#307 Steven Brightfield: How Neuromorphic Computing Cuts Inference Power by 10x

This episode is sponsored by AGNTCY. Unlock agents at scale with an open Internet of Agents. Visit https://agntcy.org/ and add your support. Why is AI so powerful in the cloud but still so limited inside everyday devices, and what would it take to run intelligent systems locally without draining battery or sacrificing privacy? In this episode of Eye on AI, host Craig Smith speaks with Steve Brightfield, Chief Marketing Officer at BrainChip, about neuromorphic computing and why brain inspired architectures may be the key to the future of edge AI. We explore how neuromorphic systems differ from traditional GPU based AI, why event driven and spiking neural networks are dramatically more power efficient, and how on device inference enables faster response times, lower costs, and stronger data privacy. Steve explains why brute force computation works in data centers but breaks down at the edge, and how edge AI is reshaping wearables, sensors, robotics, hearing aids, and autonomous systems. You will also hear real world examples of neuromorphic AI in action, from smart glasses and medical monitoring to radar, defense, and space applications. The conversation covers how developers can transition from conventional models to neuromorphic architectures, what role heterogeneous computing plays alongside CPUs and GPUs, and why the next wave of AI adoption will happen quietly inside the devices we use every day. Stay Updated: Craig Smith on X: https://x.com/craigss Eye on A. I. on X: https://x.com/EyeOn_AI

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  So neuromorphics is really how our biology does computations in our brain. So we don't have circuits, we have neurons, biological neurons, and they're highly connected to other neurons, and they communicate to each other with pulses. We might call them spikes. So you see a spike of signal go from one neuron to the next. And if that spike is strong enough with other spikes coming into that same neuron, that triggers that neuron to spike downstream. So it's a sequence of each neuron is getting spikes inputs and deciding whether it spikes the output. So it's a kind of simple concept, but you put a billion of those together and connect them all together, and you have really what the brain is. We're augmenting our business model

• 0:45Speaker 00

  so our customers can see the value of our IP before they make the risky decision to go and integrate it into a chip developer that might cost them tens of millions of dollars.

• 0:56Craig S. Smith?

  Build the future of multi-agent software with AGENCY. That's A-G-N-T-C-Y. Now an open source Linux Foundation project, AGENCY is building the Internet of Agents, a collaborative layer where AI agents can discover, connect, and work across any framework. All the pieces engineers need to deploy multi-agent systems now belong to everyone who builds on AGENCY, including robust identity and access management that ensures every agent is authenticated and trusted before interacting.

• 1:40Craig S. Smith?

  AGENCY also provides open standardized tools for agent discovery, seamless protocols for agent-to-agent communication, and modular components for scalable workflows.

• 1:54Craig S. Smith?

  Collaborate with developers from Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat, and more than 75 other supporting companies to build next-generation AI infrastructure together. Agency is dropping code, specs and services, no strings attached. Visit agntcy.org to contribute. That's A-G-N-T-C-Y dot O-R-G.

• 2:27Speaker 00

  Hi, I'm Steve Brightfield. I'm the Chief Marketing Officer of Brainchip, and my background has been in the semiconductor business for three decades plus now. I was trained as an electrical engineer, Midwest, Purdue, and I went off and was focused by the first, two-thirds of my career on designing digital signal processors into semiconductors and solving use case problems in, from consumer to military use cases with digital signal processors. I worked at, you know, a lot of large semiconductor companies, but my longest stint

• 3:06Speaker 00

  was at Qualcomm. And at Qualcomm, I worked to launch their first smartphone chip. And it's always interesting, I was in product management at the time, is that you never know what product is going to be successful in the future. And there was a lot of debate and arguments

• 3:23Speaker 00

  about why would we want to build a smartphone chip? Because nobody's going to want to watch videos on their phone. And, you know, my phone makes good calls right now. So we fought with the management and we launched that product. And of course, you know, everyone's got one in their pocket today. And I think if I look forward to Brainchip, everybody was going to have a Brainchip in their pocket, or their lapel, or their hat in the future. And they're thinking right now, why would I want that? Right? So it's

• 3:55Speaker 00

  It's been an exciting journey for me, seeing, you know, products that I worked on that used to consume a large, you know, refrigerator, then go to a box, then go to a board, then go to a chip, and then go to a little tiny piece of IP that's inside of a chip. That was my journey when I started working on GPS back at the very beginning, when it was literally the size of a printer was your GPS receiver. And then we worked on making it smaller and smaller. And now it's in practically everything, and you would never imagine that, right? And I think that's what's going to happen with AI. You know, we see AI today as these big data centers, supercomputers with nuclear power plants next to them. But really, what's happening in the industry is AI is dropping into all these everyday products that we have today.

