BrainChip Holdings Ltd (BRN) Earnings Call Transcript & Summary

Okay, everyone, I'm going to jump in now. It's about 6:00 here in California. Welcome to everybody in Australia and Asia. I know it's your morning. Unfortunately, we can't have folks participate from Europe. It's middle of night for those folks. Today, I have with us Peter and Anil, 2 founders, CTO and Chief Development Officer. Sure, you know each of them. They're going to make some comments about the product development as well as research. [ Clive ], if there's any problems with audio, you just let me know.

Will do.

It's always a little bit uncomfortable because I'm sitting here looking at this -- and there's -- we've got 200-plus people listening. So if there are any problems, please let me know. First, I'd like to say, throughout the presentation, I have tried -- we have tried to address all of the questions that have come in through e-mail and the update. If we miss anything, I'll certainly try at the end of the call to make sure that we haven't missed anything. But if we have missed anything, certainly just reach out to me directly. I'm sure most investors have my e-mail address. You can reach out to me. You can reach out to Roger, Peter or Anil, we can kind of fill in the blanks. So I'd like to thank the folks at the Financial News Network. They do a great job in helping us orchestrate this. This is not an easy task. I'm in Northern California. Anil is in Southern California. Peter's in Western Australia. And so I'm hoping this goes well. We have Peter alive or live on the phone as well as Anil, and certain slides that we go through in this deck, I'll turn the dialogue over to them. I'm going to move forward. This is the standard disclaimer you guys -- folks have seen it time and time again, but it's a prerequisite for us as a public company. Read it at your convenience. Many of you know exactly what it says, but I'm not going to spend a lot of time on it. A quick agenda. We're going to talk a little bit about health, safety, communications and productivity as an update because many, many questions about what's going on with corona. Frankly, you'll see, we'll talk about it in detail. It's really not been disruptive to our business. It's not been disruptive to communication. It's not been disruptive to productivity. A little bit about sales and market update. Financial update is going to be very cursory because as you probably know, we pulled this call in. We would normally not create an update until after the 4C, because here it would be late April. I think April 30 is the 4C. And then we would do an update with -- amended to the -- appended to the 4C in a conference call. But there's been so much curiosity about where we are. And I think, rightfully so where we are with the development of Akida, where we are with sales and market traction and the implications of what's going on with the kind of global issue with the, call, pandemic or coronavirus, that we decided that we would pull this in and provide an update now. So not much in the way of financial update, but still closing the books on the quarter. I mean the quarter is not even over, but we'll be closing the books very shortly. Anil will address the product development update, and Peter will address the research update. Again, questions, we've tried to cover the questions here. If we miss anything, absolutely feel free, as I know, you all do to send us an e-mail. Let's go forward. On this issue of corona, I'm not going to dwell on this a lot because, frankly, it has not had a great impact on us. I mean on a personal level, of course, I've got an 88-year old mother. I've got a stepson with diabetes. But with respect to how we run our business, because we are a relatively distributed company to start, Northern California, Southern California, Toulouse, France. Our partners in Shin-Yokohama in Tokyo. Our communication, in some respects, I feel like it's actually accelerated. People are working from home. We instructed our employees even prior to California, New York, New Jersey, Illinois and whatever is going on with international travel restrictions, we instructed our employees to work from home several weeks ago. It has not been a disruption in communication. In fact, my e-mail lights up at 2:00 in the morning, 3:00 in the morning, all throughout the day. I think people are -- they're stuck at home. And I think we're actually getting a little bit more productivity, a little bit more communication than people in the office, taken off for lunch and then doing what you're doing on a normal office day. Our communication with partners has been relatively unaffected. In fact, Socionext is not in lockdown. They are actually going to work every day. They are in the office in Shin-Yokohama. So that partnership has not been affected at all. Certainly, there's been some challenges with respect to customers. It varies by region. I was supposed to be in Shanghai and Greater China for the month of February and Seoul, Korea as well. And we have lots of activity going on. But this is your engineering dialogue. And this -- conference calls, you know traffic. So I would say, little impact, certainly, a different kind of communication. But nothing that I think really affects how we're making progress with target customers. A question I was asked by several was, what's the implication for the supply chain, as we move into wafer fabrication, as we move into back-end assembly test operations? I just literally