DECEMBER UPDATE WEBINAR PART 3
Those data to spike converters are really what allows us to take virtually any application where you've got large amounts of data, you want to correlate it, you want to recognise repeating patterns and correlate. The execution engine also includes the training methodology. I think very unique is that this will have many, well not many, several training methodologies. We'll be able to do on chip inference which is basically deep learning large data sets. You train with, you train on middle network on a large external data set. Then you put classifiers. I've got, you know, 10, 100,000 images. This is 1000 images that I'm looking for, I'm training the network. What's very typical on the deep learning environment and convolution middle networks. It's going to be very unique to Akida.
We'll do unsupervised, autonomous learning. We'll do supervised, autonomous learning so that if you put Akida in an edge application, if you first want to train it with a known data set, that's great. If you want -- and Akida you can do this -- you want it to learn incrementally as it ingests knowledge... We'll see a repeating pattern. We won't know what it is but we'll flag that pattern. Then someone will label it and say, "Well, that to me looks like a cat." Or, "That's a dog," or, "That's a person." “It’s a tree.” You know, all of those would probably have been in the original data set, but you get the example. Now you've incrementally enhanced the knowledge of the AI edge device. That's going to be very, very unique. We've tried to make this as easy as possible for end users. All of the interfaces and tools are very well known, very well wornpath -- Python scripts. We'll have a modelzoo where we'll have things that we've developed internally or collaboratively. These are millimetbasic applications. So it works well on MNIST, CIFAR-10, GoogLeNet. When the cyber thing is done, we'll add cyber security. This is basically a repository of developed applications.
I'm not going to go through what's on the left, this is the Akida architecture. I will say that it's very complex. What we managed to squeeze into this chip is really remarkable. You know, we're right now we're back to the bullets. The FPGA design is virtually complete. I do want to remind everybody that the FPGA design is for us internally, for us to debug and make sure that the logic works. FPGA has got 100 times less density than the basic that we'll built. You know, some customers may want to play with it but they won't be able to run a full application on it. You know, if you're going to take 100 to one, you're going to reduce 1.2 million neurons by 100 times. They're going to reduce 10 billion sinaps by 100 times. For us internally it's what we do to make sure it works before we move to the next step which will kick off in January, which is called a hardened IC, the logic design for the Akida device. It'll give us more confidence that everything is risk reduced and then we'll go into the logic design. You know, we've got something that we've really got our arms around.
Many facts are repetitive, well there's a few questions about this and apparently there's something called WikiChips which I did go take a quick look at when somebody sent me the link. From recollection, most of it looked reasonably well done. At the bottom, I guess there was some blanks that had to be filled in. Whoever generated that Wiki page said that we had selected TSMC as a vendor. TSMC is a fad. They wouldn't really be, they would only be a vendor for wafers. We are in the middle of that vendor selection process. Not just a vendor selection itself, you know, TSMC, Landry's, pick your choice. The process technology is something that is really important for us to destil right now. I don't want to go into too much science conductor talk but Moore's Law, as things get smaller and smaller and smaller and smaller, and today the bleeding edge is 14 millimetres, that's not actually a bleeding edge anymore. Close to bleeding edge is 14 millimetres. In fact, summer actually down to seven millimetres now. It's great because you make things very small but the mask sets are extremely expensive. The wafers are extremely expensive.
We have to look at these cross over points of what die size based on what process technology. Can I hit the lowest power possible so that I can be on an edge device? If we're package limited, meaning that we need somebody that interconnects on package, if there's interconnect mean that the package has to read 15 millimetres by 15 millimetres, then why try to get the timing's die in the package? Go for the cheapest mass set and go for the cheapest wafer provided that you can get low power and all of the electrical characteristics. We're in that process now. It'll probably conclude in January. Might stretch out a little bit longer. It's not, it's not that important in the development as much as it is getting our team in place and making sure that the design of what we're doing is following the design rules of whatever fad would be in play.
You know, manufacturing process for semiconductors, there's a lot of pieces to it. You know, when we go from logic design which we'll call RTL. If we go from RTL, you know, then you have to go through synthesis and verification. Those are two very big steps. Then you have to go through physical design which is placed in or out. It's a very expensive part of the process and it's, we're literally laying out all the wires so to speak. Once you got that done, you have to cut your mask set. Once you cut your mask set, then you launch that into a fad. It comes out, you slice it and dice it into chips into die. Then it goes off to package. We call it assembly. It goes into assembly. Then it comes out and goes into test. There's a whole bunch of vendors.