• 4:40Speaker 00

  Your Fitbit, your smartwatch, your mobile phone, your... Every consumer product you're touching is going to have a little bit in it. And I think that's what I'm excited about Brainship, because Brainship is focused on those small consumer devices that put AI in there to help the individual that's wearing it, not these big, you know, AI data centers that are gobbling people's information up and using them to, you know, sell advertising, let's put it that way.

• 5:10Craig S. Smith?

  Brainship neuromorphic from the start? And let's define neuromorphic for listeners who aren't familiar with it.

• 5:18Speaker 00

  Sure. Well, let's talk about what neuromorphics is first, and then it'll be easier to explain Brainship. Because neuromorphics is really a study of the brain. It really is. And it means a lot to different things to people, and it's kind of good to baseline that. So neuromorphics is really how our biology does computations in our brain. So we don't have circuits, we have neurons, biological neurons, and they're highly connected to other neurons, and they communicate to each other with pulses. We might call them spikes. So you see a spike of signal go from one neuron to the next. And if that spike is strong enough with other spikes coming into that same neuron,

• 6:02Speaker 00

  that triggers that neuron to spike downstream. So it's a sequence of each neuron is getting spikes inputs and deciding whether it spikes the output. So it's a kind of simple concept, but you put a billion of those together and connect them all together, and you have really what the brain is. And it's really, it's been an evolution of millions of years to come up with this architecture, right? And the key metric of it, it's super efficient. And it had to be, because if you look at evolution, evolution was about survival of the fittest. So the animal that could think better would survive better.

• 6:42Speaker 00

  The challenge was he had to have it low enough power so he could cool it off, right? And he could power with, by eating and so forth. So, you know, I think, you know, your brain is about a 25 watt light bulb, if you think about that. That's the power of computing, right? So how do you kind of take those advantages that the brain has and put it into the chip business, right? And that's where BrainChip really came from. It was designed by a guy, Peter Mandelbaum, one of our founders, who was the brain of the company. And he was studying the biology of the brain and these spiking networks. And then we had a Neil Mandar, he is the chip guy who worked in Silicon Valley here for many years,

• 7:26Speaker 00

  building modems and other chips. And together, they basically formed BrainChip to emulate or inspired by the brain's architecture, they created a digital architecture of a computer architecture. And what we do is we take the spikes and we convert those into digital signals, really.

• 7:47Craig S. Smith?

  And can I just, for people that don't know, the neural networks that dominate AI today, when, when

• 7:59Craig S. Smith?

  When one layer or one node of the network passes information to the next node, that information is, is, goes through a computation and is passed on regardless of how important it is to the output. But in neuromorphic, uh, the, the, each node has to sort of gather uh, inputs till it reaches a threshold before it passes information along. Is that a fair?

• 8:27Speaker 00

  I couldn't have said it better, Craig. You understand this. I'm glad that we had our, we've been talking about this. So, yeah, I think people, you know, when people think about AI, they think about data centers. And then the next thing they think about is Nvidia. And if you gotta look back how Nvidia got here today, they didn't create this to build AI. They had a graphics processor that they had to compute every pixel. So they couldn't do that with a conventional processor. So they designed this graphics processor that you can have lots of processors in parallel doing the pixel computations. Now what happened was Jensen Huang learned that

• 9:09Speaker 00

  the scientific community was using his GPUs to do scientific computations. They were using it to do physics problems, do math problems. And those maths problems were matrix multiplies, could be paralyzed very well on the GPU. So he decided to invest heavily in making them better at doing linear algebra or matrix multiplication math. And I think it was prescient of him to find out that when in, you know, the uh, 2012 to 2014, when the great inventions of deep learning happened, they have used the same matrix multiplies. So he had a ready-made platform for them. Uh, the problem is, is that a matrix multiply, like you said,

• 9:53Speaker 00

  every data in the matrix has to be multiplied by every single data in the matrix by the way they're architected. And the problem is, is that if half of the values are zeros, it doesn't matter. You're multiplying zero times something and you're gonna, you know, you're gonna get zero, but they do it anyway. Because it's, it's, it's just a brute force way you do algebra, right? Um, and uh,

• 10:19Speaker 00

  To create brain chip, we just could not follow that path. So what we did was we create more of like a data flow processor where the data flows in.

