Let's get started on this afternoon session. I'm Vivek Arya. I cover Semiconductors, Semi-cap Equipment at BofA, and really delighted to have Fermi Wang from Ambarella, CEO of Ambarella, join us this afternoon to share his thoughts about the company. I'll go through my Q&A, but please feel free to raise your hand if you want to bring up some of your questions in between. Welcome to you, Fermi.
Thank you, Vivek.
We appreciate you joining us, and maybe for people who might not be familiar, right, with Ambarella, if you could give us a quick overview of your kind of growth strategy and product pipeline, that'd be very useful.
Right. you know, today, if you look at it, our technology is really applying AI to video content, and particularly focused on edge devices, because our differentiation is really performance per watt, and which is critically important for any edge device. Doesn't matter whether it's security camera, automotive, for those applications that not only you need high performance, but you need very limited power consumption. That's where we focus on in the last few several years, and our CV revenue, our AI revenue go from, you know, 45% total revenue last year to close to 60%. That's really our core business these days. We're going to continue to focus on these two market with one market called IoT, which is dominant by security camera, and the other side is our automotive business.
The current revenue ratio before this inventory correction was roughly 3- 1, 75% on the IoT side, 25% on the auto side. Like I said, the technology differentiation is the only reason that we can continue to compete in this both space that we're competing with a company like NVIDIA and Qualcomm and some other smaller companies on the low-end side.
Got it. Thank you. Maybe we can start with a few kind of near- term, get those out of the way, and then really focus on the longer- term Fermi. Just in terms of the near- term, what are you seeing in terms of demand for IoT and automotive products in the current quarter and as you look into the back half versus the assumptions you had at the start of the year?
Right. First of all, the demand is still strong, especially on the IoT side. For example, I heard that Mobileye was here just this morning, give a. Their business, based on the video product, continued to grow, I think, 15% a year, that's consistent with our expectation at beginning of the year, that didn't change. What's different is really that our revenue from Mobileye, definitely is we're seeing a downturn compared to last year, that's basically because of inventory correction. That gap between their demand and our shipment, that's where we see our revenue weakness. We see the same similar effect among almost all of our IoT customers. For auto, that kind of impact is much less, but still there.
I see. Is there a way to quantify for me how much IoT inventory is out there in terms of number of weeks or months or?
It's hard because every customer, for example, certain customer, they are willing to share with us their data. We know that, but, you know, we have a lot of customer out there. We have 20 meaningful major IoT customer, not all of them willing to share with us their inventory situation, so it's very hard to generalize it. The only comment we gave before was we continue to watch our booking number. If when the booking number start growing again to a healthy level, then I will say, call it the sign of that, the end of inventory correction, but we are not there yet. Other than that, we haven't really give other prediction, except that we believe our CY revenue is the bottom, based on the current inventory correction situation.
I see. In terms of automotive, how do you see the trends there? Do you have inventory in that end market also, or how would you know, look at that specific market?
I think for that market, the inventory level is much lower than IoT. We do see some pushout demand requests from our customer and, but the size and the scale of the pushout is much lower. I would say that auto side haven't been impacted as severely as the IoT market with for the inventory correction.
I see. Can you maybe remind us, what are your key focus areas within the automotive today, and where are they going in the future?
There are a few areas that we definitely focus. The biggest focus is how to monetize our CV3 investment. You know, Right now, the 60% of our AI revenue all from CV2, and we invest heavily on CV3 for several years already. How to monetize that technology, most critical time. For automotive market, we made two major announcement. One is we announced with Conti that one of Level 4 customer using our silicon and the software that Conti and us co-developed to going to give design win. That's our first design we announced was our silicon and software, so it's a really major milestone for us. The other one that we mentioned in our script this time is we believe our CV72, which is a CV3 derivative chip, will become.
Well, we are using CV72AQ to bid on several Chinese Level 2+ design wins. We believe that we have a good chance to win some of them. Most of those project are on the schedule to be in production at the end of next year. You know, the design cycle is 18 months versus-.
Right
Four years in any other area. That's where I really think that, we will monetize our CV3 technology in China automotive market first. That's two major area we are focusing on. In addition to monetize CV3, we also continue to make sure two things. One is we want to make sure the software, including our VisLab software, AD software, plus our Oculii radar solution, will run on our CV3 chip, and that we, what we want to demo this year. That would be first time we demo a single chip solution with our own silicon, our software, and I think that's a meaningful milestone for us, and that's something we continue to develop for. That's another thing we should talk about for the automotive market.
You got it. We'll come back to automotive. On the IoT side, I think you mentioned that you are under shipping demand by 2025, you know, 30% or so. When do you see that kind of coming back? When do you see kind of sell-in and sell through start to match? Is it possible this year, or do you think that's more likely to happen in 2024?