There's an opportunity for us to go down a different path which is a pretty well worn path for formidable fabulous semi conductor companies where they'll turn key with, I'll give you names of companies. We haven't selected anybody. You could do it with Toshiba. You could do it with Samsung. There's a company called Open-Silicon. They do cradled from beginning to end of that process. They deliver you a landed IC. Your cost for the IC is going to be a little higher. You will have saved the capital cost of the expensive mass set. You'll pay them NRA, non occurring engineering. Those are decisions that we're going through now. Neil's done this many, many times. I've done this many, many times as has Bob and Peter. We've got our matrix all built out. We'll select what we think are the best vendors and what is the best route. Either go turn key or try to, you know, build an, as I said, and all of the infrastructure that goes in place, the added risk that goes in place. It really feels like a turn key provider is probably for a first IC the most likely choice but we haven't made that decision.
A little bit more about Akida. I want to make sure that people really understand the opportunity for Akida is incredible. The AI market today is really based on what's to the right, a little RNJ I guess you'd call it. It's in the data centre. You've got companies like intel with their CPU. They also acquired a company called Altera which is an FPGA company that they use for acceleration. They've got a couple with Xilinx which is an FPGA company for acceleration. Envidia has moved it to a GPU for acceleration. We've got startups that are probably the same kind of page that we are, companies like Graphcore and Grock raising lots and lots of money. You know, they have hundreds of people working on this stuff. In a lower data centre, it'd be very high SP. It really be based on their accuracy and whatever latency, low latency that they can deliver.
I'm going to talk a little bit about how Akida as a software solution could potentially actually generate revenue for us even before the IC comes out. Let's put that aside for a second. In a data centre arena, you've got big data which is, you know, financial technology, agricultural technology, manufacturing technology. You've got cyber security for infrastructure and small business. With what we've got out of the three university cyber security which is a native SNM. It's not a standard SNM conversion. What we're excited about is the data to smart converter is proven to take cyber security data and turn it into spikes. Then process those spikes with high accuracy. That could be an opportunity for us to do something like a GPI solution where we take our software, we put a wrapper around it and we pick a customer or two that really are on the leading edge or dominant players in the cyber security mark and partner with them. I'll touch on that again.
What I want to focus on is the edge. It's a really large volume, if you look at IOT itself, the number's mind boggling. In 2025, the forecast is 200 billion connected IOT devices. When you think about 200 billion, there's going to be a lot of that int consumer space, extremely low price but there's a lot of opportunity for us in vision, surveillance, robotics, drones, advanced driver assisted systems. These are places in vision where we've got a great expertise, again hearkening back to my comment about SpikeNet. IOT, industrial consumer, medical. That is going to be our primary focus. That's going to be our primary targets. I'll show you some representative customers. Not leading you to believe that these are our customers yet but I'll give you some sense of who it is in that universe that we would be talking to.
I have to put it in one pretty picture, ADAS, you know, autonomous vehicles, it's kind of a sexy thing to talk about. We're not going to, they'll make no money on autonomous vehicles for many, many years. You know, it's going to be a slow roll out. It'll start with very controlled environments, port authorities and other places where there's not people walking around and other cars driving around. ADAS, there's five levels. ADAS level one, two, and three are in large volume now. You buy a car today and it's got radar in its bumper so that when you're backing up and you're trying to park, it can give you distance between you and the curb. Some of the cars these days can actually park themselves. There's many, many sensors in a vehicle that are not necessarily related to autonomous vehicles. It's about driver assisted systems.
I'm going to talk in a minute about a few ways for us to get in here. We can deal directly. I mean, we are talking directly to all the major automobile manufacturers. In the case of the sensors, they tend to buy those modules from first tier vendors. We've talked about companies like Active and Valeo and Continental. It's far easier for us to intersect with module manufacturers. Then if you peel it back even further, there's the sensor manufacturers who provide integrated sub systems for the modules. Then you've got companies like Analog Devices, NXP, ST Micro, ON Semiconductor. They're looking to add more value so they can harvest as much out of that module universe as they can, as well. Our focus is going to be level one, two, and three ADAS. Certainly we'll continue to work on autonomous vehicles. It's fundamental and the same applications. We want time to money.
