In this episode, Anna Petrovicheva interviews Loic Leitar, the CEO and a co-founder of GreenWaves Technologies. GreenWaves Technologies is famous for its ultra-low-power processors called RISC-V, which is able to process rich data using AI. In this interview you will find out more about this fascinating technology, as well as the company culture behind it; also, Loic will share his opinion on some perturbations taking place in the hardware market, and his insight on the prospects of technologies amid the current situation.
The episode was recorded in October 2020.
Anna Petrovicheva (AP): Hello and welcome to the OpenCV AI for Entrepreneurs podcast. My name is Anna Petrovicheva and I'm the CTO of Оpencv.AI. Today I will talk to Loic Lietar, who is the CEO and co-founder of GreenWaves Technologies. They are a company based in France, they create a processor named GAP8, which can run on battery power and do artificial intelligence computations.
So, this is a very ultra-low-power technology that is going to disrupt the market of intelligence sensors. I’ve been working with Loic and his team for quite some time now and absolutely admire the product that they make as well as their work culture. So, please welcome Loic Lietar! I'm really happy to talk to you in our podcast, and the first question that I’m going to ask is about the company that you are running — about GreenWaves Technologies. Could you please tell us how you came up with the idea and how you decided to found the company that creates such an innovative and low-power cheap technology?
Loic Leitar (LL): First of all, thank you for the invitation! We are a semiconductor company, we make processors. The company was founded almost six years ago, and it was not found as what it is today. Originally, we had to plan to make high data rates for all of our trial power wireless connections. This was the first year of the company and our CTO Eric was not on board at the time — he was a good friend, he knew what we were doing, he was not on board; but as we were looking for a processor architecture to run those algorithms for high data rate wireless connection, Eric wanted to set up his own company and ask me to join on board and be the CEO of that company. I told him I could not be two CEOs at the same time. We had an office, and he was more than welcome to sit with us, and, having worked side by side for a couple of months, he told me his processor idea would be a good fit for my communication needs. We studied the matter with due diligence, probably it took us a month or two, and afterwards, we concluded that he was right. So, the processor, that today is our first product, was a good fit to run those communication algorithms, and Eric joined the company. We had dinner, as we usually do in France to make decisions, with my chairman, Eric, and myself, and we decided to merge the two projects. Eric joined as another co-founder, but as a matter of fact, we raised money on his project which was focused on making an ultralow-power AI processor for the very edge of the network and the communication technology that we had, and later on, the market decided there was much more traction on Eric's original idea than on a communication idea. So, GreenWaves became what it is today — a company that makes ultra-low-power processors for analyzing rich data with AI methods. It’s necessary to highlight that from day one we thought that the traditional VSP was and would remain important on any heavy-duty processor, so today we do AI, but we also do traditional signal processing. And interestingly enough, the DSP side of the processor is very appreciated by your customer.
AP: So, my next question is going to be about the company culture. How do you build a team that is capable of creating such a breakthrough technology?
LL: First of all, we decided with Eric that we want to have fun, we want to enjoy life, we didn't want to repeat what we dislike in large companies, so we had a very strong vision of it, and I had this morning my monthly meeting and I repeated that the company is made of agents, so everybody is empowered. They know what they have to know, what they have to do, they have to align with each other, so it's more of a bottom-up organization than a top-down one. We think about it very deeply, it's very demanding intellectually because there is no available model to refer to, but so far — so far so good; it has worked pretty well.
I had this morning my monthly meeting and I repeated that the company is made of agents, so everybody is empowered.
The second point is that I think there is a great combination of senior people, people who have understood the whole industry (semiconductor is a very big industry), and young enthusiastic people. Now we have all the competencies that we know we need to do the job, so it's a matter of how you mash those competencies efficiently, and this is about the execution side. Your question I think was more on the innovation side, but you can’t build anything unless you execute well. We have this romantic view that startup is about new things, new ideas; sure, but this is only five percent, 95 percent is about execution. And because you are a bunch of people against gigantic companies, you have to execute perfectly and very efficiently. On the innovation side, it's really coming from the relationship we have with two world-class academic research centers: in Zurich and the University of Bologna, and a number of our guys are coming from there. We work seamlessly with them and the reason for which we are able to work seamlessly, without research modi getting into the picture, or intellectual property getting into the picture, is that we are leveraging open source projects.