• 10:29Speaker 00

  It charges up these neurons and if they fire, then it goes to the next level and computes it. But if it doesn't fire, the next 10 layers of that computations, they're not computed at all. So that's essentially the the secret sauce of neuromorphics versus the traditional AI. And I like to call it, you know, traditional AI is brute force math. And neuromorphics is a brain inspired elegant computing, right? It's very elegant how it works, right? And and it took millions of years of evolution to get there.

• 11:05Craig S. Smith?

  Yeah, I had on the program, maybe a year or two ago, a guy from Australia at Western.

• 11:15Craig S. Smith?

  Geez, I can't remember the name of the university. They're doing a brain scale neuromorphic computer. That will have as many connections as the brain. Who knows what good it'll be, but it'll it'll be an interesting. Exercise for for studying the brain or for studying neuromorphic computing. But but what brain chip is doing, because Intel has a neuromorphic chip. And I think Intel does IBM certainly does North Star, right? And but those are largely research.

• 11:56Craig S. Smith?

  Products, isn't that right? Or are they putting them into commercial products?

• 12:02Speaker 00

  Intel has what's called a low high chip. And that low high chip is neuromorphic computing chip. And IBM has True North, it's also a neuromorphic computing chip. And both of those guys had big research organizations and they kind of funded that, right? Brainchip actually didn't have a product or a chip for the first half of its existence. It was incubated in 2009 and I think 2011. He started, you know, opened his garage up and started earnestly building on it. Peter Enneman. And it wasn't until 2014 that he brought on Anil, the chip guy. And it still took him three to five years of research before they said, okay, we got something. We want to make a product.

• 12:43Speaker 00

  And then the last six years have been productizing that architecture and getting it designed into some very interesting use cases.

• 12:51Craig S. Smith?

  I'm sorry, that's the Akita processor.

• 12:54Speaker 00

  That's correct. Akita is Greek for spiking. So that's spiking network that we talked about, right? So the difference between what Intel and IBM is doing is, is that since they kind of had that research, they never committed to a product with it. But they did manufacture some chips. And a lot of researchers have done a lot of work on those with very interesting and promising results. And I think you'll see that even IBM is scaling up and building a very large neuromorphic computer with their chips. But it's another government research program. The primary difference between Brainchip and the Intel and the IBM solutions was they were analog.

• 13:39Speaker 00

  So they truly tried to match the analog waveforms of the brain. Whereas the Brainchip made a digital equivalent of the analog waveform. So now you could easily manufacture a digital computer chip using the approach. The chips that you, the analog chips that are made today for neuromorphics, they're notorious for, you know, you have to have them biased and temperature stabilized. And there's all the problems with analog, which is the reason we don't have a lot of analog computers today, are the problems that they're faced with, with their neuromorphic chips. So one of the major innovations is changing it from an analog spike to a digital event.

• 14:23Speaker 00

  So we call it event-based computing in a way to differentiate it from a pure-play analog neuromorphic computing technique.

• 14:31Craig S. Smith?

  Yeah, and and you can by putting these on the edge, you can move move inference to the edge. Is that right? Are they?

• 14:39Speaker 00

  Absolutely. So most of the training is still done brute force with the matrix multiplications. Right? So that's where Nvidia is selling a lot of these big brute force boxes to train on massive amounts of data. But once you train the model, now you can put that in a tiny device on the edge and do what what you said was inferencing, which is actually, you know, trying to infer, you know, what you're looking at or what you're hearing or what you're, you know, sensing, right? So that inferencing is done constantly on the edge. And it makes a lot of sense to do the computations at the edge. So you don't have to send the data from the edge

• 15:21Speaker 00

  all the way to the cloud, compute it, and then send it all the way back. And one is latency. It's a round trip. Two is cost. It costs money to go run on that server and bring it back. And the third is privacy. That data goes someplace and you don't know where it goes. And I think we've seen a lot of cases where large companies doing this, they're they're not respecting the data, the privacy rights. And this is still, I think, one of the most interesting surveys I saw about AI was 70% of the people think AI is good and it's going to be helping do things better and easier. And then the same population of 72% said they're fearful for their privacy of their data using the AI.

• 16:04Speaker 00

  So it's like a double edged sword. You know, it's going to help you, but you also know it's exposing you at the same time. When you compute at the edge and you're not sending your data out, it kind of helps solve that other side of that problem. So I mean, you do see the whole industry moving to the edge. It just takes longer because when you have it a data center, you don't have to plan anything. It's it's like an ocean of computing. But if you have a bathtub of computing at the edge, you got to make sure you can fit your problem into it, right? So it's fitting the the inference problem into the amount of compute you have.

