclusion. First of all, we have 275 unique CV customer who is a unique CV customer, meaning those customers has paid to buy either SDK or ref design or engineering samples. And out of the 275 unique customer, more than 100 of them already reached production with us. And among those customers, they already built more than 150 unique CV products and reached production. And in fact, we showed a similar set of number 6 months ago. It was 240, 58 and 87 6 months ago. And that -- I think that shows you we continue to build our momentum on this customer base and continue to help our customer build better products. With these 2 sets of data, I think there is a very strong and growing evidence of market acceptance of our CV product line. In terms of global operations, we have a total of 874 employees, more than 80% of them are engineers. And among all engineers, more than 70% of them are working on software and algorithms. And in terms of our global supply chain, I would like to say that we are lucky to work with some of the best partner that we can find. For example, Samsung for foundry, ASC for packaging and Seagull for testing. And in the last 12 months, as you can -- you know that there is a huge supply problem. Although the long wafer lead times and the material shortage persist, but I have -- I'm glad to say that Ambarella will sell in the bottleneck of our customer product delivery plan. On the contrary, I think the biggest challenge we are facing is most of our customers have a hard time to collect, to get enough other parts, other components on the products they are delivering. It definitely creates some problem for us. For example, it creates -- it makes our revenue forecast become more difficult. And also it might has -- create some pockets of inventory in our overall sales channel. And definitely, we are cautious and we're going to continue to watch out of these 2 problems moving forward. At the end, I want to have a quick summary. I just talked about the momentum we got with our current CV product and we are going -- we are introducing a brand-new product line, which will expand our SoC and software offering to our customers. I think this is a good time for us to think what we need to do to reach $1 billion revenue for the company. I think the first thing we need to do is clear, that we need to execute on the strategy I just talked about, both SoC and the software strategy. But more importantly, we need to continue to drive our innovation. In fact, I'm proud to say that in the past 17 years, we start competing with TI and HiSilicon, that now we are competing with NVIDIA, Intel and Qualcomm. At the same time, we are keeping our gross margin around 60%. And the reason we can do that is our innovative products. Les will give you many examples, and I would highlight a few of them. We need to continue to extend our video processing leadership. We're going to talk about how to use new network to improve image process single processing. We need to unlock the synergy with computer vision process radar technology. We need to extend our power-efficient leadership with new silicon and the algorithm technology. And the third thing we need to do is continue to scale organization both in engineering and business development through organic growth as well as acquisition. I think the acquisition of Oculii is a great example that we can acquire not only the technology, but also talents at the same time. And that's going to be one of the business model that we're going to continue to focus on. Last, I think reaching $1 billion revenue cannot be the only financial goals that we have for this company. When we reach that goal, we need to continue to drive positive operating margin for our investors, for our company, and I think that's also a very important goal moving forward. With that, I would like to introduce Les to talk about our new technologies. Thank you.

Right. Thank you, Fermi. I'm very happy to be here to introduce you to our CV3 family, which is a project which is kind of near and dear to me because we've been looking forward to this day since our VisLab acquisition in 2015. At that time, VisLab self-driving car had a trunk full of server PCs in it, which heated up the interior and caused the rear of the car to sag down. So obviously, not a production level solution. And so we've been working towards a chip family that can bring true autonomous driving to mass production state in the most efficient and cost-effective way. This new family is our most ambitious project to date. And it allows us to extend the performance that we can achieve on a single chip, up to 42x faster than our previous high-end chip, the CV2 family. So that's a huge step-up in performance, and this will allow us to address everything from ADAS to L2+ to L4 autonomous vehicles with a single unified SDK. And that's another very important thing that we've seen from our customers is that they have multiple different requirements and different price points that they need to hit, but they want to have one common software architecture for all those vehicles. And with the CV3 family, they'll be able to do that using a single SDK. And at the same time, CV3 will extend our leadership in power efficiency by offering 4x more performance per watt than what we had with the CV2 family. So just to recap, the evolution of our product line, we started off in the human viewing centric applications. Then we introduced the CV2 family, which was our first production computer vision family. And that has been used in a number of different automotive applications, including front ADAS, in-cabin monitoring, as well as L2+ and L4 vehicles, as the perception processing engines for those vehicles. With the CV3 family, we will now extend this to be able to run the full stack processing all the way up through fusion and planning. Our philosophical approach to new chip development has been really the same since the beginning of the company when we started off with image processing and compression. And we've iterated that over many generations to continue to improve that. But it's based on a core set of algorithms that we have developed and we use those algorithms to optimize the architecture design to make sure it's efficient for those. And then we have extended this approach into computer vision processing with our CVflow processor which again has been optimized for running computer vision algorithms. And in order to understand how these computer vision algorithms look, we did an acquisition of VisLab in 2015, which had an autonomous driving stack. And we've been working with the VisLab team ever since to extend this approach all the way up to the full stack. And then in October this year, we acquired Oculii, which allows us to augment our vision-based processing with radar-based processing, a very unique and high-definition version of radar processing, which we believe, when combined with the vision processing is the foundation for the robust sensor suite that you need for autonomous driving. So let's look at where CV2 is relative to our competitors because that's something that we can benchmark today. Actually, 2 years ago at CES, we showed a demo of CV2 versus the N companies' GPU and they have a 30 tera ops GPU, which was their kind of processor designed for automotive applications. We're on the exact same network on both chips, mobile net, SSD, same resolution, everything. And when we benchmarked it, we found that our CV2 processor was actually slightly faster than the GPU processor. And this is -- shows that even though we actually rate CV2 at 12 eTOPS, it actually is performing at the level of a 30 TOP GPU. At the same time, the power consumption was 1/5 of the GPU chip. Now this CES this year, we're showing a CV5 demo where we again compare with the same GPU chip. And with this CV5 processor running the latest YOLOv5 network, we're seeing a 4x speed up in CV5 relative to the GPU. So now we're 4x faster than the GPU and we're about 15x more power-efficient than GPU. So even with CV5, we've continued to improve the gap with GPU processing. So how are we able to get this level of efficiency? Well, if there's anything that I've learned in over 40 years of working on computer architecture, is that it's very easy to throw down a lot of talk, to claim I have a huge number of multipliers and adders on the chip, and I have 1,000 tera ops. What's hard is to make the architecture efficient. Can I really utilize all those multipliers most of the time? Or most of the time, they're just sitting idle, not doing anything? So the key to getting a high-efficiency architecture is to eliminate bottlenecks in the processing. And those bottlenecks really come about due to the requirements of fetching all the data that you need to keep all those operations busy. And we have spent a lot of time on optimizing our CVflow architecture. I won't go into all the details that are on the slide there. But basically, those architecture improvements allow us to flow data through the data path and the data paths themselves are designed to require less data per operation than the traditional general purpose processing that you see in a GPU or CPU. And so as a result, we operate at a much higher efficiency. So what -- the trade-off we've made is basically optimized for AI and sensor processing and not optimized for things like cryptocurrency or gaming. Now with the CV3 family, what we did was =to build on all the experience that we had with the first 2 generations of our CV processing. And so by now, we have hundreds of networks that we've looked at that are coming from the open source community, from our own internal network development as well as customer algorithms. And we went through all these networks and we look for what are the bottlenecks that still remain with this CV2 generation processing. And by fixing all of those bottlenecks and incorporating some new developments, we were able to get a 3 to 4x improvement in power, area and DRAM bandwidth utilization versus CV2. So this is the foundation of our whole next-generation family of products which will continue to lead in terms of efficiency. At the same time, we wanted to boost up our CPU performance. So we went to the latest ARM Cortex-A78, the highest performance automotive ARM core so that we can run all the rest of the high-level functions that the full AD stack requires. This family will have a single upwards-compatible SDK from CV2. That means software that's developed on the CV2 family can be easily ported onto the CV3 family. And it will also be used for our security and robotics markets in the future. So why do you need such a powerful processor for running full stack? Basically, it's because in an autonomous driving car, there's a lot of sensors that are required. For example, on our EVA car, we have a total of 10 cameras and 5 radars. And several of those cameras are operating at ultra-high-definition resolution. So we currently have 16 CV2 chips inside of that car to do just the perception processing. So you can imagine when you combine the perception, the fusion and the planning together, you need a large performance upgrade over where CV2 was. And that's what we're delivering with the CV3 family. So the first chip that we've developed and that we'll be introducing is our CV3-High chip. And this is the flagship of the family. It's the highest performance chip, and it delivers this 500 eTOPs for 8-bit and it actually can run twice as fast for 4-bit network layers. So it can actually run up to 1,000 eTOPS for that. And so it's basically 42x faster than our CV2 processor. And with the ARM A78 cores and a lot more cores on the chip, it's going to be 30x faster in ARM performance versus the CV2. And so even with that large performance increase, the power is still only 50 watts of power. So that's where you can see roughly 4x improvement in power efficiency relative to CV2. This CV3-High chip can support up to 20 cameras. And the other key capability that it brings to the table is allowing true LiDAR class resolution with the latest Oculii algorithms running on CVflow. So we've already -- since the acquisition, have been working very closely with the Oculii team on evaluating their algorithms. And even with this first chip, we will be able to take full advantage of their most advanced algorithms with 360-degree radar perception. And going forward from here, we're going to further enhance the radar performance efficiency with specific hardware optimizations. So this is the block diagram of the whole chip. You can see it's a full SoC with all the peripherals that are required to both interface to all the sensors in the car as well as into the control subsystem for controlling the autonomous driving functions. It also has high-bandwidth I/O that allow you to connect high-performance storage or other processor systems like infotainment systems. It has 16 ARM Cortex cores, which support ASIL-B applications as well as safety island lockstep applications. And it has a GPU for running 3D visualization for surround view. One of the functions that we wanted to enhance for a full stack domain processing is improve the security capabilities of our chip. And to do that, we've developed a new dedicated hardware security module, which does a number of functions, including crypto acceleration, secure storage management. But what I'm showing on this diagram is a new capability to support multiple domains running on the same chip. And what this allows you to do is isolate the different safety levels from each other, so you can guarantee that nothing in the ASIL-B domain can corrupt the safety domain, nothing in the QM domain can corrupt the ASIL domains. And another very important thing that it allows us to do is protect software coming from third parties or from Ambarella from being spied on by customers. So basically, it allows secure software deployment in a way that nothing that you can do on the ARM processor can break that security. So summarizing CV3, it's a full domain controller family that scales up from ADAS to L2+ to L4. With -- combined with the CV2 family, it allows us to offer a range of 1 eTOP to 500 eTOPs. So 500:1 range in performance, which we think is the broadest in the industry, and it offers industry-leading power efficiency. So given this new capability, what are some of the things that we can do on the software side to take advantage of that? So as Fermi mentioned, one of the areas we've been working on is how to leverage CVflow to augment the image pipeline that we've been developing over many generations. So what I'm showing here is a side-by-side demo of our image pipeline -- traditional image pipeline on the left side. And on the right side is the augmented AI-based processing version of that pipeline. And you can see that there's a huge difference in the level of detail, contrast between those 2 pipelines. So even though we've -- we spent many generations of development with our classical pipeline, using AI is fundamentally a more -- a much more powerful way to augment image processing. And with the CV3 AI processing capabilities, we'll be able to apply that to more applications and higher resolutions and higher frame rates than what we can do on the CV2 family. Another way that you can use AI processing and image enhancement is for high dynamic range images. So traditionally, when you have this kind of backlit indoor, outdoor scene, it's very difficult to have good detail in both the highlight area and the shadow area. So on the left side, you can see highlight area is okay, but the shadows are kind of blacked out, hard to see anything. Whereas on the right side, you have full detail in both highlight and shadow areas. And that's much easier to do with AI-based processing. So the other major area that we've been focusing on in the software development is our Ambarella stack. Alberto will be telling you a lot more detail about it, but I just want to review what the stack is doing and some of the high-level improvements we're making. So the stack basically help us to combine [indiscernible] together [indiscernible] There's still a bicycle there that it's keeping track of and monitoring whether it's going to be able to turn off here. So basically, it's predicting where the bicycle is going to be going and deciding it can safely take that exit and go into -- leave the roundabout. So that's the kind of processing that you have to do to run a full AD stack, and that's what we're planning to do on CV3. So people sometimes ask why develop your own AD stack. Well, the original motivation for developing that was that we need to understand these core algorithms so that we can develop the right chip architecture for doing that. But what we've found along the way is that, obviously, as we're developing these algorithms, we're optimizing them to run on our chip architecture. And because we kind of know the chip better than any of our customers, we can optimize those algorithms to run on our stack to a level that might not be possible for our customers. And so we want to be able to help customers take advantage of some of those software modules, if they wish, to get the highest performance processing that they can out of CV3. But there's another very important thing that has been enabled by the Oculii acquisition, which is deep fusion of vision and HD radar. So, so far, radar has been kind of a separate subsystem, which runs on its own processors out at each radar module. With CV3, we're going to be able to centralize all the radar processing into a single processor, along with all the vision perception processing. And when you do that, that enables you to fuse the radar data and the vision data at a lower level than is possible with separate subsystems. We believe that, that will lead to a significantly more robust sensor suite than anything that's possible with current generation sensors. So not only do you have this LiDAR class resolution, but when you combine it with the deep fusion, we believe that you'll have the most cost-effective and robust sensor suite for all conditions. So with that, I would like to introduce Alberto Broggi, who is the General Manager of our VisLab team.