Well, I think it's possible this year, but we're not coding for it. We're not saying that when is that going to happen, because I think every customer of us will have different timing.
Mm.
Some customers start the process to clean up their inventory earlier, they'll probably finish earlier. There are some customers just started, probably, I would say, a quarter ago, that will take it even longer. Overall, it will be average out. I think that. I'm hoping that for this year, but we never give a guidance about what's exactly that happens.
Of course. In that IoT, how much would you kind of call industrial versus consumer IoT? The reason for me, I ask the question is that, you know, sometimes you get, like, a seasonal rebound, right? In the second half or anything that is consumer-related. What is that industrial versus consumer mix, and is there a seasonal benefit that we might be able to see?
Right. For the IoT space there, the biggest portion is a security camera, and our professional security camera versus consumer is roughly 2 to 1. That's the ratio that we're looking at right now. You know, historically, that Q4 for consumer IP cam, there's a definitely seasonality, but we are not sure whether it's happening because when you have inventory, the seasonality may be our customer will use the inventory tool for the season if they have plenty of that. We are waiting to see, and we should see the very soon, because if they want to really have a season for Q4, they probably need to stop sending PO in June, July timeframe so that they should have a. They can have a chip for the last shipment.
I think that's where we're waiting for to see whether there's a upside or, you know, normal seasonality this year.
I see. On the Automotive side, how would you describe the competitive landscape? You know, one aspect of this is that, you know, you have very innovative products coming out. The other aspect of this is that, you have these really large competitors, whether it is Qualcomm or Mobileye or NVIDIA. I think it will help to understand. I'm sure the competitive dynamic is different versus each of them, so it'll help to understand what it is.
Right. I think the biggest problem we run into in this market is just like because our scale. That's the reason we really, at the early on, we focus on to bundle, to partner with the big Tier 1, like Conti and Bosch. That's a strategy because we know that if we go to bid on a project with one of those big Tier 1s, that solve, kind of address that issue. Your question is really about how we compete with each one of them. For NVIDIA, I think it's clear that the graphic engine just too power hungry, and that the performance per watt is really their biggest problem. The product they are shipping right now, need a water cooling.
I think from that point of view, to get a performance per watt right for the automotive industry, it could be. That's our biggest differentiation. For Qualcomm is similar. In fact, I would say that we have been competing Qualcomm with IoT space, too.
Right
... 10 years now, the reason we can win is the same thing. You have better video quality, better compression, and also better power efficiency. That's where although that Qualcomm many times tried to sell the AI processor into our IoT space, they didn't get much of a momentum. I think the similar things what you are seeing in the auto space, too, is really that while you try to repurpose AI processor, which was not designed for the video AI application, the inefficiency is very noticeable. For Mobileye, definitely, they have a strong momentum on the ADAS market, and they really dominance market by a black box approach, right? Customer doesn't even need to understand engineering, or they sell the whole solution. It's really just manufacturing.
That's why, for people who want to guarantee the production time, they choose Mobileye, because it's guarantee of go to production. However, for most of the Level 2+ OEMs, they want to have their software. They want to at least control portion of the software, and that's not viable. At least Mobileye is not offering some viable solution on that. I think from that point of view, when we compete with the Mobileye, it's really about whether if customer really want to have their own software stack, or they don't want to be, they don't want to use a black box solution, that those kind of approach are usually bid on by Qualcomm and NVIDIA and us.
I see. If I'm NVIDIA, I'll say, Look, I have, you know, somebody like Mercedes partnering with me, right? They understand the specs, and they are comfortable with the power consumption. Why isn't that a, you know, good pushback to the competitive-
Yeah. Yes, but however, its water cooling system, the cost is high, the complexity is high. Mercedes might decide that they can take that into production, but that's for high-end chip, high-end product. How about for low-end? If you know, low-end cars, you still need to have water cooling system, that obvious become a problem. I think for high-end car, okay, I can tolerate that, but how about for your whole stack. Power consumption, power efficiency is important. You know, you just cannot tell people, Continue to for all the car, you put another water cooling system for that, right? I think, yes, I agree with you that on high-end car, you can get away with that, but you may not get away with that on the, you know, in general.
Got it. Then, who is the decision maker usually in the process? Because, you know, when I talk with, you know, NXP and others, they will say, Well, we are also designed with, you know, Continental or-
Yeah
other Tier 1s." What really matters is, has an OEM actually put purchase orders, right? The Tier 1 will have relationships with many different suppliers, right?
Right.
It's not exclusive-
Right
With one, because that's their job, to have as many options as possible. How do you take that relationship and get the comfort and visibility that it's gonna translate into X dollars of revenue?
Right
in some predictable time frame?