• 16:39Craig S. Smith?

  And you were saying that someday, everybody will have a brain chip in their pocket. It's computing, the data that it's computing on is from sensors at the edge, right? It could be vision, or audio, or heat, or pressure, or anything. Is that right? What are some of the use cases that you guys see?

• 17:06Speaker 00

  Yeah, we can support any sensor data. So anywhere from vibrations to microphones, to cameras, to radar, lidar, ultrasound, even chemicals. We actually had a demo where we could do smell detection with our AI. And we could detect what kind of beer it was by tasting the beer and looking at the chemical sampling of it. So what really happens in the, when we say the edge, sometimes we say on-device. So it's right on the device that you have in your hand. It's not going off the device. The other thing, term that's popular now was called physical AI. That means that this is AI that's interacting with the physical world.

• 17:50Speaker 00

  It's not in some tower someplace with a bunch of data computing away. It's actually grabbing data continuously from a sensor and computing it right there on the fly. And we call that streaming data. And when you have data streaming in, it's going to demand you do something with it. Because you have only two choices, you either compute it, or you store it. And, or I guess three, and then you, or you transmit it to someplace else to compute it and store it. So it's a lot cheaper just to compute it right there at the edge.

• 18:24Craig S. Smith?

  Are there things that you can't do, computations that are just too heavy?

• 18:30Speaker 00

  That's correct. Yeah, that's correct. We designed our architecture so really focusing on the detection and the classification, using the neuromorphic principles. If you go to LLMs and these other algorithms that are beginning really popular in generative AI, they're not necessarily a great target for that. Because they have a different computational profile, right? And They they actually deploy those matrix multiplies effectively. We do have technology at Brainship that we think is a different technology, which is a a different neural network architecture than transformers. Transformers is the the the foundation of

• 19:12Speaker 00

  of large language models and generated AI today. But we've adopted what's called state space models, which is actually a an innovation of transformers that's more computationally efficient.

• 19:25Craig S. Smith?

  Yeah, and it's kind of the the the the leading edge right now. A lot of people are shifting to state space because it it has better memory and and recall and and all of those things. So

• 19:41Speaker 00

  It's got less computations, too. Yeah. One of the things that I think people need to understand the edge is is that just because you have a microphone there or a camera attached to your computer, right? Sometimes nothing's happening, right? There's no audio or there's no there's audio, but it's just noise. There's not anybody speaking or anything. And the camera, the camera could be staring at a blank white screen, waiting for something to fly through it. But, you know, the level of activity in the camera is quite, you know, a lot lower than the amount of data coming out of a camera. So this, how much information

• 20:22Speaker 00

  is in the data stream is what we exploit with neuromorphic computing. The analogy I always like to give is if you're staring at a blank screen, your computer isn't working overtime calculating every pixel. Your brain would overheat, right? It's but when things move in that screen, you can detect it and process it and classify it instantaneously, right? Now, cameras today were designed like television sets. Every thirtieth of a second, they put a new picture up and they fake your brain out into thinking it's moving, right? So when we do inferencing, people put cameras hooked up to AI and they did the same thing. Every thirtieth of a second, they give a whole frame of data and they'd say, go compute

• 21:07Speaker 00

  it all. And if nothing happened in that in that thirtieth of a second, it would still compute every pixel as if it was gold and then come back and say, well, we didn't see anything. Neuromorphics is saying, compute the changes in the scene. So if you see a change, it's you generate a spike. So that blank screen is going to consume almost no power in the AI processing because there's no spikes going into the processor. As soon as spikes come in, it lights up the network and it starts generating spikes and pulses. And how those propagate through the network is how you recognize what that is being seen in that scene. Same thing in a microphone.

• 21:47Craig S. Smith?

  You, last time we spoke, you gave a very good explanation or example of doorbell cameras and, and how they, you know, they just have to process this continual stream of data, even if it's a, if there's no movement outside. Is that a, one of the use cases?