got off the phone with our partner. There is really no supply chain interruption. TSMC, this is 28-nanometer fab in Mainland China. Frankly, probably the safest place to be, it's in class 10 facilities. So you can't get sick in that place. But there's no indication of supply chain interruption, either through wafer fabrication or backend assembly. So nothing to be concerned about on that front. In summary, I think productivity is really unaffected. I think, ironically, in some respects, communication because we're all working 24 hours a day around the globe, maybe productivity is up rather than down. And the lack of ability that's travel. Yes, certainly, there were some meetings that should be done face-to-face. But at this juncture, that's really not affecting our performance. Move on to the next slide. Is that right? So a little bit an update on sales and markets. And I'm keeping this relatively cursory. Certainly, AI Edge, as I think you all understand, is our target market. That includes automobile manufacturers, module suppliers to automobile manufacturers, Tier 1 global market manufacturers for Smart Home and Smart City. Tier 1 automobile manufacturers, we've got the trade. We can't name names, but we've got the trade, where we're close to a proof-of-concept agreement. That is kind of the first step. We'll get a little bit of money. But more importantly, it's validation that Tier 1 automobile manufacturers are interested in developing a solution, which includes an Akida device in a module or in some part of their infrastructure in the automobile. Automotive module suppliers, probably the most target-rich environment for us. Most of our mobile manufacturers don't develop their own modules. They push that down to Tier 1 module suppliers. We have a proof-of-concept, and it's been frustrating. I've talked about this before, but it's frustrating. It's a large European Tier 1 module supplier. Contracts taking 4 months -- almost 5 months now. But it's moving forward. In the module arena, you can think about LiDAR. You can think about ultrasound, you can think about radar and you can think about standard pixel-based cameras. In our case, what we're finding is a sweet spot in the LiDAR environment. So that's moving forward very nicely. Unfortunately, again, you're dealing with large multinational conglomerates, the legal issues sometimes take longer than you would like, but those are being resolved quite effectively now. Smart Home, primarily, in this case, it's really South Korea for us. But Smart Home is a big deal. Anil will touch on some of the aspects of what we can do with keyword spotting, what we can do with incremental learning. These are things that the smart home manufacturers, the global manufacturers in Smart Home and as well as Smart City, those are probably 2 together are very excited about. Okay. So now let's talk about Socionext. We put a press release out and Socionext actually is releasing that same press release. They translated it to Chinese, that will go out on the Chinese news wire I think today or tomorrow. And this is a reference design. Socionext has a very high performance, 24-core Arm processor. So you can think about it as an Intel processor or AMD processor. In this case, it's a homegrown ARM core -- 24-cores. What they do currently in this reference design is they offload the multiply accumulate for the mathematics that are required to a DLA or Deep Learning Accelerator. All it does is the math. What they found very attractive and what they're offering with us in concert is move the entire network onto the Akida device, so that the SynQuacer doesn't have to run the network, it can do what it's expected to do, which is the analytics, the middleware, the user interface, let Akida do all of the hard work of the neural network. In our case, we don't have to do all the multiply accumulates, but it basically offloads the SynQuacer from having to do the neural network. It's a great application. It happens to be an edge server for video analytic application. They will market directly to OEMs and the market -- in concert, the markets' ODMs or the subcontractors that builds for the OEMs. The revenue will come directly to us. The reference design will show our chip. They will show the SynQuacer and whoever builds it, whether it's the OEM directly or the ODM, will come to us and buy the chip directly. So it's a -- it's not a reseller kind of situation, they will come directly to us. But we're also exploring further commercial opportunities in cooperation. As we've talked about previously, Socionext is the second largest ASIC supplier in the world. They know our IP very well. The opportunity to offer our IP in their portfolio as they approach their customers. That's another step in the process. And potentially, it could be a reseller of our IC and we're not going to be able to cover the globe. Whether it's Japan, Mainland China, Southeast Asia, we're not going to be able to cover the globe ourselves, we have to partner such associates that should be a big benefit. So I think this has been a very powerful validation that the Akida device in concert with this 24-core arm processor can be a very, very powerful, it's interesting to use word power because we're trying to focus on low power, but a very powerful solution for the OEM and ODM requirements for H servers in -- particularly in this case, the video analytics arena. So I think I've touched on several of the questions that were asked about the