Thank you, Les. Good afternoon, everyone. The presentation that I will be giving today will cover 2 aspects. So the first one is a description of the announcements and the progress that we did on the autonomous driving stack in the last few years. And the second one will be the description of the demos that you would be able to see outside later today. So I will start with the first one, by showing you what happened 2 years ago. So 2 years ago, we were giving demonstrations with our EVA cars, and we were showing autonomous driving and autonomous parking. You could jump in the car and the car would drive outside of the parking lot, drive around, around the hotel here in Las Vegas and then get back to the parking lot, find an empty slot and then park itself. So that was done 2 years ago. And in order to do that, we were using HD maps. So everything was pre-mapped with very high accuracy. You could map all these surroundings here, and then you would drive thanks to these high-definition maps. And since I mentioned high-definition maps, I would like to show you what's the difference between a high-definition map and a standard map. So in this picture, this is a satellite picture showing one of the junctions actually around the corner here in Las Vegas. And you would see that in orange, we have information coming from a standard navigation map. So that means that you can see that there are a connection between the lanes, connection between the roads, so the topology of the lane, of the road. Typically, you also have other information connected to these lines. For example, you have the number of lanes or you have the speed limit, which is connected to that specific segment. And this is what you have in your normal and usual navigation software. But if you want to see the difference between a standard map and a high-definition map, this is how our HD map looks like. So it's way, way denser than the previous one. And if you really want to appreciate the differences, I can remove the background picture. So you can see that we have a lot of lines, a lot of information in the HD map. If you want to enlarge portion of that, you can see the different information that we have. For example, we have lane markings, we have curbs, we have a stop lines where we have to stop when somebody is crossing the road, we have pedestrian crossings, traffic lights. So all these things have their precise geolocation. So the precise position in the world. And this is our HD map. So it's pretty dense. And the question is how do we create this map. So 2 years ago, we were doing this -- we have this process of creating the HD map, which actually required a little bit of human intervention. So you had to fine-tune the position of the objects and the lane markings. But today, we have a tool that is able to help us. And I would show you what the tool looks like in the -- in this video clip. So you just drive around with your car, with -- and you take images from the cameras in your car. For example, we have a stereo camera in the front and you take these kind of images from the stereo camera. Then you process these images, for example, with this type of engine that you have in the camera itself. And you get this kind of information. So this is a 3D point cloud where every single pixel has a color information and has a 3D information in the world. And you also do the processing of these pixels and you label these pixels as road, lay marking, curb, traffic sign and so on. And you do that for all the sequence that you acquire and multiple sequences. So if you put everything together, you can easily map the whole town. And this is, for example, Parma. And we've been driving around and we've been mapping the road and the whole city in this way. So the -- of course, if you need to drive on an HD map, that's -- your driving would be very efficient because you already know how many lanes you have, you already know how -- where to expect the traffic light, where to expect the lane marking or a junction. So you already know everything in advance. So driving on an HD map is very efficient. But of course, you cannot assume that you can have a map of the whole world. So it will happen that you have to drive in some roads that does not have any HD map. So this is what we introduced in these years. So we have a way to drive also where you don't have HD maps or you only have standard maps. And in order to do that, we -- what we do is that we just perceive the environment, and we drive with just what you can see, so with the real-time perception. And this video clip shows that. So in this case, we were driving in a non-HD map area. So we're just using those standard maps. And we were just acquiring information and driving with what you see. So as I mentioned, driving with HD map is way more efficient. But actually, you can also drive without the HD maps. Plus the system understands by itself when it needs to switch to either HD driving or standard driving. And of course, if there's a mismatch, for example, you have been mapping one road, there's a mismatch between what you see and what you recorded maybe 2 months ago before -- 2 months before or maybe 1 year before. So if there is a mismatch, the system will understand that and it will switch back to standard driving. So -- without using the information, the old information of the HD map. Of course, I've been talking about mapping, but mapping is just one part of the autonomous driving stack that we have. So we have perception, as was already mentioned. Data fusion, and we are also fusing data coming from the Oculii radars. Object tracking, we do the prediction of the motion of other objects. Localization, planning, trajectory planning and maneuver selection. So these are all the pieces that we have into the autonomous driving stack. And just to show you something that you will be able to see in the demo outside, these are some clips that show some of the capabilities of the system. So you would be able to see, following a slower vehicle merging in a road with higher traffic, you will have to handle pedestrians, traffic lights, unprotected left turns. So these are all the systems, all the capabilities that you would see into the car. So just to summarize the improvements that we did since the first demonstration we did 2 years ago. First of all, we added new behaviors in the HD map driving. Second, we extended also to non-HD maps, so we can drive in known HD map areas. Of course, we started with limited complexity roads and parking lots, parking lots are pretty complicated, and you will see it during the demo. And also, we added the possibility of exploring the parking lot. So when you enter in the parking lot, the car will start wandering around and searching for an empty slot. And then when the car will find an empty slot, it will be making the maneuver to get into the parking lot and so on. And the same on the way back. So when you exit from a parking slot, it will try to find the exit of the parking lot and then get back to the road. So that was for the first part. The second part is the description of the demos that you would be able to see outside. So we have 2 type of cars. The first car is the EVA car, the same car that we showed 2 years ago, but actually with announced functionalities. And it would be driving around in autonomous world with that car. Plus we have another car, a little more advanced, say, EVA 2, if you want, that it would be driven manually so that you can experience the fusion that we do between stereo vision and radar because that car has the radar -- the Oculii radar installed in it. So with the first demo, I already mentioned that we will be driving both with HD maps and without HD maps. We picked 4 hot points here around in Las Vegas. There are 4 parking lots. We'll be able to select the route. So I want to go to the museum, I want to go to the university or whatever. So the system will create the route and will start. And you will see that everything has been mapped here, so all the roads are mapped, but not the parking lots. So when you would be entering in the parking lots, the system will understand that there is no HD map and the system will get back to standard driving. And each parking lot has some pecularities like a simulation of a drop-off of passengers or pick-up of passengers, so you will see in each specific parking lot. And then this is the second car that we have. This car integrates the Oculii radar in the front. So you'll be able to see, as I mentioned, the results of the fusion between stereo vision and the Oculii radar. And if you have a look at the car in the trunk, actually, if you turn the car, you will see that this car has only one box for the processing. And this is why we actually designed the car to be CV3-ready. So once CV3 will be ready, we will just swap the boards, and that would be it. And as we already mentioned, it's slightly different from what we had a few years ago when we had -- our trunk was full of PCs. We had 16 servers in that, and the power consumption was something like a little more than 3 kilowatts, just for the processing in the trunk. So this is what you will see in the demo area outside. So with that, I will thank you for your attention. I will turn to Steven. So Steven is our VP for radar technology and the former CEO of Oculii. Steven, the floor is yours.