• 22:13Speaker 00

  That, that is actually, um, you know, originally when those cameras came out, they would stream the data continually from the wifi over your internet connection to Amazon or somebody. And then it realized, we're not making any money at this. It's costing us more to compute it than it is that we can charge the customer, right? And the customer, they didn't want to pay a hundred dollars a month to stream the data, just in case somebody walked by. So they've got intelligent. They're still using like this brute force technique, but they have a little detection that, you know, that says, wait until we see something move, then we'll start crunching on it, right? But even then, you know, you can miss things and you're still looking every thirtieth of a second. So, you know,

• 22:56Speaker 00

  it, it, it, it isn't like the brain where the brain can instantaneously detect something rather than waiting for that next frame to show up. So what we're seeing is they've made a lot of innovations to do a lot of local processing in a, in a blink camera and everything. But the problem is this, even with all those optimizations, you gotta go change the battery of that thing every, you know, month or so. And, and if it's hung up on a wall and it needs a ladder, it doesn't get changed or it costs you to do that. So you get another 10x reduction in the power with neuromorphics to do it where it's always on. And that can translate to 10 times longer battery life. Instead of every month, every year you change the battery. Now that becomes a, that's a game changer from the consumer,

• 23:40Speaker 00

  right? Uh, just think if your smart watch lasted a week instead of. Nine and a half hours, right? At the end of the day, my smartwatch has ran out of energy before I did it. I'm like, hold on.

• 23:53Craig S. Smith?

  Yeah.

• 23:53Speaker 00

  So extending, yeah?

• 23:56Craig S. Smith?

  I was just gonna say, I have an Oura ring and I've got a like, you know, one of those rings that reads whatever and but but it keeps running out of battery. I don't imagine they were using neuromorphic.

• 24:12Speaker 00

  No, they're not, but but I can't really disclose, but those are exactly the targets that Brainship has in a wearable. So wearables is a huge focus for us because if you wear something, you don't want to take it off all the time. You don't want to have to charge it all the time, but you want it to work all the time. And if if you look at an Oura ring now, it's got a microprocessor in it that it's constantly calculating all the time, just in case something happens, right? And that just sucks the power. If you have something neuromorphic in there, it'll wake up when some data happens, right? And it's going to reduce the computational power. So we think wearables extends into medical,

• 24:54Speaker 00

  industrial, defense, as well as all these consumer products. We have eyeglasses that'll have computations in them. Your earbuds will now become medically certified hearing aids just off the shelf with AI in them. The ring that you've got now, instead of having to send all that data to the cloud or to process it, it'll process it on the ring and it might light up a LED that says, hey, you need to do, you need to drink hydration or you need to exercise or you need to do this without having a $800 smartphone in your pocket and an account to a cloud service that pays you, you pay, you know, $20 a month for it just so that it can grab the data off your ring and give you results, right? So those

• 25:39Speaker 00

  are great products and they built them quickly because they could leverage all that infrastructure. But the consumer would be great if I could use this without some of the limitations it has.

• 25:50Craig S. Smith?

  Yeah. How were, I mean, you know, autonomous vehicles is another obvious application. Are they using neuromorphic now? How are they processing?

• 26:04Craig S. Smith?

  Lidar and image data, they can't be sending it to the cloud, obviously.

• 26:11Craig S. Smith?

  For.

• 26:11Speaker 00

  No, it's it's it's computing. In fact, I think early in the days of autonomy, you open the trunk of your car, it was a supercomputer hidden in there, right? So it kind of significantly took away your baggage. And so much power. You know, I think in the early days with cameras driving the vehicles, they were collecting a terabyte of data a day per car. And it was so much data that they basically had to have these huge magnetic tapes that they would pull out of the back of the trunks every day and load it to the data center for training. Now, one of the great advantages of neuromorphics is is that it's dynamic.

• 26:54Speaker 00

  It's not waiting for every 30th of a second picture. It can one thousandth of a second, it can detect a change in the signal, right? That means you can do lower latency detection, which is really critical in autonomy, right? And you can you can identify an object in, you know, less than a millisecond rather than 16.6 milliseconds, which is the dwell time you're waiting for that camera frame to show up. The other is it can be lower power.

• 27:25Craig S. Smith?

  Are you guys selling the IP or are you designing chips?

• 27:31Speaker 00

  We're doing both and part of the reason is is that people need to see the proof points of the neuromorphics. If we had just IP, it'd be too hard for them to go, how do I solve my problem with this? And try it out before I commit to it, my product, right? So we just announced this week, we're going into volume production on a neuromorphic chip from Brainchip. And we're entering the volume semiconductor business, not so much because.