announcement. Again, if I miss anything or we miss anything, please feel free to send me a note. Again, financial update, I'm not going to do a lot here because we are -- we have a 4C that's scheduled for 30th of April. We'll talk about cash then. We'll talk about expenses. We'll talk about forecasted expenses. But rest assured that we are maintaining continuous expense control. We're controlling our head count. It doesn't mean we're reducing our head count. It means we're controlling our head count, not adding new heads. One of the interesting things that, I think, will resonate with shareholders that are interested in kind of the technical nuances of the manufacturing process is we have decided that the wafer fabrication will be done on a multi project with -- that is historically, I've called it a pizza mask, take a wafer, and then we get a slice, somebody else gets a slice, somebody else gets a slice, it expedites our turn time, our lead time through fab and it reduces our cost. I'll touch on that within a moment. Travel, travel expenses should be -- are going to be down dramatically because there's no place we can go. Maybe that's good, maybe that's bad, but it's certainly just helped the cash flow. And we have certainly taken a great deal of effort in reducing discretionary legal and advisory expenses. Okay. So it's kind of my slide, which is lots of questions about what do we do with respect to capital? We will require revenue and do our investment to fulfill our mission of commercializing what is really groundbreaking technology. We live in a world because we're a public company, where -- if we were private and we were venture capital backed, people will be throwing money at us. In this case, we want to be very, very careful and very, very circumspect about how we raise money, so that we can reduce the cost of capital as best as possible and increase shareholder value at the same time. So we are looking at strategic investment. We're looking at debt, we're looking at convertible debt. We're looking at structured finance. We're looking at equity. And certainly, we're driving for as much revenue as possible. But I think it's -- it would be disingenuous for us to not be upfront and say, yes, we will probably have to do a capital raise. There'll be some offset to that and minimize it, but whatever revenue we can bring in. But we want to reduce that cost of capital as much as possible, and let our shareholders enjoy as much of that benefit as possible as well. Financial update. This is just giving everyone a sense of, kind of, the diversity of where we're located and what we're doing. We've got Silicon Valley. We have sales, marketing and an executive office here. That's the yellow on the top left-hand corner. Southern California is where Anil runs fundamentally all the hardware, software and research. Toulouse, France, again, software and research. Hyderabad, India, software development. And that's a contract services organization right now. Perth, Western Australia, which we'll speak to it in a moment. And new innovation center that is for advanced research as well as applied research to support customers. We are evaluating Shanghai, China that probably would have moved more quickly. China not shut down on us in the last 60 days. Head count, 34 full time, all 6 contractors. One in Western Australia and the other in Hyderabad, India. There's a bunch of questions about competition. Specifically, there was an article about Intel and what they're doing with the kind of second-generation or Loihi? Look, as CPU is slow, it's inefficient. It's inefficient to use of critical resources. The CPU should be doing other things. That's why you see DLAs, Deep Learning Accelerators. It's why you see GPUs being used to offload the matrix multiplication that is required by standard deep learning architectures. SoCs were System-on-Chip basics from other reported competitors. What we're seeing is an inefficient use case coverage. There are companies that are focused exclusively on voice. And they're small networks. It's not scalable. It can't be used in any general purpose sense. There are some that are more targeted at video applications. There's some that use esoteric processes. They're using analog multiply accumulate functions or they're using sub threshold logic. These are not portable to different fabs. They're not scalable to different geometries. Akida, on the other hand, is highly efficient, it's ultra-low power. And these are the things that when we talk to kind of the major lead customers, potential customers in Korea or China, in Europe or the U.S., highly efficient ultra-low power. And we're talking about hundreds of microwatts for certain applications, hundreds of milliwatts for most applications. And even when you get to the very, very large networks, a couple of watts, when -- if you use a CPU, you're burning critical resources. If you use a GPU, you're burning tens, if not hundreds of watts, and a DLA, which really is not scalable across multiple applications. So it's flexible, and it is a complete network. It does not require a host processor, does not require external memory. It runs on its own and it offloads the CPU to do what it's supposed to do without the inefficient use of power in GPU, the inefficient system design of the DLA or the inefficient use case coverage of an SoC. Okay, here, I'm going to turn over -- I'm hoping that we've got this technology now, Anil, I'd like you to jump in here if we've got you live.