All right. Well, it's my pleasure to be here today. And as you've heard from all the fantastic presentations, there's a lot of synergies that we believe exist between Oculii and Ambarella, and we're excited to share some of those synergies that we've already unlocked with you here today. In many ways, radar is really a very natural complement to camera-based processing because of a lot of physical characteristics. Number one, radar operates at a completely different physical wavelength than optical light does. And so environments and situations which might blind or degrade the performance of the camera perception, these are actually scenarios where the radar can enhance the perception such as low lighting conditions, weather conditions in rain, fog, snow or even obstructions on the lens, radar can actually see through all these different environments because of the longer wavelength and so it naturally complements optical-based camera processing. Now the challenge is traditional radars are very poor when it comes to spatial resolution. So although the radar can see through the weather, it can see through the obstructions, what it sees in a traditional automotive radar system is not very clear. And so this is what our fundamental AI software addresses. It takes and preserves all of the advantages of radar but it makes the resolution significantly higher so that it can be a peer to the camera-based processing and allow for a sensor fusion of 2 very high resolution and efficient sensing modalities. Before I talk about how the Oculii software works, I think it's important to first frame why traditional radars suffer from poor resolutions, particularly in the automotive domain. Traditional radars are what we call not intelligent sensors. They typically repeat the same signal over and over and over again, constantly and repetitively. And so as a result, a traditional radar requires many, many, many antennas, hundreds, if not thousands of antennas, to have high resolution because this is how each antenna measures the environment and each measurement provides it with information to create higher resolution. Now unfortunately, more antennas also increase the cost, size and power exponentially, while the performance actually only increases sublinearly. And so this has effect, creates an unattractive trade-off where the sensor cost and size and power increases, but the performance does not. If you look at the upper end of radar performance, for example, in the military, these sensors are extremely high performance, very, very high resolution, very long range, but these sensors cost tens, if not hundreds of millions of dollars and are more expensive than the platforms themselves. So this is obviously not a scalable solution to deliver high-performance radar. And so this is exactly what Oculii has addressed with a fundamentally different approach using software and intelligence rather than with more antennas and with hardware. Oculii's software fundamentally breaks the assumption that the radar should be dumb. The radar should just constantly repeat the same signal. We use an adaptive waveform that learns from the environment and embeds different information at different times. This information that we're adaptively embedding and learning allows us to design the antenna array in the radar data cube in a very different way, leveraging sparsity and then computation to effectively fill in the missing information that the sparsity does not provide. And so this is, in many ways, a software-based solution to a hardware-based problem. And what this does is it actually fuses and complements very well Ambarella's strengths because of the computing capabilities that they've already been able to demonstrate on CV2 and now that we'll be able to demonstrate even more effectively on CV3 with multiple radars and multiple sensors. Just to give you a sense of what we've been able to achieve on existing commodity market-proven radars already, traditional radars as I mentioned, as you increase the number of antennas -- the X-axis here shows the number of antennas in a radar system. As you increase the number of antennas in the system, the resolution does not increase very much. And as a result, even as you're going from 12 antennas, even up to 200 antennas, you're still not in the class of resolution that you really need to be in order to have this be a peer to an optical system. Oculii's software, already on existing embedded silicon, provides up to 100x increase with the 3 different platforms that we've already built and showcased. The FALCON sensor is a corner radar designed to provide up to 20x performance and the RAPTOR on the very right-hand side is a sensor that has up to 100x performance improvement and is already getting close to LiDAR-like performance but at several orders of magnitude cheaper cost, much lower power. And in a market-proven radar solid-state sensor that's very, very easily integratable into mass market automotive systems. With the new CV3 architecture, one of the things that's fundamentally going to change is not just how much processing capabilities we have now for the radar processing, but by centralizing and processing all of the radar data on the same SoC and on the same fabric that you're also processing camera data on, this will enable OEMs and Tier 1s to build a completely different type of intelligent radar system that is able to leverage this technology along with the capabilities of CV3 to deliver even better performance that can dynamically shift to where the processing actually needs to be as the vehicle is performing different types of maneuvers. So for example, if the vehicle is driving at high speeds on the highway, you want to be able to see as far as you can 400, 500 meters down the road with very, very high precision. But when you're taking an unprotected left turn, for example, these types of maneuvers will require you to look in different directions with higher resolution in different ranges. And our software, combined with the capabilities here in the central processor, will allow you to shift where you want to use processing for to increase resolution where you need it. Most importantly, though, this CV3 processor, as you've already heard from Les and Fermi, is an order of magnitude better than what we've already been running our existing software on the embedded side. And we expect that with this performance that we'll be able to take advantage of, this will unlock even higher resolution, even longer range for the radar perception. But most importantly, it will do all of this while preserving the cost, size, power and solid state capabilities that have made radar already such a prevalently and widely deployed sensor already in mass market vehicles. Just to give an example of what the radar is able to showcase today. This is an example of what the radar data is able to provide from a point cloud perspective compared next to a 32-channel LiDAR, which is probably 2, if not 3 orders of magnitude more expensive. On the left-hand side there, you see that the LiDAR data, although high resolution between 0 to 50, it can't really see past 100 meters even in this type of good condition. The radar on the other hand, is already able to see almost 400 meters. And most importantly, not only can it see, it's also able to measure the precise speed in every single point, which allows you to very efficiently determine whether the target is approaching, leaving it as a threat or if it's something stationary that you're actually needing to use for mapping and localization to build up the HD maps that Alberto was sharing earlier in the AD stack. And lastly, as I mentioned, the radar still retains the traditional advantages of the radar sensing modality, namely that this capability is still achievable and reliable in all weather conditions, and this sensor fundamentally a solid state, low-cost, mass manufacturable and already ready for mass market automotive deployments. So today, here at CES, we're going to be showing 3 demos that I think will showcase the capabilities of this radar in a variety of different use cases and scenarios. As Alberto mentioned, we're going to be showcasing the radar on the EVA 2.0 vehicle, and so you'll be able to see what this looks like. This is another video here showcasing what that point cloud looks like in a full 3 dimensions, the previous video we just topped down, but this video is actually showing the full 4 dimensions actually, it's X-Y-Z plus Doppler. And what you can see here is that the radar is actually able to get extremely high resolution, a lot of detail and a lot of sensitivity. And all of this is going to further improve significantly as we move this processing onto the CV3 processor and unlock the capabilities that the processing enables along with what the Fusion can do. So we're very excited to showcase the initial kind of implementation here on the EVA 2.0, but this is going to improve even more significantly as we further integrate into the CV3. We're also going to be showcasing the radar's use in a lot of different other verticals and applications as well. As Fermi mentioned, one of the big markets and use cases already for Ambarella silicon today is in security. And in particular, in security use cases, you encounter in -- face many of the same problems that you do in the automotive environments with respect to low lighting conditions and weather and rain and detriments there. And so again, here, the radar provides a natural complement to camera-based systems, very efficiently detecting motion, very efficiently seeing through low lighting conditions and obstructions. And most importantly, all of this can be done, fused inside of the CV2 or CV3 processors, enabling a very cost-efficient way to add these capabilities while still retaining a lot of the core advantages that radar can deliver. Lastly, we'll also be showcasing the beginnings of using these radars on other autonomous systems outside of automotive as well. As you can imagine, self-driving cars get a lot of the attention, but they're not the only autonomous systems out there that are benefiting from all the advances in perception over the last several years. In many ways, a lot of the autonomous systems such as these autonomous mobile robotics systems and indoor and outdoor systems, they actually have an even smaller cost, size, power envelope that the entire perception stack has to fit in, in order to meet the requirements of the end users and service providers. And so in many ways, we actually think that, again, camera plus radar processing is a very, very natural fit for a lot of these autonomous mobile robotic applications. Whether it's for outdoor delivery, indoor logistics, these systems have to perform with precision and with safety, especially when they're interacting with humans. And this technology that we have from Oculii is really going to complement automotive, security and robotics use cases across the spectrum. And so we'll have a third demo here showcasing the robot there on the top left-hand corner is our own robot, but those are some of the other robots of our partners that are using this sensor. But it will showcase our sensor modules being used in these robotic applications so you can get a feel for what the other applications are outside of automotive as well. Great. And with that, I'll hand it over to Chris, and he's going to talk a lot more about some of these use cases and markets and very exciting to hear from him as well.

Thank you, Steven. I think before I start talking, we have a little video to show you of CV3, I hope. One more. [Presentation]