Something that is a product for us is a test chip for them, and the testing results will be described in academic publications, but it's the same object anyway. This open-source allows us to work together on the same object. This combination isn’t easy: it's an everyday effort to make it work; we also have a number of our employees that are part-time employees of the university. The head of those labs is a shareholder of the company, and I think this is where the innovation and the stream of innovation are coming from. Now as the company has grown, we are more and more intimate with some of our markets and some of the very deep innovation: we have a recent example in the audio space, this is property to the company, it's not coming out of the relationship with the research centers and this innovation stream. It is coming from our understanding of the market.
AP: That's fascinating — I mean this combination of industry and technology. I think I agree that this is the point where all the innovation happens.
LL: Right, but open source is obvious in the software industry, but it was not the case 30 years ago — probably you're too young to remember, but in semiconductors, this was a leading-edge stuff. RISC-V was a big buzz those days. RISC-V is not about open source, it's about free instruction set architecture, but it has enabled ETH Zurich and the University of Bologna to build this open-source project out of which GreenWaves Technology has come out.
AP: So, let's talk about the RISC-V itself. What do you think is the future of the technology, because right now the technology is emerging at a rate that is probably unprecedented in the world for a hardware technology like that? Could you please share your opinion on what is the future of the RISC-V standard as an architecture?
LL: I have a very strong opinion about it: first of all, there is a big buzz. There is a huge momentum now, it has been adopted for deeply embedded applications which means that you find RISC-V in many of the big chips of big companies. But they are the processors that they use for themselves to control their chips, they are not processors available to their customers to develop programs and run them on RISC-V. Commercially, this is the highest use of RISC-V today. RISC-V is a fabulous tool for research. One of our academic friends was saying that the number of academic papers on processing architecture has been multiplied by 10 since RISC-V was out because it's a great platform for researchers to start from something that is already held back and innovate out of it. I must admit anyway that if you look at what I call “commercial processors” or processors that can be programmed by third parties (exactly what we do — we sell processors and our customers are programming them), then the offer is very limited. I know a couple of companies in China, I know Turbo company in the US because of the very ambitious project based on RISC-V, and there is our company, but they are very few in total. I think several people are making their own as-is, but there is a limited number of open commercial offers. It's difficult to build a commercial processor: you have to come up with a complete offer, you have to explain it to people and justify why it is like that. And frankly, why would you drop it if the architecture that you offer is not radically innovative? RISC-V by itself is slightly better than ARM, but the fact that it's freely available doesn't bring anything. The guys who are selling the processor don't have to pay a few cents of royalties to ARM. It doesn't change the economics. Where RISC-V changes everything is the ability to innovate. To my surprise, there are very few companies that innovate architecturally, and we are one of them! Maybe it will take a bit more time, maybe things will come, but just replacing let's say in a microcontroller, an ARM microcontroller by a RISC-V microcontroller, and keeping everything else equal, in this situation I don't see the benefit for the market, and as a matter of fact, this is not happening. We have seen a lot of test chip announcements, but in real life, in production, there are very few commercial offers with RISC-V.
AP: I see a new angle to this market, because this recent use of Nvidia acquiring ARM may affect the future of RISC-V in a very positive way. I think that one of the possibilities is that as an aftermath of this deal ARM becomes a US-based company, and faces sanctions from China, right?
LL: My comment was purposefully ignoring geopolitics, and our project is beyond that, but of course it happens. In China, RISC-V is big, in Europe it starts to be big for some applications, for sure the guys that are in the space business, military business are very strong for RISC-V. But I would not build a rationale based on that: the people who launched RISC-V don't have that in mind. There is no doubt geopolitics attempts to control the industry, and for RISC-V being open-source is the way to address the issue. It is true for the processor itself, it is true for the architecture on top of that, it is also true for the software tools we use to design the integrated circuit, and for the time being there are three companies only in the world and all of them are American. And it's a big threat. But the open-source solutions are not at all at the level at which we need to have them to be used in commercial products.