Yes. You can hear me okay, right?

Yes. Perfect.

Yes. So product development update. We are actually done with all the work that we are working with Socionext, and they are ready to hand over the, what is called, a GDSII data that TSMC can make the mask, will happen on next Tuesday. Our wafer start -- will start on April 8. That's the MPW that Lou talked about, which gives us faster turnaround time. We will actually also bring the back-end process for assembly and all that Socionext [indiscernible]. We also have a couple of our reference board designs, all the software ready, when the chip comes back, tested. So all of those things are on track now, and this is the schedule that we are running to right now.

Let me interject because, again, here we touched on the multi-project wafer, which is a -- by default, it's a hot lot. So it goes through a [ patient ] faster than the normal wafer. And we talk about engineering samples in Q3, our target is very early Q3. We put Q3 as kind of a round number. But given cycle time through fab, cycle time through back end, it should be really early in Q3.

Yes. That's correct. So we are on track. And I think we have very -- feel very good about where we are with the project.

Okay. So I jump to the next slide. We put a big circle around the center section. Maybe you could describe that middle ground between traditional data-based convolution, and at the far extreme, when the world attaches itself to a native spiking neural network. There our opportunity and advantages in event-based convolution.

Yes. So what we do is there are multiple ways of solving the inference in neural network, either you can do a data-based convolution, where people do matrix multiplication, it's all systolic array, there's multiple MACs. It depends on how many multiply accumulates we have. Some people have 1,000-multiply-accumulate engines or 64,000 like Google TPU has. But those are just accelerator. The neural network actually runs on the CPU. And on the other extreme is like Intel Loihi or IBM TrueNorth is pure spiking neural network. You can actually do spiking neural network, but you need spikes to be coming in. Most of the time, you are getting data. What we do actually is -- what we have done, while our Akida is actually event basis spiking neural network, but we have actually taken a middle ground where we convert standard CNN and we bring them into a event domain, which is called spiking domain. So we are able to do today -- today's neural network that are DNN neural network into spiking domain and run them very efficiently. At the same time, in future, we can release spiking neural network with the same hardware. We have actually -- I'll talk about in next row, which is our development environment. So we can take a standard neural network, map it through our Akida environment. Of course, we can take any network right now with our tool flow. We take advantage of sparsity of events, both in activation and build. Most of the neural network that is event-based cannot take advantage of the sparsity. And because of that, we run the full network on our Akida. We can go between 500 micro Watts to 4 Watts, like what I talked about before. So what we have done is we have taken our neuromorphic computing elements. But we have mapped it to solve today's problem that people are having, which is by doing convolutions also on event base very efficiently in our hardware. So our IP is small, high performance, very low power. And of course, we added learning to it, which I'll go into details in my demo. So we are targeting a market that's today people want to do inference on the IoTH device in the power budget that they have. Not all the spiking neural network technology can do that, so we actually are bridging the gap between what DNNs do, what do spiking neural networks do. We solve today's problem with our event-based convolution actually in Akida very efficiently. Of course, the same hardware also does -- I'll show you also a spiking neural network with this demos later on.

Okay. So Anil, moving on to the next slide. Maybe describe a little bit more about the workflow. Because the ADE, the development environment, the field-programmable gate array, which provides emulation and soon-to-be engineering samples, are really just flow the customers are looking for.

Yes. Exactly. So we are quite unique in the industry that we have a complete software development environment, which we call ADE, that also has a Akida chip simulator that actually emulates what a chip will do. And this is what -- this actually uses complete industry standard flow using tens of [ flow gas ] and Python. So the people who are using standard is today are familiar with it. We can actually map any of the networks that people have and Akida in a software simulation and evaluate how many Akida cores we will need, what will the [ performance ] be, what the power will be. So the customer that Lou talked about earlier, the automotive customers or the automotive ODM customer, model customers, we can -- they don't have to wait for our chip to arrive. We can completely analyze it now in your software. Actually, I'll show you simulation results of one of the software. Now once -- the reason for doing that is once they are -- the -- we are flexible, we are actually -- we can scale our IP to whatever size that is required. Our customers can run the different types of network on our software environment, evaluate how big Akida IP that they will need, how many cores are required. That can be analyzed and then that helps them to really decide the size, power, budgets and everything else for the IP that they will use. This is the -- we are unique in this. That's complete software development available right now. This is free actually from the website, quite a few of our customers are using it. And we also use it internally to evaluate all of our network analysis that we have done. Of course, all of this Akida IP has been tested completely in FPGA for last 6 months. We actually have gone further and actually connected multilayer network that fit on the FPGA, like 6- or 7-layer networks, which are like keyword spotting. They're running on our FPGA. That gives us a lot of confidence that what Akida chip that we are doing is verified, works in FPGA. We have tested all the learnings and everything else on the FPGA platform. Some of our customers might -- when they license IP, might want to use our FPGA for verification. Once the chip is there, then of course, they will be using our development platforms and evaluation boards that are currently being ready.

Okay. And I'm going to move forward because -- folks I hope this works because we've got a live demo here that Anil is going to try to speak to. I may stop and start it so that we can catch up.