Okay. That concludes my presentation. Actually, you don't get away that easy, I'm afraid. So I'm going to talk to you about the target markets and applications on the road to achieving $1 billion in revenue. And I'm actually going to start looking at our IoT business. I think all of the presentations up to now are focused on automotive, particularly with the CV3 announcement and Oculii. But let's first look at the SAM for IOT business. As Fermi already presented, our total SAM for fiscal year '22 was $2.3 billion and is forecast to grow to $6.9 billion in fiscal year '28. Now I'm going to discuss the revenue SAM for IoT, excluding the automotive. The bottom 2 graphs here show the SAM for enterprise and smart home security. Sorry. Thank you. The bottom 2 bar graphs here show the SAM for enterprise and smart home security. Today, these markets represent about 50% of our total revenue. And together, they grow to about $1.5 billion by fiscal year '28. Above that, you can see markets that represent new opportunities for Ambarella, ID authentication or access control and robotics. The other category is mainly consumer products, which we still serve in terms of drones and 360-degree camera products, such as those from DJI and Insta360 and sensing cameras of all clients that I'll review shortly. And lastly, we have radar. Although we based the Oculii acquisition primarily on the automotive opportunity, we are excited with the radar opportunity in the IoT market, and we will have to revisit these numbers which are admittedly conservative at this point. Our product portfolio of AI vision SoCs is the widest in the industry. Let us introduce you to CV3 based on our new third-generation CVflow architecture. But we also have a full lineup of SoCs from the low-cost CV28M all the way through ASIL automotive chips up to our 8K CV5, which we introduced last year. Of these 7 families, 5 are in production today, with CV5 expected to be in mass production by the end of this year. CVflow offers the best performance per watt and is combined with our strengths in image processing and quality. Our customers like the ability to have a common software SDK and development tools so that they can develop multiple camera families with the reuse -- with much reuse between the different family members. And all of the family members share advanced image signal processing pipelines or ISPs that provide outstanding imaging under challenging lighting conditions, which is a critical differentiator in security and automotive applications. And we will be improving upon these with our AI-based ISP developments in future, as Les demonstrated earlier. We have introduced sensor fusion, for example, with vision and radar on CV3. But it's important to know that we have been doing sensor fusion in nonautomotive markets for a while now, and I'll show you some of these examples. In this slide, we have many sensor types listed on the right and images of the -- sorry. Sorry about that. I shall start again. We have introduced sensor fusion, for example, of vision and radar in CV3 earlier. But it's important to note that we have also been doing sensor fusion in nonautomotive markets for a while now, and I'll show you a few examples. In this slide, we have many sensor types listed on the right-hand side and images of the fused results combined with these sources on the left. Actually, audio could also be added. In these examples shown here, we are combining vision data with radar in robotics applications, time-of-flight in robotics and sensing applications and thermal data in security cameras and structured light and access control. Also, our AI performance, combined with IR image processing makes us unique in many of these applications. Earlier last year, we introduced our CV5 family. This is state-of-the-art for video security and our first chip in 5-nanometer process technology. With CV5, customers no longer had to trade off between AI performance and video resolution. The ability to process 4 4K images on multi-imagery cameras is unique and addresses the fastest-growing market segment. We also introduced a low-cost 4Kp60 derivative, the CV52 which offers state-of-the-art imaging and AI in a single or dual-sensor camera, performing AI and 4Kp60 video and under 3 watts of power. And we introduced an 8K version for consumer applications that encode 8K video in less than 2 watts, making it suitable for small form factor action cameras. CV5 can also encode 14 separate streams for automotive recording applications and we will demonstrate that to you at CES this week. We have multiple wins at key customers and mass production is expected to begin by the end of this year. Let's look at the IoT security camera market. It has largely transitioned from analog, analogous to 1G, to network cameras or 2G. CV represents 2 types of opportunity in the IoT camera market: one in new product cycles in existing markets; and two, acceleration of growth in installed base, in particular, for monitoring applications. Now the transition is towards AI or 3G here, which grows the value in the mid- to high-end segments where Ambarella is focused. In the bottom bar graph, you can see the split between the enterprise of public market segment and smart home security with about 75% being on the enterprise side. The transformation of the market towards AI is accelerating with cannibalization of existing cameras with new ones with intelligent AI features. Cameras now combine both human viewing and AI-based software such as person detection, tracking, license plate recognition, et cetera. AI adoption in the smart home security market is just beginning and is perhaps 18 months behind that of the enterprise, but we are seeing customers' demand in smart monitoring features and companies like Ring shipping cameras with AI hardware acceleration, features include person detection and notification, vehicle detection, pet detection, et cetera. The enterprise camera segment is currently serving smart cities and infrastructure applications. The trend is towards making cameras intelligent. Some of the applications are shown here, for example, and include traffic management, accident detection, license plate recognition, finding missing persons and reporting unusual behavior. And the replacement cycle is accelerating as people demand these features. All of our customer discussions today and new designs are for AI cameras. And the ASP for these AI SoCs is approximately twice that of existing non-AI SoCs. Additionally, there are new incremental opportunities to grow the installed base and sensing cameras, for example, applications in smart retail and occupancy monitoring. Here, the cameras are purely for data extraction, for business analytics. There may actually be no video streaming or storage at all. And applications here include product placement, warehouse product tracking, business intelligence, people counting, social distancing and property management. The ability to do AI processing in the camera also provides the ability to preserve privacy by avoiding sending video to the cloud. This slide shows a few of the logos of some of the companies we are doing business with both in the enterprise IoT arena and the IoT smart home market. We are doing business with basically every major enterprise camera supplier. Customers such as Motorola and Bosch have been based on our solutions for over 12 years. And for some of the new opportunities I discussed on the prior slide, large incumbent security customers are aggressively moving into these markets. For example, Motorola acquired Openpath and Access Control, a customer that is using our CVflow applications. On the smart home side, we do business with most of the major consumer brands and these major players include Ring, Vivint, Alarm.com and Comcast. Now let's look at the smart home security segment. More homes are installed in security with seamless integration with other security systems and more cameras have been installed per home, driven by new form factors such as doorbells and outdoor cameras. The industry has been transitioned from pure video streaming and storage to analytics and people are demanding features that allow home monitoring without these false alarms and allow easier monitoring on their smartphones. For example, people and package detection and Vivint's new swimming pool alert are examples. The trend now is for analytics to move to AI-based solutions and new customer to use AI for presumptive security. For example, the Ring? Doorbell and Wired Pro 2 -- and Floodlight Wired Pro cameras use our CV25 for person and package detection today. And additionally, there are new form factors like Ring's Always Home flying camera where the AI processor will be required to do slam, object avoidance and path planning in addition to traditional camera functions. If we consider how the growth of the IoT camera installed base can potentially accelerate, you can see on this exhibit that the sensing camera opportunity can be more than twice that of security alone. There are now multiple new IoT applications being driven by AI. These include sensing cameras where cameras are used for data collection and decision-making and may not include human viewing at all. On the right-hand side, you can see a few of these new applications, all of which Ambarella is looking to address. And these include mobile payment, access control, robotics, new retail, health care, machine vision, smart occupancy, logistics and smart security. In the smart building category, we have been working with our partners on semiconductor, Lumentum on a reference design for 3D access control, recognizing faces and detecting spoofing. This leverages Lumentum's structured light technology and sensor fusion running on CVflow. We have also a battery-powered reference design for smart locks based on the Smart partnership -- on the same partnership and we'll be demonstrating that to you this week at CES. In future, AI will allow building automation, including anti-tailgating occupancy sensing. The occupancy sensing can be used by HVAC control and efficient meeting room allocation. Other new markets for us include robotics and machine vision. In home robotics, cameras are increasingly being used in robotic vacuums to enable them to become smarter. Our CVflow SoCs are ideal for these applications because they confuse data from multiple sensors. And in industrial robotics, cameras are again being deployed to make autonomous robots smarter. And in these applications, we can also leverage our CVflow SoCs and the technical developments from our Automotive business, including radar. And in machine vision and Industry 4 applications, cameras are increasingly performing AI at the edge in the camera, making systems faster, more reliable and more easily scalable. Here, we can leverage our IP security camera experience and technical strengths in image and AI processing. During the show, we will be demonstrating our beam counting demo here at CES, and I won't make any jokes about that being a favorite of our accounting department. Based on our success with our CVflow architecture, we've been building an extensive ecosystem of partners. These include hardware platform partners shown here at the top. We have been partnering with AWS to host our tools in their cloud and enable easy network development and deployment, and we have a wide range of third-party software vendors covering various AI applications, including security and robotics. Many of our customers, especially the big security camera makers, also work on their own software, all supported by CVflow's open architecture model. I'll now talk about the automotive market. So let's first look at the automotive revenue SAM by application. The total SAM here is growing from just under $2.5 billion in fiscal year '22 to close to $7 billion in fiscal year '28. The bottom 2 bar graph segments are for OEM car recorders and dash cameras, both of which are markets that Ambarella has served for many years and above that, our new electronic mirror and in-cabin monitoring segments, which include both driver monitoring and combination products such as driver monitoring plus interior monitoring cameras. Our automotive SAM looking forward is now dominated by sensing applications rather than viewing ones. The largest portions of the sensing SAM are for Front ADAS, Level 2+ and Level 4. These are markets that Ambarella is serving with our functional safety, CVflow SoCs, including our existing CV2 family and the new CV3. And lastly, we have radar, a market which Ambarella is now addressing with the acquisition of Oculii, providing both software and module solutions. For the first time, we are now adding the Oculii SAM to the Ambarella SAM. And by fiscal year '28, that's about $600 million, mostly software with SAM module business. This slide shows the trend in vehicle sensor suites and specifically the increase in the number of cameras and radar in vehicles with high levels of autonomy, as highlighted by the industry example shown. In the past, the increase in number of cameras per vehicle was the opportunity for Ambarella. But now with Oculii and with CV3, we are not only providing camera perception but radar perception and processing for fusion and planning. The average number of cameras per car across all vehicle types is increasing from approximately 1.5 today to about 3 by fiscal year '28. But in the high-end passenger cars, it will be in the range of 8 to 12 per vehicle and potentially more than that in Level 4 robo taxi applications. These numbers do not include cameras for electronic mirrors or driver monitoring or backup cameras. This slide shows a quick snapshot of some of our automotive camera logos. Our automotive business began with dash cameras, which became an OEM car recorder option. Since then, we have expanded into forward-facing ADAS design, starting with Chinese commercial vehicles and more recently fleet manufacturers. And more recently, we have won significant ADAS wins with new electric car makers and autonomous vehicle manufacturers. We have also entered the electronic mirror market with partners such as Gentex, the world's largest mirror maker and our shipping in in-cabin and driver monitoring applications, including our wins at Great Wall Motors and Dongfeng in China. I'd like to highlight a few of the recent design wins, showing our progress towards providing the central process and solution in these vehicles. In our recent earnings call beginning of December, we mentioned that the new Rivian R1T truck was used in our solution, but we can now also tell you that the Rivian R1S SUV and the Rivian vans are also using our solutions as well. Our CV2 SoCs are being used for the drivers plus autopilot and Gear Guard security features. The vans also use our CV2s for stereo forward-facing cameras. And as we announced earlier in the year, we have been chosen by Motional, a joint venture of Hyundai and Aptiv for its autonomous vehicles and Aptiv for the vision systems in its buses and vans. Ambarella's open platform approach allows our OEM and Tier 1 customers to differentiate. This contrasts with Mobileye's fixed vision IP bundle, which provides limited room for customers to add their own software. Our Front ADAS approach enables our customers as well as third-party software vendors to run their software on our open platform based on our CVflow SoCs. And these enable more functionality within the single box design due to their lower power consumption, allowing their use within the thermal limits of a single box. Our high-resolution processing provides longer detection distances and wider field of views for cross traffic, and we offer dedicated hardware support for stereo vision and optical flow. Furthermore, our platform enables a combination of different functions. For example, addition of driver monitoring, recorder and viewing functions and the ability to combine best-in-class software for multiple software vendors. It's important to note that vehicles will increasingly be expected to add software features over time, for example, to meet changing Euro NCAP requirements and our high-performance open platform fully supports these requirements. I'd like to show you an example of what we have planned for a reference design targeting the front-facing ADAS market. This builds on the synergies between Ambarella and Oculii and was not previously on the either company's road map. We are now one of only a few companies that have the advanced vision and radar technology under 1 roof. The Rebel II is a 1 camera, 1 radar Front ADAS reference design with very low power and embedded sense of vision -- fusion. It includes AI vision perception, Oculii radar processing and sensor fusion, all running on a single CV2 functional safety chip. The combination of Vision plus Radar Fusion enables superior range and scenario coverage and higher performance versus a vision-only system. Availability of the Rebel II will be in the second half of this year. We are continuing to build an extensive ecosystem of software and algorithm partners for each market segment. These include Level 2+ forward-facing ADAS, interior perception, including driver monitoring, electronic mirrors and AVM, amongst others. In the driver monitoring category, we will be demonstrating our partnership with the Seeing Machines later in the week here at CES. Of course, in addition to the third parties shown here, many of our automotive customers are also developing their own software in-house. I'll now talk a little bit about Ambarella's software business model. Ambarella currently addresses software value or capture software value by selling chips with different part numbers and price points while enabling software features via an electronic fuse lock in the chips. We've actually done this for many years and across all of our different market segments. We also sell modular SDKs, base layer of functions, middleware reference applications and with different operating systems, including Linux for security cameras and real-time operating systems for automotive applications. For certain markets, we sell mature software applications to shorten time to market. For example, in car recorders and electronic mirrors. We will be demonstrating our autonomous driving stack as presented by Alberto on our EV vehicles here in Las Vegas this week, and hopefully, you'll get a chance to ride in them. We will provide complete AD stack implementations as a reference for evaluation, benchmarking and test, and the stack contains significant IP, which we will be discussing with our customers. Some modules such as stereo calibration and AI-based ISP, we expect to license for mass production. With the acquisition of Oculii, we will now be selling RADAR software. This will be our first move into a pure software play, selling software that runs not just on Ambarella chips but on other companies' chips as well. We will collect NREs, receive per unit license fees as well as annual maintenance fees. And we will also offer Oculii radar modules for IoT and lower volume vehicle markets like, e.g. L4 and off-highway vehicles. And finally, I'd like to briefly address some of the other new automotive applications that Ambarella is serving. The first image shown here is a KeepTruckin AI dash camera for the fleet market using our CV22 SoC. The camera integrates one camera for Front ADAS and incident recording and a second RGB-IR camera for the driver monitoring system with driver recording, and we have this been demonstrated at CES this week. In the electronic mirror example, our CVflow SoCs are simultaneously processed in rear left and right camera images with very low latencies and high frame rate, while we perform AI-based object recognition and blind spot detection recognition in the left and right cameras. The driver monitoring solution shown here as Cipia, which leverages both our AI performance and RGB-IR processing. The truck images from the CoROS van where we automate logistics using our CV2 stereo processing to automatically read the bar graphs and paragraphs on packages. And lastly, we show the new Rivian rear guard security system. And one image that I didn't have time to include, I would have included a new product announced today that is Nextbase, which is the largest European dash camera supplier. They announced the Nextbase iQ product, which uses our CVflow SoCs, and it's one of the first new dash cameras to use AI intelligent process in it. So all of these new applications leverage Ambarella's significant technical advantages, including optimal system solution, leveraging our low-power CVflow SoCs and comprehensive software SDKs and applications, computer vision with industry-leading performance per watt and mature CV development tools and best-in-class video processing include an excellent imaging and low-light conditions, highly efficient video encoding and color RGB-IR processing. In summary, we see significant new opportunities in the automotive market and believe that we have a highly differentiated product portfolio of both SoCs and software to successfully address them. Thank you very much. I'd now like to hand over to Louis Gerhardy.