• 28:04Speaker 00

  We're changing business models. We're augmenting our business model so our customers can see the value of our IP before they make the risky decision to go and integrate it into a chip developer that might cost them tens of millions of dollars. So for example, we have a customer right now using that chip in a wearable device where they have it in smart glasses and they can detect from your brainwaves activities, whether it be migraine. In this case, they're detecting epileptic seizures before they happen. So it's a very therapeutic kind of product, but it's really enabled because you got to do that computations all the time with the person. You can't sample the data, send it to the data center and then

• 28:49Speaker 00

  come back and say, oh, you might have already had an event, right? You want to detect before the heart attack happens, not after it happens to tell you, I just want to confirm that you had what you think you had, right? And on the other side of the spectrum, we found people in the in the communications business and the defense in radios where they're looking for signals and Neuromorphic is perfect for that because it can find these signals in a lot of noise and it can pull them out with very efficient power. We had a customer swapping out an Nvidia Jetson chip and putting our chips in there and he got like a tenfold improvement

• 29:33Speaker 00

  in his power efficiency. And for him, that was the key metric that he had to accomplish because he had a mobile platform. So we, when you, let me round that back to your question, autonomous autonomous vehicles. It's not autonomous vehicles. It's everything that's autonomous, whether it be a robot, a car, a ship, a plane, a drone, anything that's moving, any machine that's moving constitutes a collision risk with a human. So how do you do that? You have to really have this solved.

• 30:05Craig S. Smith?

  And who's who's manufacturing the chips for you?

• 30:10Speaker 00

  We're getting a manufacturer here in the United States by Global Foundries. They have a foundry up in upstate New York. It's a 22 nanometer chip. Yeah, I think that was IBM's facility up there at the time, right? Fishkill.

• 30:30Speaker 00

  And the beauty of that product is it's it's it's a silicon on sapphire semiconductor. So that means two things, we get low leakage. So we don't have a lot of wasted current for these very low power. And the other is you can do radiation hardening. So one of our customers is is doing space missions. They're creating the first rad hard AI chip that's going to go into satellites, manned vehicles, and even any anything that goes into the, you know, space.

• 31:01Craig S. Smith?

  Yeah, where power obviously is is an issue. Power consulting. And and then you're also partnering with with other chip designers, ARM and Intel and.

• 31:16Speaker 00

  Correct, yeah.

• 31:19Speaker 00

  So we have an. We have an ecosystem of partners. Intel is a partner of ours. We've worked with Red Assus, which is a Japanese semiconductor manufacturer. They're a licensee of Akita. At a US company, Front Grade Geisler, they're the one doing the rad radiation hardened chips for space.

• 31:41Craig S. Smith?

  Okay, so you guys are the first commercial producer of neuromorphic IP and soon chips themselves.

• 31:52Craig S. Smith?

  How long do you think before there will be? This will this will be taken up by industry because they're still operating on supercomputers in the trunk of the cars, so to speak.

• 32:06Speaker 00

  There are new companies that do have. We were, that's correct. We were the first commercial provider of neuromorphic IP and chips. We've had chips for like five years, but we use those as development platform. And what the customer said, look, I can't wait to do a custom chip. Can I just buy this chip and do my first generation product with the chips? And we argued with them for a little bit and we said, yeah, I think we can do that for you. There are other companies that are producing analog neuromorphic chips, but they're kind of dedicated for a specific market segment like speech wake up, right? Or a biological wake up. So they're like function specific neuromorphic chips.

• 32:48Speaker 00

  We have a very digital programmable chip that can use any kind of sensors. So we're kind of unique in that aspect.

• 32:54Craig S. Smith?

  Build the future of multi-agent software with AGENCY, that's A-G-N-T-C-Y. Now an open-source Linux Foundation project, AGENCY is building the Internet of Agents. A collaborative layer where AI agents can discover, connect, and work across any framework. All the pieces engineers need to deploy multi-agent systems now belong to everyone who builds on AGENCY, including robust identity and access management that ensures every agent is authenticated and trusted before interacting.

• 33:38Craig S. Smith?

  AGENCY also provides open standardized tools for agent discovery, seamless protocols for agent-to-agent communication, and modular components for scalable workflows. Collaborate with developers from Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat, and more than 75 other supporting companies to build next-generation AI infrastructure together. AGENCY is dropping code, specs and services, no strings attached. Visit agntcy.org to contribute.

• 34:20Craig S. Smith?

  That's A-G-N-T-C-Y.org. Yeah, but, but do you see, I mean, you know, there is this big move to the on-device computing or to the edge. Uh, this is a solution that solves the power problem and, and, uh, I, I've been the, the form factor and it's much smaller system and the transmission systems and all those different things. Uh, but, but it's not in my house right now. You know, I have a thermostat, a smart thermostat.

• 35:03Craig S. Smith?