I agree with you on that. Particularly for Level 2+ design wins, OEM want to make the decision. That's why we continue to try to make a relationship with that. Even in China, we talk about this as a CV72 design wins. We have to go to OEM to convince them.
Right
hat this is the right chip. However, you need to have a Tier 1 to provide solution to OEMs, and we are ready to go into production. You have to get both, but decision is done by the OEM level. For other business, it's a little bit different. For example, for the ADAS, today, most OEMs say, I'll just buy from any Tier 1 who can give me a production-worthy solution." For the e-mirror or for the recorders or even for the DMS or radar.
Right
Tier 1 is driving the decision for the solution they're going to propose to OEM. Obviously, Tier 1 have multiple silicon supplier. They can provide all of that, but you have to build a relationship with Tier 1 to make sure you have a solution that they can propose for you.
Right. When do you think these relationships with these Tier 1s will become tangible? Is it 2024? Is it 2025?
Well, we just announced the first design with Conti.
Right.
You can see that the momentum is there.
Right
Want to continue to leverage on that. Also, we were trying to make sure that we continue to build our momentum to convince, like, you know, Chinese Tier 1s. That's also because Continental, Bosch, also in China, but Chinese tier, there are also different Chinese Tier 1 that we need to work with. That's definitely a effort that we continue to put in to make sure those Tier 1 can propose our engineering solution to OEMs when it's, when there's RFQs.
Got it. You think the business in China will prove sticky over time, or you think that business will face a lot more competition?
You know, competition everywhere. I think sticky is a y ou know, in China, well, that's a good thing and bad thing about China. China is, if you have technology, they will use you.
Right.
Doesn't matter your scale.
Right.
Right. As long as your price performance is right, they will give you a chance. That's beauty of Chinese market. That's why, we think that the CV72 will penetrate the Chinese market. Of course, at the same time, China market has a problem, which is that, you know, they want to use their Chinese components as much as they can. That's why, you know, like, Horizon Robotics has a home court advantage in China.
Right.
Our strategy has to be, we have to provide better price performance than our Chinese competitor, and making sure that we give our customer a reason to use us. Technology has to be there, right? We think that we have a chance to win China because our price performance is there. Also, we have a roadmap that, you know, you know, our high-end chip already
Right
You know, is multiple over an X performance, while our low-end chip is selling at, you know, we're only quoting a $50 in the Chinese market. You can see that we have a roadmap that cover really from high-end to low-end for automotive, and I think that's another advantage we have over our Chinese competitors.
Got it. Has some of the weakness in the business been because of the weakness among China EV makers? Like, is there a correlation there, or was that not business you would have gotten anyway for this year? It was a longer-term business it's not the trends this year that matter, right?
Right. You know, we do sell our other solution like, you know, recorders and also e-mirror, so type of solution into EV market. We do see that's weak, however, it's not really big portion of business.
Right
It didn't have a significant impact on us. We talk about our China total exposure is 15% being consumed in China, there's only a portion of that is in automotive. There are other business there. Overall, I think that because our penetration in Chinese auto market is not big at this point, that impact to us of the today's economy is not severe at this point.
Got it. Now, I have to ask you about the AI market. Right? That seems to be, you know, a very exciting growth area. What does that mean for Ambarella? Because, you know, when I think about it, especially large language model, right, there is the well-known data center market, and then there is also people like Qualcomm, who are saying, I can put a 10 or 20 billion parameter large language model on the phone itself. Right? What kind of opportunity do you see for Ambarella? When do you think it starts getting more tangible?
On the edge side, we are doing that already, right? For example, autonomous driving car, you know, one of the most popular network is called Transformer. Really, that's a network. If you have a Tesla, on the display, you see this bird's-eye view of your surrounding. That surrounding view is generated by Transformer.
Right.
Everybody, all of the car maker want to copy Tesla, need to have that Transformer. If you ask around about who has the best implementation, Transformer implementation on silicon, I think you'll hear our name. Also, we demo that Transformer at a CES with one of our largest Tier 1s. We know that running LLM type of network, we have the probably one of the most efficient silicon architecture. From that point of view, in fact, we have, you know, less our CTO and I have been working, looking at the server market for many times, and we decide not, that's not our market in the past, because it was dominant by the training workload.
Right.
We don't have training data, so we don't have training solution.
Yeah.
Also, even for the server inference engine, the market is segmented, and our differentiation is not really highlighted there. I think that argument was changed when the LLM was introduced. Everybody believed that LLM inference will be the dominant server workload in the near future.
If you ask the people who are really generating all of the workload, what's the most painful thing for the current solution is the price is too high, and the other one is power consumption is too high. We start asking ourselves, with that, do we have a solution? Do we have a opportunity in that space? In the last five months, we spent time to port a large LLM model onto our current silicon. The silicon we sample to our automotive OEM in August last year. We port to a real silicon, and kind of confirm that we believe Running an LLM performance on the silicon is similar to A100 and at a 1/5 of total power number.