As a reminder, I'll be the last presentation, and then we'll go into Q&A immediately afterwards. And the title of my presentation is Building a Sustainable Model of Return. And collectively, what you've heard from all the speakers today is a description of this new foundation. I'll describe for Ambarella. And from a financial perspective, it's this new foundation, we expect is going to deliver more sustainable, more predictable financial returns than we've had in the past. And so our technology, as you know, it's always been differentiated. But what's changing is that there's just a lot more of it, less described the third generation of CVflow. Alberto described a more complex scenarios for the L4 stack. Steven's overview of Oculii and how the radar technology fits into our road map all the way down to raw data fusion. There's just simply more technology, more value per system that we can access in growing markets where, again, as Chris said, we have new product cycles in existing markets, but this CVflow sensing capability enables us to reach into many new markets, enable machines to perceive the world and make intelligent decisions. Again, this is the new foundation that we're talking about. And as Fermi said, the gross profit dollars per unit on a like-for-like basis, it's about 2x. And so this is going to be really critical for our ability to continue to drive positive operating leverage, and we'll introduce a financial model that will assume that, that does occur. And finally, Ambarella is in a position of financial strength. We have ample liquidity and a positive long-term operating cash flow outlook. I'm frequently asked, after a 30-, 40-minute meeting, portfolio manager will say, okay, that sounds great, what's the elevator pitch? What's this all come down to? What evidence can you give to us that it's actually happening? And I always refer to the average selling price. So as you all know, in the semiconductor industry, it operates in a deflationary environment, the deflationary characteristics every year, you do more and you get less. Well, I'll point out Ambarella's ASP is in the mid-single digits today moving higher over the last 2 years, but we're introducing products now, whether it's CV2 family or the new CV3 family with double-digit ASPs, and CV3 reaches into the triple-digit ASP range, again, for a company with ASPs in the mid- to mid-high single-digit range today, but going up for the last 2 years because of CV. So we'd suggest this rising ASP is indicative of the change and this new foundation, and we would expect that there's more to come. Going backwards. Quality of revenue, we put forth has never been higher. And if we back up for a minute, our revenue this year will be at an all-time high for the company, and that compares with fiscal 2016. So that's a nice stat, but it did take 5 years. So what's exciting about that and what's really important is the internals. The internals at Ambarella have changed dramatically. And let me give you a couple of examples. Our largest customer year-to-date, the first 9 months of this year, our largest customer is 7% of revenue. And in fiscal 2016, the largest customer was 30% of revenue. From a different perspective, our top 5 customers for the first 3 quarters this year were 23% of revenue. And back in fiscal '16, they were almost 60% of revenue. And there's another way to look at how this foundation is changing this more sustainable foundation I've been talking about. And that would be what's the underlying economic driver of the revenue that we're realizing today? And there's been a major shift in that as well, whereas back in fiscal 2016, heavily driven by the short product cycle, discretionary consumer electronic market. You know the applications I'm talking about. Nowadays, 90% of revenue is driven by 3 things: enterprise CapEx, public infrastructure spending and by consumer durable goods, a product that might sell into a smart home and get replaced every 5 or 6 years. So this is the foundation that we're talking about and the quality has changed and these internals are quite a bit different than they were before. So let's just cover the automotive funnel, I'll describe a few new things about this. First of all, why do we even provide an automotive funnel. We do this because we can get a design win in the automotive market, and it might take -- if we get a design win tomorrow, it might take 2 to 3 years to begin production. Furthermore, we're often not allowed to talk about those wins. And so we've created this analytical framework with a very disciplined and transparent model, we tell you exactly how we build it and what the discount factors are. And this way, we can communicate to you how things are changing at Ambarella and our Automotive business. And so on November 30 when we announced our earnings and we announced the new funnel, and we'll update it once a year, this funnel increased from $600 million a year ago to $1.8 billion, it tripled. And just to put that in context, what this means is that we expect in the 6-year period of fiscal '23, which is calendar '22, to fiscal '28, calendar '27, in that 6-year period, at this current time, we expect our automotive revenue to be $1.8 billion. So how did it triple? Well, number one, significant expansion in our global reach. And we've talked about in the last 12 to 18 months, increasing our sales and marketing, FAE support in Europe with an office in Munich and more people in Detroit, that's obviously had an impact. When I get to some of the funnel stats, you'll see that. There's also more content per win. I mentioned double-digit, even reaching into triple-digit ASPs now with CV3. It's a growing market. And most importantly, we think our share is going to be growing. And if we just touch on the share for a minute, we think our market share of the SAM in auto that Chris presented is a little more than 3% this year. With this funnel, this $1.8 billion, if you look at -- go back and look at Chris' automotive SAM chart and add up the SAM in the same period of time that the funnel is, that's about $25 billion. And so divide $1.8 billion by $25 billion, you can see we think our market share in that period of time is going to be about 7%. And so that's the way you can think about how we see our market share changing. What are some other important facts in this new funnel? Well, first of all, about 80% of this $1.8 billion is computer vision products, and so about 20% is human-viewing products, which is the market that Ambarella initially penetrated the automotive market. Number two, Americas and Europe now represent more than half of this new $1.8 billion funnel. And number three, in the back half of the funnel, the last 3 years, L2+ is beginning to make a material contribution to it. So with this new foundation, we've been talking about, what's the financial model going to look like? So first of all, in terms of revenue growth, the SAM that Chris presented for IoT and auto combined is in the high teens CAGR, high teens revenue CAGR, and we expect to take market share. So first of all, we expect to grow at those levels or above as we take market share. From a non-GAAP gross margin perspective, our non-GAAP gross margin guidance has been and remains 59% to 62%. We want to be a larger company, and we're going to price, accordingly. Yes, our gross margins are above that now, but we're comfortable staying in the 59% to 62% range. And then in terms of the operating leverage, CV, while it's early in the ramp, is a major driver of this positive operating leverage we're talking about. And we're projecting at $500 million of annual revenue, non-GAAP operating margins of 21% to 24% and at $1 billion, 30% or higher. And if you calculate the incremental operating margins on that, that's 30% or higher as we scale to these higher levels of revenue. And from a free cash flow perspective, as a fabless company, the most important thing for us is the human capital. You heard that from Fermi, you heard it from Les. CapEx, purchase of equipment is going to remain in that 1% to 2% range, which has very positive implications for our free cash flow outlook. With the changes in our business, we are going to report revenue in a different way. We will not begin to do this until May or June when we report Q1 of fiscal '23. So for Q4 and for all of fiscal '22, we're going to continue to report the 3 revenue segments on the left side of this page. But given that other has declined and some of the other changes in our business, we're going to put security and other together, call it IoT, non-auto IoT, and then we'll continue to report automotive as it always has been. We will continue to report subsegment information for IoT -- non-auto IoT, for example, enterprise, public security, the smart home security. But one of the reasons we're doing this is that there's a number, as Chris mentioned, many new emerging segments of the non-auto IoT market, where as they become more material in the coming years, we'll break those out as subsegments as well. A couple of examples could be access control, which we've spoken about in the past as well as mobile robotics could be another one, but there's many others. In terms of the ample liquidity I described earlier, after Oculii, Ambarella has about $150 million of net cash. The acquisition closed on November 5. But most importantly, this company has incredibly robust track record of delivering positive operating cash flow, 13 consecutive years. So it goes back before the IPO. And this includes a period of time that Fermi described the last 5, 6 years, where around $600 million was invested in computer vision R&D, and it really just started to become material in the last 1.5 years. So that contribution from CV, we think is early, and it bodes very well, we think, for our ability to generate cash. It's also important to point out the computer vision investment has been completely funded by internal capital markets. In other words, the video processors. So Ambarella since its IPO has not had any secondaries. It has not issued debt. The gross profit dollars have financed this CV investment. And what's left over has been deployed this way. So the accumulated cash flow that Ambarella has generated has been deployed to M&A transactions of VisLab and now Oculii. It's been deployed to stock repurchases, about 35% of it and then the rest to CapEx. So all of these arrows remain in the quiver or on the table going forward. And we'll continue to update our capital deployment plans as we move closer to those milestones of $500 million and $1 billion of revenue that we talked about earlier. So it hasn't always felt like it, but the stock has been doing well. And Ambarella's IPO was in the fall of 2012. And it's been the best-performing semiconductor IPO since that time, as the chart on the left of depicts and against major indices on the right, the ones relevant to us, Ambarella has also performed very well. So with that, we're going to stop for Q&A. I think we're going to do a slight reconfiguration of the room. Is that right? And you can get your questions ready. Julie will have a microphone. And maybe just give us a minute. Come on up, and we'll start to reconfigure the setup here for questions. While we're doing that, let's talk about the rest of the event. And immediately following Q&A, we're going to have a management reception right outside here. And you may choose to mingle with the management and ask additional questions or you can go see the demos. We'll have Ambarella host in front of the demo room, and they will take you through and explain what you're seeing. And some of you do have reservations for the EVA car as well. So please go out to the parking lot in the tents and wait for those. Again, if you don't have an EVA reservation talk to the front desk here and we can arrange it, if not today, later in the week, if you're going to be here longer. Do you want to kick over here because I have the computer -- I have to get the computer.