AP: Thanks! Could you please tell our followers about some of the applications that you foresee to be the most popular for the RISC-V architecture and RISC-V devices?
LL: RISC-V is a processor like any others, so as I explained that deeply embedded controllers are the natural spot for it, and here it will be unless someone would ever again buy an ARM processor for the embedded systems. This is anything! Maybe it's easier to list what in my opinion RISC-V will not get in: in an established market of smartphones, I see no rationale for the industry to switch from ARM to RISC-V and the same in many areas. RISC-V has a chance where red classic architecture cannot make it, when you have energy efficiency requirements, for example, that are beyond what a RISC architecture can do. This is our case — okay, we are also a RISC-V company, but our merit is not coming from RISC-V, it's only an enabler. What is unique about the architecture: it is relatively parallel, we have typically eight cores loosely coupled working in parallel, it's hierarchical, so you have one plus eight cores, and many other features. We are not running applications thanks to RISC-V; RISC-V is enabling us to develop an architecture that allows us to be much better than a regular RISC processor, whether it's ARM or RISC-V. We can do it with RISC-V, we could not have done it ARM, that makes the difference.
If you are not 10 times better than the established competition, you just don't make it, people would rather wait for the big guys to deliver the next generation product than work with you.
Our target market is battery-powered devices that have rift sensors, typically an infrared camera, a microphone, a radar, where sensors produce a lot of data, and those data are analyzed and transformed locally. The wireless communication has to deal with a reduced stream of data, which we call metadata — the outcome of the analysis. And here you are back to the classic machine learning or deep learning inference type of solution; we are not dealing with learning today, and certainly our processors today are not capable of doing that, so it will be about the simplest but also one of the most successful applications — counting people in a room. You stick an infrared sensor on the ceiling, you snap a picture of relatively low resolution (80 by 80), and our processor will count the number of heads recognized in the picture. What we bring is that we can do it on a small battery for five years, one picture per minute for five years, this is huge! Also, we can recognize faces, we can locate the face, we can do face ID, and we do that with you — thank you, Anna. We also do keyword spotting, so we recognize words, and so on. If you compare this type of architecture versus more specialized architecture, the picture will be as follows. Our competitors tend to pick a vertical architecture, they think is important, and they will ultra-specialize the architecture for that, for example, keyword spotting. We think that in the microcontroller world which we are in this is not the way to go. Those machines have to be good enough, to do many things, for two reasons: the first one is that frankly, we don't know what would be the killer app; now we have a better idea, I’m happy to come back to that after. The second one is that on the same transistor because this is the only processor of the system, you will want to run more than one algorithm so you might want to run say a machine vision algorithm to recognize a person, but also to recognize the voice — on the same architecture, and the processor has to do a good job on both. Or you might also want to add a gesture recognition capability based on the radar signature recognition, and here again that machine has to do all that together. And our last product is finding very good traction on the earbuds, and here you have a heavy-duty neural network algorithm running, but also heavy-duty traditional audio DSP running, and a processor does it all pretty well. Indicatively, if we compare ourselves in most of the markets to the competitors, we are 10 to 20 times more energy-efficient.
AP: That's a magnitude! That is going to disrupt the technology once it gets widely adopted!
LL: Right, but if you are not 10 times better than the established competition, you just don't make it, people would rather wait for the big guys to deliver the next generation product than work with you.
AP: Okay, thank you! My next question is about the weird times we all are living in now. How did the pandemic and this global situation influence the hardware market and RISC-V devices in particular? Has it been affected at all?