We can look in the top right hand corner, medical devices. This happens to be a hot spot for me. There are an increasing number of remote medical devices, whether it's heart monitors, blood pressure monitors, diabetes insulin pump, constant glucose monitors. These things are all becoming IOT devices. They talk to the iPhone and the iPhone talks to the cloud. It can tell you what to do when your blood pressure hits 145 over 110. Bottom right hand corner is really just the IOT universe. We can see all of the sensor inputs. Many of which are appropriate for us. Some we'll have to take a deeper dive into. When you look at position, machine and vision in particular is a hotspot for us, acceleration. You can look at the slide. I do think that at some level, at the edge, there will be cyber security.
I talked about us taking the Praxis University IP, productising it within the Akida development environment. Maybe productising it, let the GPI solution for a couple of key customers. They're still doing things at the enterprise, not at the edge, but cyber security is going to be important at the edge as well. The denial of service attacks that hit the US government, I guess at this point probably 15 months ago, took their major, major players because they launched a denial of service attack from like 10 million IOT devices. We're able to get malicious code during firmware updates into IOT thermostats, IOT doorbells, IOT devices of all sorts. At the time, led all to launch requests to a specific DNS all at one time and basically brought that facility down. You know, I don't remember the major players but I think the Twitters of the world, they were all impacted as well.
Akida at the edge which BrainChip is and Akida at the edge, there's not a lot of competition because the spiking neuro networks are, we believe, the best way to reduce power dramatically and provide enough processing power to do some very sophisticated things at the edge. You know, they're an automobile. They're not sending four 4K camera and video streams which hog the bus and suck up CPU power in your trunk. A lot of that decision making could be done at the edge or at minimum, the metadata we said, I just took them four video, four 4K video streams. Really all I saw was there was a kid in front of the car. That's the only data that has to go back. You don't have to send the entire data stream from all 4K cameras. You just send the metadata back.
The next slide, again, these are representative potential customers. I wanted to give you some names just so you get a sense of the universe. When you look at the edge, this could be Akida as an integrative circuit, one that comes out. Another opportunity for us to fester time to money is in some of these applications, they're not going to buy an integrated circuit. When you look at a cell phone, you tear it open, there's probably three integrated circuits in there. They're not going to add another out of sympathy. As we talked about with one of the Chinese cell phone manufacturers, they would entertain to get our IP block, integrating it into one of their customer IC's and then pay us a royalty, an upfront license fee and a royalty. This wouldn't be just limited to the cell phone environment but there's lots of environments where they do their own custom chip because they want to reduce this, you know, the IC count in whatever the application is. It can be a revision, some of these camera manufacturers we'll talk about in a moment. At the edge would be Akida as an integrated circuit or potential an IP block.
To give you example of names, Envision, company called Hikvision and Dahua. Those are the two largest surveillance camera manufacturers in the world. Both are in a place called HangzhouChina. They've been in business for a long time. They would be potential IP plays. They might be integrated circuit plays. ON Semi is used to be part of Motorola 20 years ago, spun out. It's a semi conductor company, very large. They are a major player. I won't quote Dominic because Dominic uses another word, but they're a major player in any senses. They have the ability to co package, sold together. They integrate on IP. I just use ON as an example. There's several in that space, as well.
You know, just as one, maybe first tier suppliers in the automotive industry Valeo. Again, it could be Active, it could be continental. There's a handful of first tier suppliers. Then IOT, again, ON because they're very big in transducer, transducer conditioning. At the edge, analog devices, smart transducer is really one of their strengths. They've been doing that for a long time. Then there's Bricks, we'll call them Bricks. NXP, that's the semiconductor division, formally electronics. It's kind of what you do in the French, Italian company which is also very good at sensing those sensor technologies.
Data centres that said Akida software is, looking at companies like Silence which was just bought by Blackberry, Cisco approved point, I could name 15. I just picked three kind of off the top of my head. Big data, financial technology, agricultural technology, manufacturing, data is data patterns or patterns. With the acquisition of the IP and at Praxis, we probably accelerated our opportunity in cyber security by a year or two. Again, I think it was the best $35,000 we spent.
We've got some questions on the competitive landscape. Across the top are, you know, kind of the private or in our case public but public startupAkida. There's a company called AS Cortex. They do a device with interesting front end use, and analog multiplier, and what they call sub threshold logic which is adjunct gates at.2 volts. It's a very low powered device. I don't see them as a competitor. I don't think they have enough horse power. We've got a lot of horsepower in a very low powered device in Akida. You asked about, they call them GrAI Matter Labs. I think that's the way they pronounce it. It's really, this is really a universal research project. ETA Compute, ETA Compute is still doing mathematical computation. They're not doing the spiking all that work. They do have very, very low power consumption. Again, from what we've seen, they're not going to be a competitor for the applications that we're going to claim. They're just not likely not going to be enough horse power.