So can we start? So I think we'll start the Q&A. Yes, please.

Thank you for putting this together. Great to see everyone in person again. I don't know if this question is for you, for Fermi or Les, probably both of you. But you both mentioned today that your larger competitors obviously have different business models. And the way that you do things is with the algorithms approach first. Why do you think your largest competitors don't do that yet? I mean I understand they have different model -- business model. They leverage their R&D differently. But we're getting to a time now with AI where you would think that they start to catch on to algorithms-first approach.

Les, do you want to give a try? Come up.

I think a lot of it has to do with the DNA of the companies. NVIDIA's background is in graphics world, and they've been building GPUs since the beginning of the company. And so their mindset is always how can we extend the GPU to cover these new markets. And if you look at how much area even in their latest -- or in GPU. You look at how much area they go out to GPU functions versus dedicated AI acceleration. It's the vast bulk of the areas is dedicated to GPU functions. And having worked in large companies before, I know it's very difficult for a large company to make a fundamental shift in direction. So I think that's the basic reason.

I think -- I would like to add another comment here. I think if you look at those large companies, they already have established business based on GPU or app processors. And any time it's not only changing the silicon. When you change your silicon architecture, you need to change the complete software stack. In fact, for NVIDIA, they've been pushing that their software stack as a main differentiator from the business model. So actually not only to change the silicon but changing the complete software stack, I think that's a big challenge for them. Because still, 90% of the business is still inside the GPU application. So I think that's another reason for a big company, like Les said, it's hard for them to change the direction there where we are. Sorry, go ahead. That's okay. Next one. Okay. No problem. Go ahead.

This is an incredible presentation. And one of the most exciting things to me is how Ambarella is moving towards creating more software for -- or I guess, like decoupling like the software product from the hardware product. And one thing I'm really curious about is just how you're going to see -- like how will the unit economics of that work with respect to the ASP you're seeing today in CV3 or like CV2? And like how could -- how should we think about like the software modules as they're being sold with respect to ASP and perhaps would be a hypothetical customer or something?

Right. I think I'll take that question. I think the -- Louis talked about the CV2 was like a double-digit ASP and the CV3 can be triple. But I think that's all still on the hardware side. On the software margin, when you license, you usually need to attach different price online. For example, when we acquire Oculii, the RADAR processing, we talk about from the $3 to $15, depends on the performance. And you should expect similar things like, for example, we talk about we license our blind-spot detection to our email customers, which caused them probably a few extra dollars. So it really depends on the performance and uniqueness of the offering, right? So I don't have a [ suppress hike ] for everything yet, but you should expect that there will be extra price associated with each software modules.

So I'm trying to wrap my arms around the domain controller kind of opportunity. I guess going from CV2 to CV3, does that expand? And is each domain won by 1 person? Those guys might win one, you might win one. Or how does that kind of shake out? I'm also curious about kind of the new car guys that in why they're liking you versus some of the existing.

Les, you want to?

I think you should stay here. So I think what you can say right now is it's quite a wide range of opinions in the OEM space about how to construct their central domain controller. I think in general, people are moving to a central domain location that's running all the software. Exactly how many chips and how they partition it is going to vary from OEM to OEM.

A question online related -- so Les, you can stay there. So with Ambarella's heritage and video and imaging, there's companies like Tesla that have developed their own chip. And how would you compare and contrast what Tesla has in volume today with their architecture and just the competitive dynamics between what you talked about today with CV3 and what Tesla has?

Right. So I'm going to assume that's from a technology perspective, right? So I think Tesla has said that their current chip is around 70 TOPS of performance. So if we use that as a metric, you can compare that against the 500 TOPS -- eTOPS number that the first CV3 chip will have. And also, I think our performance is higher. So in general, I believe our chip is significantly higher performance and more power efficient than the Tesla offering. And another differentiation is the -- unlike Tesla, we do believe in radar and the -- especially high-definition radar as that being integrated into the sensor suite is going to be very important to have a robust true anything, let's say, beyond L2+ level of time you need more robust sensor suite.

I had a follow-up on CV3. I think, Les, you mentioned that the EVO car has 16 CV2 processors, but it's set up to be retrofitted with CV3. How many equivalent CV3 chips will you need for that? And then maybe this is a question for Fermi. You've got obviously some pretty high profile front-facing Level 2 wins with some automotive disruptors out there. How much of your time these days are you focused on these EV disruptors versus Tier 1s and traditional OEMs? And going forward, is it just a timing issue? And can we also see -- expect to see traction with some of the more traditional automakers as well?

Okay. So the first one is an easy answer, it's 1 chip, 1 CV3 high chip we'll be able to go on the entire stack.

Well, I think it's really a different time line we're talking about. When we talk about electrical vehicle announcement we made, in fact, we start engaging with Rivian in late 2018 and the 4 years later, we talked about production. And recently, if you talk about how I allocate my time, in fact, in combination with our office in Detroit as well as in Europe. In fact, in the last 2 months, I travel there for 4 times, all during the pandemic. That's just show you that there's activity, and we need to continue to show up to our customers, and that's definitely one major area I spend a lot with our customers. Please...

An online question, if I can?

Go ahead.

The question is for L2+ and above, how do you see radar and LiDAR competing or cooperating at L2+ and above?

Yes. So I think for L2+, you can get away with almost any type of vision plus maybe nothing in the case of Tesla, right? So that's because the human is always there to take over control if something goes wrong. But as soon as you go to L3 or above, and the eyes are off the road, you need to be very sure that the sensor suite is not making a mistake in the perception. And you need to augment vision with at least radar. Whether you still need LiDAR is something I would say is still an open question in the community. But from our point of view, we believe that with the Oculii radar technology running on CV3 that we will not need LiDAR.

Go ahead.

My question is around computer vision as a percentage of your sales. So you said 45% of your sales will be computer vision in calendar year 2022. That implies about 115% growth rate off a very high growth rate last year. So wondering how sustainable that is in that? Does that incorporate cells from Oculii being integrated and the new CV3 chip? And then secondly, when you scale to $1 billion in revenue, what percentage of your sales by then do you assume it will be a computer vision?

So for next year revenue, I don't think CV3 will contribute on that, maybe a little bit sampling and so on, but probably not a material revenue. On the Oculii side, this year, we talk about $3 million to $4 million revenue, and we expect the growth there, but it's not going to be jump up to a level that we just talked about. So majority of the CV revenue growth were coming out from our CV2 family chips next year. And we do believe that the growth will continue to be there, but the ratio is hard to maintain 100% every year. But definitely, I think if you look at the other set of data, we just offer you, we have 270 customers. 150 -- 100 of them already in production. I hope that majority of the other were going to production soon and each customer will introduce more than 1 product. So I think with that, I'm still confident that the CV revenue continue to grow in a very healthy way. Yes.