  I have, you know, a smartphone. I have hearing aids. Uh, but when, when you, and I want to talk about hearing aids, when do you see this really entering, uh, the computing, uh, infrastructure around us?

• 35:22Speaker 00

  Well, I think, I think we're seeing rapid adoption now. One of the challenges was is that since we weren't doing conventional networks, that there was a programming barrier, right? It was adoption, right? AI was really, here's all this open source code, it works, you push the button and you brute force compute it and wow, you got a good answer. It's kind of like ChatGPT today. No effort, it's all off the shelf, you type into it, out comes the answer. It's still really expensive for them to do and they're spending venture capitalist money to give you a free taste of this before they get you signed up, right? That's what's going on. But to your question, I think we're trying to ride the neuromorphic computing and brain chip in particulars,

• 36:07Speaker 00

  trying to ride the coattails of the overall market moving to the edge. And when we look at market research reports from companies, they're saying about 10% of these edge products, embedded devices are running some AI software on them. But within the next four year, four to five years, 30 to 35% of those products will have AI on. And I think if we look out the next five years, 90% of them will have it all embedded in it. And there will be a neuromorphic computing in probably half of those devices, because it's going to be more generally available, it's going to be more understood. And, you know, you needed a product out there for people to

• 36:49Speaker 00

  just quickly grab off the shelf and adopt. And they couldn't kind of make a calculated investment in it. We're changing that by going and offering chips as well as the IP.

• 36:59Craig S. Smith?

  Yeah, you know, one of the things that gave Nvidia a lock on the GPU market was its CUDA programming language that people became very familiar with. And, you know, I've spoken to Cerebras and Sambanova and some of the other new GPU manufacturers or inference chip manufacturers. And that's a barrier for them because people don't want to learn a new programming language. And, and

• 37:39Craig S. Smith?

  And as a matter of fact, Cerebras in particular is just shifted to putting its its chips in the cloud and offering inference services. Is that, do you face that same hurdle?

• 37:54Speaker 00

  Yeah, I think the whole industry does. It's that ease of use. It's that push button capability that allows a large number of engineers easily adopt it. You know, CUDA wasn't developed for AI. It was developed for those, those scientific computational guys that were trying to forecast the weather and design weapons and, you know, analyze the structural integrity of a building, right? And those scientific computations was where CUDA was born, right? And it just happened to be, you know, convenient that the main operator in AI is described in the CUDA language. So, and it is, is, you know, it is, is one of the things that we see in our industry

• 38:39Speaker 00

  is getting people interested in normorphics is how can they go from a CUDA programming environment to an Akita programming environment, right? And the answer is they're not going to make that step easily. They've got a lot of investment in code that runs on CUDA. So one of the things that industry I think is doing, and I think Nvidia is supporting this, is making CUDA-like software APIs at the edge. So that I can take some of my code and I can use those CUDA primitives, but I can have them run not on an Nvidia GPU, but they can run on an edge device, whether it be a brain chip device or another manufacturer of an edge AI device.

• 39:24Speaker 00

  And I think that's going to actually accelerate the edge adoption too, because it's going to create a bunch of new capabilities that are transferred from the, the big box Nvidia environment to this ultra-low power edge environment.

• 39:39Speaker 00

  And it's actually good for Nvidia to do this because they keep on CUDA and they actually make sure it's going to dominate the data center space because now they've got the edge rooting for them too and leverage them. And Nvidia had to make a choice, are we going to do edge, are we going to do cloud? And I think they said, well, we're going to do cloud and we're going to allow that fragmented edge market to be serviced by others.

• 40:07Craig S. Smith?

  Well, explain how that happened. So you have a brain chip in your device and you want to

• 40:17Craig S. Smith?

  put a model on it or have it compute a model of that is also on device, I guess, either on the chip or memory or something.

• 40:33Craig S. Smith?

  So in order to do that, you need to use your proprietary programming language for Akita, right?

• 40:46Craig S. Smith?

  So explain how Nvidia is supporting that with CUDA?

• 40:50Speaker 00

  They're not, but what they're doing is they're enabling people to write, take code that was written in CUDA and in areas where Nvidia isn't interested, that they can run it on different kinds of hardware. Now, you asked a previous question is, you know, what are limitations of neuromorphic computing in Akita? And there are, you know, we can't do all the different kind of operators, right? That was the beauty of what Nvidia does is they could, whatever the math operator was, it was supported in CUDA, right? So what we've done is we've combined our Akita with a host CPU and we can run some functions

• 41:32Speaker 00

  on Akita, but if it doesn't run on us, it just passes the CPU and it runs it. And actually, this is what Nvidia does too. Everything doesn't run on the GPU in Nvidia. It runs on ARM CPUs that are embedded into their devices, right? So.