Of the big companies, you've got MovidiusTegra and we've got, Intel bought a company called Movidius. Again, these are using matrix multiplication and lots of computation. They'll never get the power man to the place that it will be. I think in applications or use cases where you can run an SMN to solve the problem. You know, we'll be at the shoulders. This is all about our event based which we are. Or you're trying to do math based at the edge of IOT solutions. RVO, we're getting lots of validation from our initial customer contacts. You know, if you can be in the event based world, solve the use case and win us the lowest power possible, that's a win, win, win, win. Low power, low cost, small die and, frankly, lower development cost. I know it's costing us millions of dollars, you know, a quarter to run this company. We think about the private companies that we do on Enterprise or cloud stuff. It's always at two or 300 million dollars, with two to 300 employees, 400 employees, probably going higher. While the burden's never comfortable, we're far more cash efficient than many of the players that we've talked about.
This is a summary of I believe we've got a strong team. We've done a lot of work this year. Our research team is five or six deep. The development team is primarily focused, located in Newport Beach. We've got well protected, intellectual property. This provisional patent has, I think it's gotten 18 filings in it and we'll hire and we'll subdivide into multiple patents. We have a good foundation at BrainChip Studio. I was disappointed as anybody about the lack of revenue in 2018. I think we just did not understand the sale cycle, the trials are there, the design ones are there. Eventually that will turn into revenue. We just really, we really missed it on the sales cycle and the process for law enforcement and homeland security or national security.
There's probably interest that we're getting for Akida to edge from global companies, it is remarkable. I mean, to use the old expression, the phone ringing off the hook but it would be hard for me to pick a major global company that's interested in the edge that we haven't had contact with. You know, contact means we traded emails, we've made visits, and we're moving to you know, can we take some use cases as proof of concept, work with your on our development environment, show you what the benefit on performance is. All in all, revenue was disappointing. I'll say it again and again. I think that this year was a very big year for the company building the strong team, protecting its intellectual property, and getting Akida well on its way to being a reality which is really a vision Peter had for this company from the beginning and why all of us joined.
This is always awkward for me at the end because I feel like I'm talking to myself. Let me, just to make sure I did not miss anything major. I got the list of questions here. This is an interesting question, can you guarantee to the shareholders that none of BrainChip's technology has been stolen by SM Tech? I don't know if I can use the word guarantee but we're highly confident. That was part of, when we say audit books and records, it also included auditing software. Now, as I said, there were deficiencies in what we got from them. We'll see what we get back from the FOIA request to Lockport School District. This was... The very first thing that came to mind for me and when this dispute started so our legal folks have been all over this.
This is an interesting question, is there a possibility that any French agreement could be delayed to the unrest in France? I don't know. I do know that there's a lot of push back on privacy concerns. Maybe those privacy concerns will now intersect with the unrest. Here, given the original ban imposed on ZTE and the recent arrest of Huawei’s CFO or are they still talking to the Chinese cell phone manufacturer, we'll talk. We'll continue to talk. We're not going to show them anything. At the very least, we're going to go back to school and understanding what the cell phone from the standpoint of AI at the edge what the requirements are. You know, maybe we take those cycles of learning and we go to Korea or we go to Cupertino, some place else.
Do you feel BrainChip's profile is a specialist in SNM technology prior or otherwise and being noticed by the right people like the benefit shareholders? I think absolutely. As I said, the conversation started with major companies in basically all those vertical segments. Talked about Quantum. We talked about Quantum and what's going on at Quantum. I did mention, I don't know, maybe two or three quarters ago at this point that when the press release about our collaboration with Quantum went out, literally the next day we got a phone call from one of the top five in the storage, in the top five areDell, Hitachi, Nedap, IBM, and HP. I can't name which of the five but they've continued to move very nicely. We actually just got off the phone with Bob Beachler this morning. They expect to shift an accelerator card and then playing with the studio software on a server and now want to understand what kind of sweep that they could get with the accelerator sweep. I say it should this week, maybe it's already went to the media, be out next week.
I think that covers pretty much everything. If we get any feedback, again, so we're going to sign off now. Again this was kind of an interim call. Everybody in Australia, enjoy your hot summer. We'll be back on for a conference call after we put out the 4C in late January. Thank you very much.
Ends
(20min delay)