Two questions. First, since we're in Las Vegas, I'll start with a fun question. Can you give us the over under on when you think you'll hit the $500 million and $1 billion revenue model?

I think you look at the growth of the SAM, we said we expect to do that plus by taking market share. So I think that's the closest we can get there for you.

Okay. I didn't think I'd get an actual answer. The more serious question is as you look at introduction CV3, how should we think about that expanding your TAM? Was the domain controller already in your SAM numbers, and so CV3 allows you to come in and just take a higher share of that SAM? Should we think about it as really new incremental dollars now coming into the company? And sort of, again, what would the time line be on material revenue from CV3? I mean my guess is that's 3 to 5 years out, is that sort of a reasonable expectation?

CV3, in terms of timing for revenue, it could be as early as second half of calendar year '24. And in terms of the incremental SAM, if you go back and look at Fermi's -- he had a pyramid that showed where we are today in that pyramid, which depicts basically L2+ or L4 software stack and providing camera perception. And that is what you know Ambarella for today in the automotive market. Now with Oculii, we can go horizontally, as Fermi said, and do radar perception as well. But CV3, what that brings in addition to the incremental perception opportunity with the radar, it also allows us to go up the stack and capture the fusion and planning layers as well. And perhaps Les can talk about the amount of processing required to do fusion in a car where you've got so many sensors and so many HD camera and radar sensors. But just to answer the dollar question before Les does that, if you look at the SAM that we've provided and we've added 2 additional years relative to what we've had out before, look at how L2+ begins to take off as an opportunity for Ambarella and the higher layers. And while we can do camera perception, like we have with Motional and Arrival and Rivian, there's an incremental opportunity now to do that fusion layer and the planning layer, and that's a big part of that incremental L2+ opportunity. So maybe Les can describe now the incremental processing needed to do fusion and the planning.

So if you look at the stock, it's actually a pyramid also. Most of the processing is actually at the perception layer, but when you're doing all of those sensors on 1 chip, that's a big multiplier factor, the 10 cameras, 5 radars, 15 high-bandwidth sensors coming into the chip, choose up a lot of processing. And then the next layer of the fusion layer is also a significant amount of processing, which will leverage the CVflow engine. And the planning layer -- current planning is actually running primarily on the ARM processors. And that's one of the reasons we beefed up the ARM processing a lot in CV3 was so that we can run all the traditional kind of planning software, which requires a lot more ARM performance than what we had on CV2.

Just had a quick follow-up. I believe you said it was going to be [ $600 million ] by fiscal '28 [indiscernible] software. How much of that would be pure software versus software coming from your own chips?

The question was for the -- repeating it because it broke up, of the $600 million of incremental SAM from Oculii that we showed for the first time, how much of that in our model is software versus their module business? I believe, is that right?

Right. And then how much would be third-party paying software license?

And how much of that would be third parties paying software license was the second part of the question? So of the $600 million, a majority of it, say, $450 million or $500 million is coming from the software licensing. And we've been very conservative in terms of modeling the radar module business and how it will ramp. But there are some interesting opportunities out there, and we'll just have to keep you -- we'll give you some updates as to how that business is progressing. We've mentioned the interest in the security camera market, where obviously Ambarella has a lot of good customers. But most of it, to answer your question, is on the licensing side, and there was an online question, I'll go ahead and tack on to that, which is what is the range of ASPs for your radar licensing business. And that can range for the Falcon from $2 to $3 per unit up to $10 to $15 for, say, an Eagle, and we haven't talked about Raptor yet.

And then we haven't talked about with the radar, obviously, running on CV3 with much higher performance in the resolution, the price will be different, too.

I had a question online from the -- regarding the go-to-market Chris talked about for software. And is there a recurring revenue model in here for Ambarella, do you see?

Well, I think being in a semiconductor industry for 30 years and many of our competitors have tried that before, haven't seen one being successful charging recurring revenue, and I don't think that is part of our business plan either.

Yes, Suji.

I was just curious for Dr. Broggi perhaps. This HD map, all the data you have there, is that a proprietary Ambarella format? Or is that standardized? And can customers populate that information versus you guys? And I guess, I mean, does that help -- does that level of data help, for example, faster highway driving, I'm just curious?

Please.

Actually, right now, it's our own proprietary maps. And we are not thinking about doing any -- we are using that as our own stack. So we haven't been doing any homogenization without any other software sources and sources of [ MEMS ].

Where is the other microphone? Julie, do you have it. And while you're taking the microphone...

Hold on, hold on. Les has something...

Let me just add one thing to that. I think the map format is something that's very easy for us to adjust if there was a reason to use a different format. We're really focused on the core technology right now with how to generate that information and how to use it inside the stack.

Go ahead and pass the microphone, I'll ask this question. When will the presentation be made available? Shortly after this meeting is over. Next question is with regards to radar specs, NXP and TI are currently limited to 12x16 or 16x16 channel resolutions, a recent spec described 48x48, where does Oculii fit into all this?

Steven, do you want to -- Yes, go ahead.

That's a great question. And in many ways, I think you're seeing a lot of the asymptotic returns on the hardware-based solutions starting to top out. When Louis referred to the 2 numbers there, the first number is number of transmitters on the system. And the second number is numbers of receivers in the system. And those 2 numbers multiply together, give you the total number of MIMO receivers. And already, what you're seeing is that even as you go from a 12x16 system, which is really where TI, Infineon and NXP are capped today to -- there's other companies doing a 48x48 system, Mobileye, in particular. The amount of compute that's required to do that in a brute force way is already going up into the tens of TeraOPS just for 1 radar module to process that many number of physical active channels. That's not to mention all of the number of ADC channels, power amplifiers and all the analog circuitry that you need to enable 1 module as well. So obviously, that cost, size, power increases a lot. But if you look at the specs and the resolution numbers of what these companies are claiming even with 48x48 systems, there's still about 1 degree by 1 degree, which is actually what our system that only has 6 transmitters and 8 receivers is capable of delivering today. You guys can actually see a demo of that. The module is about the size of a postcard, only 6 transmitters, 8 receivers and it gets the same amount of performance as a 48x48 system. So our software effectively allows you to achieve about a 50x performance improvement on a 2-chip design today. And on the 4-chip designs in the future, we're planning about 100x improvement. And that's not to mention what we're going to fuse on the Ambarella SoCs to enable even higher levels. And so again, this software-based approach is completely scalable to actually complement any other system that has more antennas. As you see, we have 3 categories of products we have a Falcon, Eagle and Raptor. Those have increasing number of antennas. But we really think that the sweet spot is going to be fewer number of antennas because that means smaller size, lower power, smaller form factor, lower cost, and that is what's going to be mass scalable with the right software, I think, with CV3. So...

I just had a short follow-up on the Map question. I was just curious like where does all the map data gets stored? Or who is responsible of storing the Map data, is it like currently Ambarella just like locally looking at pharma and storing the maps there? Or like is it -- or is this like this data service going to be available to customers? Or how should we be thinking about that?

We haven't been optimizing that. So the map data is now restoring to the car. Actually, we have a big server in pharma. We have all the maps. But when you drive, you can just download a little portion of that. But again, we haven't been optimizing that. So yes.

I think the kind of high-definition map that we're using is actually quite efficient compared to the kind of dense point cloud maps that are used by people at Google because it's -- as you could see from Alberto's talk is primarily landmarks plus lane information. So it can be stored on the car for fairly large areas. It could be streamed from the cloud as well. The infrastructure for handling that is something we would expect our customers to deal with and not us.

There's an online question about CV3 derivatives. Can you talk about what those might be? And what are the adjacent markets that those would serve?

So I think just like CV2, CV3 will be a complete family of chips, and CV3 highlight and less introduced today is the higher version of the family. And you can imagine that CV3 is capable of doing Level 4 type of autonomous driving. And you can imagine the performance requirement for Level 2, Level 2+ and Level 3 probably are different. And we should have a family of chips, a different performance and therefore, a different price points to address those key applications. Based on the same number that Chris presents the 2 biggest markets moving forward is ADAS, Level 2+, and also maybe in the future Level 3, Level 4. And you should expect us, we have separate chips share the same unified software SDK and to serve our customer, but that's for automotive. You should also expect the CV3 architecture will have a separate chip for our other markets, which will require higher CV performance, particularly with today's introduction of new network-based IDSP. And you can see that the quality difference at a very low life situation, which is critically important, not only for automotive driving, but also for security camera. And we expect that more people will -- or more customers move to the this kind of direction, which require the solution we provide to them with a higher new network performance, and you should expect that there is a derivative chip from [indiscernible] family, we'll address that market, too.

There's another online question saying that you made it clear about GPU compute -- comparisons, but could you describe how CV3 compares with more application processor type of approach?

Sure. We have done comparisons with both Mobileye and Xilinx architectures. And in each case, we saw kind of a large advantage in terms of power efficiency and overall efficiency. So I think it's -- like I said, it's easy to have TOPS, efficiency is hard. And so even some of these other so-called application-specific approaches not necessarily that efficient. You really need to have a combination of silicon expertise, algorithm expertise bring those 2 things together very tightly.

[ Quin ]? Yes, [ Quin ]?

Just a clarification, probably more in the $1 billion model than the $500 million model. You talked about the software module opportunity, which is a longer-term opportunity, you've got the Oculii software. Is all of the software revenue fully incorporated in the gross margin targets in that $1 billion model or to the extent that you start to recognize software, especially from autonomous driving modules, would that potentially be upside or change the margin model over time?

Yes. So I would describe 2 software businesses. There's an existing one that I think Fermi described as lower in a stack, and that's in those numbers. It's in the SAM, it's in the funnel. But then there's the new emerging software strategy that is not in the SAM that is not in the funnel. And should that update our long-term financial model, that's just something we'll have to keep you up to date with.