So before you start, let me explain what we have done. So we -- I talked about ADE. So we actually taken something called a standard MobileNet V1 network that has been trained on the GPU CPU that is actually used for classification of images after converting it into -- on the Akida and running on Akida simulator, what -- of course, it does classification from the original network, but I'll -- I'm going to show you how the learning for Akida, which is our unique feature is actually used to show that on a -- from a one-short learning. So what we have done is the network does classification, but we have removed the last layer -- the last classification layer and replaced with a Akida -- with a learning layer. Okay, will you please start the demo? [Presentation]

Could you set up the next demo, which is the next learning, where we're taking spikes directly and not having to do the conversion promo?

Exactly. So last demo that I showed you, we were taking standard frame basically camera that actually send or HD camera, you'll get 2 million pixel per frame, you got 30 frame per second, lots of data you have to utilize, but we were taking the data on a bench -- Akida on chip will convert into events and we're processing it. Now we have this demo that shows the -- there's different type of camera that actually doesn't send the full frame, but it only sends events or pixels with intensity change that was spikes. We will start the demo. This is a Samsung camera. We thank Samsung for providing this to us. It actually and does send events directly, so we don't need to convert in from frames to events. And this -- you can start the demo. [Presentation]

Okay. Let's move on. Peter, I hope you're live, can you speak?

Yes. Yes. I'm here.

Okay. So just a quick update for everybody. There's a lot of curiosity about what we're doing in Western Australia. So maybe you could speak to this slide and give some sense of what our intentions are and what our goals are.

Sure. The Western Australian research center, we'll be looking at the future of artificial intelligence. We are looking at where the market will be in, say, 3 years, 10 years from now. So that means that we are looking at building devices that are more like the brain, chips that can make decisions and they can learn like human beings. And as you just saw, we are already -- in the current device, we are already doing that to a large extent. We will expand on what we have at the moment. We've built on the capabilities of Akida I. So in the WA research center, we will perform advanced research into artificial intelligence, which goes well beyond recognizing an image or avoiding an obstacle. For instance, when something isn't contemplated, chip has never seen and it needs to take the right response. So this is now in us AGI, the next-generation of AI. And the next-generation of AI is AGI. AGI stands for artificial general intelligence, which means learning more like a human being. So Akida II will be an advancement of Akida I, more capabilities, each generation of Akida. So Akida III, Akida IV, et cetera, will be an advancement on the previous generation, and each is developed with a vision of the -- of what -- where the AI market is going. But we don't develop these things in isolations. We develop this in dialogue with potential users. To do that, we employ a number of -- we will be employing a number of computation neuroscientists. The BrainChip research center will also work with the companies and institutions in Perth. We are in discussion with several companies here now. And the aim is to make the BrainChip research center self-funded through joint development projects, government grants, tax incentives, that sort of things. We will have 4 PhDs in computation neuroscience and 4 PhDs with a background in AI applications. And these people will be recruited from the universities around Perth, like Curtin, The University of Western Australia, Murdoch and Edith Cowan University, who all have excellent PhD programs in neuroscience, in the robotics, neuromorphic engineering and neural networks. There's quite a lot of talent here in Perth. And I have interviewed already 2 excellent candidates who I would like to -- we'll have join the research center on a very short term once we have secured some projects. So the facilities that we're looking at is not going to be in isolation. Either we will work with universities, things like the Perron neuroscience center and other institutions like EZONE around Perth. We're in discussion with these people. And we are looking at whether we can use some of their space. We need about 150 square meters. We already touched on communications. BrainChip is a very distributed company, and we have offices in the United States, Australia, France. And we always have been very well connected. So we can connect through a virtual private network. And I can connect [indiscernible] in the office in America. So...

Go ahead.

The -- yes?

I think we should wrap up. This has gone on a little longer than expected, but I hope it was helpful for everyone. Look, for interested investors, we did put -- and this will be posted on our website. It will be lodged with the ASX. But you can see there's been a great deal of activity and writing about neuromorphic generally, BrainChip specifically. So we just included this to make it easy for anybody to get to what's been published since. I think that goes back to January. The Akida integrated circuit schedule. Tape-out. Tape-out stopped from this November because no one used tape for the last 20 years. But the GDSII file transfer as well as wafer fab fabrication starts on April 8. Intellectual property is currently marketed heavily. We've got lots of activity in Asia, in the U.S. and Europe. And Peter just touched on advanced research. I think we've covered all the questions. I actually have them all printed out here in front of me. I don't think we missed anything, but if we have, I'm sure you folks are not shy, and you'll send us a note. So with that, I'd say, look, this was -- this is kind of a one-off. I think it was important for us to have Peter and Anil participate for us to have an update, not waiting for a 4C in April. But there will be a 4C in April. There will be an update. An AGM coming up. So there'll be lots of communication over the next weeks and months. So thank you all for joining us, and we'll talk to you soon.