LL: At the semiconductor industry level it's very strange, because first of all, on the manufacturing side those guys operate in very stringent conditions. The industry went ahead, except for probably the automotive market, and the semiconductor industry has slowed down. This is very surprising, but this has happened. In many countries, by the way, it has been also identified as strategic industries so the factories were required not to shut down, which is not the case in other industries. As far as we are concerned — of course, we are still very small. We have seen no customer consolation except one in Korea because the customer itself was financially challenged because of the crisis. What we have seen is a slowdown in the project because companies were not operating as neatly as they were before. We are in a hardware business, so when you have a customer developing a product, it's a piece of hardware, and you have to deal with a piece of hardware while staying at home. We shipped our reference design board to our customer's house, but the oscilloscope was in the wrong place, and those things resulted in delays in the project. So, this is really about the operation. Also, some projects have been delayed, because it's about experimenting in real life, for example, the people counting application. At some point, the customer has to deploy those sensors into offices to make trials, so if the officers are empty, the trial cannot take place.
I assume that we were hit by a four-month delay, and the sales that I was planning to make this year, I will make them towards the end of the year and next year, but it's not dramatic — on one hand, and on the other hand, we are lucky because in the EU, especially in France, the financial instruments have been put in place very rapidly to protect companies like ours and they have worked very well. We closed a five-million-euro profit financing round a week ago that is mostly debt and that is coming out of the instrument that the state put in place.
AP: Congratulations on that!
LL: It's the government that did the right job on that.
AP: That's great! I guess there is also a positive impact of this whole situation that to my mind in two years perspective people will think more about automating things, right about deploying more robotics into place, because right now we clearly see that some things that can be done by machines, should be done by machines, right?
LL: That's funny I don't have the same view, I think that trends will not be accelerated thanks to the crisis. Again, I’m biased by what we do as a business and where we find traction. What I see is that a lot of technology will be put at work to help people work no matter where they are. For example, if I am at home, the sound might not be that good, or if I have a baby, the baby will be in the background. I believe technologies are coming out — and we are contributing to that — to guarantee that no matter where you are, you can work, you can discuss it, etc. It's fascinating because I have discovered that the technology that we are using today for headsets is very primitive. There are many ways to provide more life to the voice that you can have in your headset. The science is there, the algorithms are there, and the only thing that is missing is the computing power with the right energy efficiency, which by the way we are bringing. This is why we are so excited, and I see that trend being brought in by the pandemic more than the robotization I must say.
AP: This was really helpful, at least to capture podcasts while people like you and me are distributed across the world and sit in different countries, right.
AP: So, my final question for today: what would you recommend to people who think about creating a startup in the very competitive and very tough area of hardware engineering?
LL: I think we are all obsessed and very romantic about technology. Technology is fundamental, but it’s even more fundamental to start with a very strong vision of where the market is going that no one else believes in. A startup is always going through missteps, trial and error, it's normal, because you are building a social body that you start from scratch. Therefore, you have to make sure that your vision is advanced enough and risky, so that it will forgive the fact that here and there you have missteps. So, those fundamentals are very strong, although now it may seem to be something very abstract. For example, in our case, we came up with our idea saying that we will need to deliver much more computing power in those battery-powered devices than we need today to analyze rich data locally. This was the vision. It turns out that six years later it’s IoT that would make our success hearable, but still, it's the same vision, so, we are just slightly changing the angle, not much more, but the fundamentals of battery power objects that have not super confusing capabilities. This has been an environment for the company for six years, and for a long time, we have been in a raider mode. Then, when you have set the vision, you have to scan and try to understand where you get a signal on your radar, where the market resonates with you. Also — and I think we almost made the mistake — you should not try to make it too difficult. In our case, for example, we are coming out with a new concept technologically, we have to sell RISC-V, it’s AI, and most of our customers have an intuition that they need to deploy AI, but they don’t know how to do it. So, it takes them a long time to make a product and, therefore, make sales for us. They are experimenting, and for a startup, this is challenging, because you need revenue, you cannot live on this money forever. So, ideally, your vision has to apply to an existing market and disrupt an existing market as opposed to helping to create a market. In New York we are very romantic about that, I think other geographies are less, so we love the idea that we will invent a new market. For a startup this is cheap: you have to build the team, build the technology, and wait for the market that you are enabling to realize that the chances to fail are pretty high. So, it’s vital to have a unique disruptive vision at the same time trying to apply it to an existing market. That would be my recommendation.
AP: It was a pleasure for me and a privilege to talk to you, and thank you!
LL: Thank you so much!