And also, I want to point out that software licensing is not 100% gross margin. John help me to understand that there's any associate investment need to be accounted as the cost of the revenue. So although I think still that software revenue, gross margin will be higher, but not dramatically higher than our current corporate gross margin. Any other questions?

Let me check online one more time. If you could reiterate the time line for CV3 tapeout and commercial deployments? And what markets and applications do you think would be the first to use CV3? And also, sorry, what is the size of CV3 opportunities for Ambarella versus CV2?

Okay. So CV3 high already tape out, thank you for our [ VLSI ] team and work hard to make it happen. And we are planning to sample the silicon and our first-generation software, probably, I would say, second quarter next year, and the production will take a while. And because that -- meaning that we need to integrate to our customers and systems and software development and qualification, I think a minimum of 3 year or maybe even 4 years progress.

Second quarter this year.

Second quarter this year. Sorry about this. This is January 4, third already. So I think that in terms of the size of CV3, I will say a lot higher than CV2. In fact, if you look at the same number or the sales funnel we talk about, I think Louis mentioned that more than half of the funnel is based on the Level 2+ and Level 3 application. And so I will say that over time, when the CV3 become mature product I think CV3 will contribute a lot more in combined than the CV2 family as a whole. So I think that's definitely its expectation not to mention is like when all the magnitude higher, which all help us to build up the CV3 revenues.

Two additional online questions, 2 different ones. So in 5 years, what do you expect the CV versus video processor split to be and then the same question from an auto versus non-auto markets?

I think in 5 years, I think majority, majority, vast majority of our chip will be CV-based. I don't think about the video-only solution can survive because majority of our customers are demanding our CV solution, fewer and fewer video-only product being required by our customers these days. So that's part from the auto and non-auto, I think that I will be very disappointed if our auto revenue is not bigger than the non-auto in 5 years. And in fact, with based on the same number and with the activity we are seeing, I think that our auto revenue will be higher than auto in 5 years.

Just one last one. A lot of different questions on Mobileye. Maybe if you could summarize how CV3 compares to Mobileye? And which Mobileye products would you compare to? And what applications is there overlap in?

Right. Well, I guess closest product that you can compare to would be iQ6. I'm not quite sure what the public information is about iQ6. But from what we have heard unofficially CV3 is well above it in terms of performance and capabilities.

Any other in-room questions or -- pretty much tapped out online. Yes, Julie, are you...

I have a question. In terms of performance CV3, would you -- it's a lot of higher performance we're talking about, is it a 10% or 2x faster? What are the metrics?

Yes. Let me repeat the question. So the question is -- correct me once if I don't get it right, CV3 performance...

How much better is it in the iQ6 or IQ Ultra, whatever it is that you're talking about?

Yes. It's a little bit difficult for me to answer that because I don't know what information that hasn't been made public yet. I can't give you a direct comparison, but let's say there's a class of products which are aiming at around 100 TeraOPS kind of processing level. And if you compare that to CV3 high, that's a 500 TeraOPS chip. And also, I think the ARM performance that we have is well above what they have. I think traditionally, as part of Intel there, Intel always want to sell another Intel alongside of the Mobileye chip. And so I kind of handicapped how much CPU performance they were able to put in there.

Additional question. Is Ambarella consciously moving away from the volatile -- the more volatile consumer discretionary markets you'd call it other that may not upgrade to CV longer term? What is your strategy in that market?

I think that market will move to CV eventually, but we did make a decision not moving away but defocused on that market because we know that the size of the market and we know the -- there were only one dominant customer in each market, meaning drone or sports camera. And our decision -- the decision we made is really about that the market itself is not enough for our growth so that we need to find a new market that can leverage our expertise in the product line, which also at Santa have a much bigger base in terms of customer and the revenue growth, and that's a decision we made. So for consumer market, I really think that, eventually, those companies need to move to CV because I just don't see how -- they can add a lot of value with the CV functions, but that's not our focus areas anymore.

What are the -- this wasn't really talked about today, but one of the most, I guess, interesting things about Ambarella is just how many iterations of chips you're able to push out throughout like a year and throughout 5 years. And you mentioned that CV3 will be backwards compatible with CV2. And as all of these CV2 chips are now entering production and customers are using them, how fast do you think they'll need -- or how much time do you think they'll need to be able to move to CV3? Because that's like a 10x ASP increase.

Right. Well, obviously, you need to write sounds software to take advantage of new hardware that we add to CV3, so it will take time. But more importantly, it's really our customer need to go through the production cycle. We're talking about automotive customer. We're not -- there are really safety concerns. And I think software development is a portion of it, but more importantly, the back end to do production testing. That's probably taking even more time. So my opinion is it's going to be very fast for us to demo the capability of CV3. But for customers going to production, that's a different question.

Andrew, we'll get to you next. Question for Fermi. What's the biggest risk to heading the $1 billion bogey in revenue?

Well, I think the biggest risk is always that we don't know what our competitor is going to do, right? So in my mind, we know exactly where we're going, and we need to grow fast and making sure that we keep our advantage. But we cannot control what our competitor is going to do, and that's definitely always the biggest worry for me. There's a question.

Sort of the same question but asked maybe a different way. I guess what's the bigger risk here? Is it competition? Or is it continued inflation/component shortages lingering or maybe even kind of this ongoing geopolitical risk with China that might be argued to be in the stock at this point, but things could still escalate?

Well, I think there are a lot of risks you talk about, right? I think some of them short term. For example, I really don't believe this shortage problem will last another 2, 3 years. That's just not possible in my opinion, so it's a short-term problem and also that the geopolitical situation, although it's going to last. And -- but if you look at our revenue composition, it's kind of derisked to a point that I think is less risky for us. I really think that now we are focusing developing automotive products. CV3 is a great example, and that we need to focus on. Now we have engineering development going, and we need to focus on the business development and making sure that we talk to all the key customers, which we do, and convincing that, although we are a small company compared to those big competitors, but we are capable not only just in terms of technology, but we are ready to support them as a partner. So I think that's where we're going to focus as more effort on in the next couple of years.

Online question on the driver monitoring market and how you see this market evolving and which software partners do you work with for driver monitoring.

Right. So I think driver monitor, the consumer level driver market continue to happen as an aftermarket. Now we see the driver market move to become OEMs. And in fact, today, at the OEM level, we provide a complete solution. But when they move to CV-based things, we are counting on our software partner. Like we have multiple software partners that we list in Chris's presentation, we're counting on them to deliver ADAS-type of a software stack for those customers. But I also want to point out there is another market related to driver monitor, it's really fleet management. That fleet management at the beginning is just putting a driver monitor to keep safety. Now they found out that they can integrate with their current telematics system, and that combination become very strong. And for those customers, most of them continue to give them a software SDK and giving some perception level software, but they are using our third-party software, customer software to finalize the product. So for that recorded business, for the aftermarket, for the OEM business and fleet management, we are doing our standard SDK solution, but we're counting on the third-party software partners to support, and we have multiple of that.

Yes. I would add from the SAM perspective, the SAM numbers you saw for non-IoT as well as auto are completely refreshed. But the overall auto with the exception of the additional 2 years didn't change much but with the exception of driver monitoring, where we did increase our estimate for how rapidly that market would come by, say, calendar year '25 or '26. So we do have higher expectations in our SAM for that market. But in terms of the overall size, as you can see from the stacked bar chart that Chris had for auto, it's not the principal driver of our automotive business. By calendar year '25, I think it was a green color, really takes over in terms of driving the largest part of our automotive SAM, and that's where CV3 really kicks in. I think we're good online. Anything else here? Last chance.

Oh, there's one more.

Yes?

Thank you for the presentation, guys. So if I think about your competitive position relative to your peers, obviously, you're lower power, you're not a black box, so you give the OEM more ability to have control if they want it. But relative to some of your peers, where would you say that you think that your -- you might have holes in your portfolio or your IP might be lacking? I know, obviously, CV3 addresses a lot of those things, and that's a big announcement today. But just in terms of infotainment or mapping, if there's any areas that you could talk about where you want to expand your IP and your advantages there.

Well, maybe Les can address this infotainment question better than me. But I just don't -- a lot of -- some of our customers -- our competitors are talking about combining infotainment system with ADAS and other safety functions into one chip, which I just don't think that's possible but not because of the business reason but technical reason. Just that considering to do a Level 2+ car, how much performance you need to jam into a silicon and how much DRAM bandwidth you're required to do that and you try to add another function which would require a huge amount of DRAM bandwidth to do that, I just don't see that it's the right partition from a technology point of view. So Les, you want to...

Add one more comment. I think the other big problem with that kind of architecture is how do you ensure that the hard, real-time processing that's being handled by this L2+ stack is not impacted by the infotainment processing. And it's very difficult to do that when they're both sharing the same DRAM, and DRAM is usually the ultimate performance bottleneck in the system.

So in fact, in my presentation, I highlighted this real-time application, which I think is something we're really proud of that under any other circumstance that our product will deliver the performance. But in many other architecture because they try to share all the things, although they're kind of a high performance on every possible aspect. But when you put it together and when they run together, you don't get the high performance on any aspect. And that's a result of sharing resources, particularly on the DRAM bandwidth.

Yes. I think this relates to the question on application processors as well to some degree because some of those competitors are trying to attack every single domain in the vehicle. We're talking about the active safety and autonomy domain, period. We're not selling this into infotainment or body and chassis engine control. We're not -- that's not what this is about. So we're focused on these high-bandwidth applications, video and radar and now fusion and planning. So that's a big point of differentiation versus that application processor approach.

Any other questions? And with that, I would like to thank all of you attending the meeting today in person. Also, I thank all the people participate online. Thank you very much and looking forward to another CMD days in the near future. Thank you very much.

Management reception, management side after this and the Ambarella hosts in front of the demo room to give you a guided tour of what we have. Thank you.

Thank you very much.

Oh, sorry, one more thing. As you check out -- thank you, Eric. We have a little bag of souvenirs for you. And so please, as -- before you leave, check with Julie in the pink turtleneck here to get your souvenirs. And there's a Capital Markets Day cheat sheet in there, too, that you might find helpful. All right. Thank you.

Thank you very much.