With that, we believe we can address both the cost as well as the power consumption problem, right? Why power consumption is so important, right? You know, each data center, you know, the power supply is a fixed number.
Right.
With 1/5 of performance means you can stuff in 5 times more AI performance into that data center. That's meaningful. I really think, especially when the workload becomes so dominant in there. I believe that there's a demand for a ASIC type of inference engine for server, and that's where we'll come in. Where we are currently at, that we need to finish, quite quickly finish our demo to our customer and start doing a business development to talk to a customer to make sure they understand our performance and also convince them that to port their LLM model onto our chip. We don't need a lot. We just need a few example that prove our performance and the power consumption, and that. With that momentum, that will probably really.
Right
Trigger us to do more investment if we can convince ourselves there's a real business behind it.
Is comparing to the A100 a fair comparison? That's, you know, part of a lot more of their training setup. If you were to compare it to the L4 or L40, right, which are more inference PCIE SKUs, do you think you would still have a favorable comparison?
Yeah, I think so. In fact, definitely, that's A100, you know, A100 is really GPU, right? If you look at using A100 to inference, that performance is still pretty high. I think we can definitely talk about how to do a fair side-by-side comparison.
Right.
It's really about to prove that. You know, we have done this, with, you know, automotive trying to compare.
Right
our other new network performance compared to NVIDIA, that we get similar performance numbers. We definitely think, you know, from our point of view, that's a fair comparison. If the people, customers think that we need to prove further, we'll definitely do that.
I see. Is that a worthwhile market for me as, you know, we were chatting before? You know, you mentioned that there's a number of players, right? I mean, the incumbents, right, it's not just, you know, NVIDIA, it's, you know, Intel, Habana Labs, right? There is AMD with their MI300 planning to enter the market. You and, of course, you have Broadcom, TPU, right, et cetera. Is that a market worth investing in when you have 3 or 4 incumbents, and when you have several others who are also trying to... Right? Like, how much of your R&D resources is going after that market?
First of all, you know, we're using our current silicon.
I see.
Right?
It's incremental.
Yeah, and also the software that everybody trying to build-
Right
We already have that software for our other business, right? Today, people talk about compiler, trying to convert neural network to our chip, and we have helped 100 customer convert 500 different neural network on our chip with based on our tools already. We have tools, we have an environment. We know how to convert, help our customer to convert neural network to our machine, and we need to optimize the software for LLM. At the same time, you know, through working with the IoT customer and automotive customer, guess what? We have been helping them to port software running on CUDA to our platform twice already, helping them to do one more time to run the, you know, server and inference engine to our chip. It's just another exercise that we have done.
From that point of view, the resource required to enable a handful customer to prove the concept, get design win, we don't need a lot of people because we already have that investment on CV3. We basically leverage the heavy investment that we put on CV3 in the last several years. From the additional incremental OpEx, I don't think that's a lot. When we convince ourself there's a real business behind it, and we can win that market, you bet that we're going to ramp up that engineering investment to try to get an even bigger market share. As today, our approach is let's convince ourself and the customer that there's a business model that we can win.
I see. On, you know, back to the automotive side, I think in the past, you have also made important investments, right, in, on the radar, right, side, and then in, domain, controllers. Could you give us an update on where we are?
Right
in those investments?
Domain control is really CV3, and we talk about that, the design win already. CV3, as a domain control, we're doing very well in terms of, you know, continue to deliver software only on our chip and demoing customer. We are talking about that we already have a first revenue in China when we get more design win there. That's happening. On the radar side, I think it's a very important thing because, you know, Tesla just announced that they're going to put the radar back, so all the copycats who want to copy Tesla design. We're starting how we're going to use a 4D imaging radar to a ground truth. That definitely will help us to do a business development.
We and also we have momentum now on the design win side with Tier 1 to really look at our centralized architecture and porting their software over. I think from the radar point of view, I think we also start continue to make progress on the current situation. Our target is really at the end of the year, we have to demo that our VisLab software, Oculii software, running on our single chip CV3. That's the important milestone, because at that time, we basically have a total solution offered to OEMs, you know, with everything they need.
Got it. How would you contrast that with, you know, say, somebody like NXP or Infineon, right, who are the incumbents in some of those markets?
I think they are still for 4D imaging radar, they still focus on edge device.
Right.
Basically, I think, building a module on the edge and using very big antenna array to do that. That's not our approach. Our approach is how to use the smallest possible antenna array, you know, 4 by 3, ideally, and we use signal processing at the center.
Right
To generate similar, point cloud and the performance in terms of range, resolution, and accuracy.