• 41:48Speaker 00

  This is called heterogeneous computing. So it means that like every processor element isn't, isn't perfect for every use. But if you have different types of computing elements in your chip, you just hand that task to the, the, the element that's most effective at it. So for example, you know, we're working, I'm at the RISC V conference in, in Santa Clara today, and I'm partnered with Andi's, which licenses a RISC V course. And we can go to a customer and we can combine a RISC V core with conventional accelerators and Akita into an overall platform and provide a programming interface for them that can stub out some of the code

• 42:30Speaker 00

  and then replace it with Akita and others, we just have it run on the CPU. So, I mean, this wasn't invented by us. This is an industry trend. And I, if I date back to my mobile phone days, the way that all worked is we had a heterogeneous computing platform. If you look at Qualcomm, they talk about their AI today. It runs on the CPU, it runs on the GPU, and it runs on their NPU. And I worked on the NPU when I was there. Now, if we look at today, we work with Andi's with CPUs. And we can offload to a GPU and we can offload even to another AI accelerator sitting next to Akita.

• 43:11Speaker 00

  Now, if the data isn't sparse at the edge, maybe Akita isn't the right accelerator for it. So you just send it to the unit next to it, right? But we can balance out a system so you, you can, you can get a lot of complex models, but they're optimally executed.

• 43:27Craig S. Smith?

  What, what's the application you're, well, let's talk about hearing aids. Explain how, how earbuds could become hearing aids with, with norm more.

• 43:40Speaker 00

  Well, one of the things that we've done is, is created.

• 43:46Speaker 00

  Wake up so we can wake it up easily like a wake up words just like you know you say hello Google or hello Siri right they have a special circuits in those mobile phones to pick up that at low power so it's on all the time but the power is much lower in a hearing aid so you're gonna have that technology we've also created you know these state space models we've created de-noising algorithms that dramatically can reduce the noise in fact when we were at CS last year we did a demonstrator our de-noising where you could listen to the you know talking in the noisy environment and you put these on and you pass it through the AI algorithm and

• 44:28Speaker 00

  you're like wow I just really can hear it well a lot of hearing aids is about selectively producing the right information to the human ear and I think these de-noising algorithms are one the other is is that people don't understand is that LLMs can actually be used in some of these processes because if the LLM knows what's being said it can kind of predict is he gonna say this word and even if it's noisy go oh yep that was that word and then I can reproduce that even cleaner so this is a trend in the industry is people are replacing digital signal processing algorithms with a machine learning and AI algorithms for these signals and hearing aid

• 45:12Speaker 00

  is an obvious solution there because currently you had a digital signal processor doing filtering to try to prove it and they would like let's tune the filter so it matches up right and customizes it with AI ML you kind of you don't need to do that and you actually get better results in fact one of my friends is starting a company doing exactly this who came out of the work we did on doing it on the mobile phones

• 45:40Craig S. Smith?

  I imagine that there's there's a lot of I mean

• 45:46Craig S. Smith?

  Translation is improving the speed of translation. There is a lot of excitement about simultaneous translation. I would guess this would be an application for something for neuromorphic. Where, where, you know, you have it in a listening device, whether it's a headphone or, or earbud or hearing aid and the trick, could it handle that kind of a load?

• 46:14Speaker 00

  We're using neuromorphics for the input signaling because of the advantages of the sparse data, but when we get to some of the large language models, we don't use neuromorphic algorithms, we use state space models. And we're looking at combining those two into a single, you know, platform so that you get the enhancement of the signal going into the speech recognition. The speech recognition with the LM can predict what next word is being said and improve the accuracy of the recognition. And then you can have a local, large language model in your earpiece that could have maybe a very limited set of, of information in it, but it's what you need, right?

• 46:56Speaker 00

  One of the use, interesting use cases is a, a memory LLM for an old person. And, you know, it would say, oh, your granddaughter's name is Shelley and she's four years old. And so that you, when you, you can, if you don't forget this stuff, boom, you have it, right? And you just need some cues sometime to get your memory back. And, and this is one of those interesting things that we, we're like the nationalists to health does, this could be really good because that helps solve a lot of the, these issues with, if you can't hear well, you start having dementia and you start having, you know, problems, cognition problems, right?

• 47:39Speaker 00

  It's very important to have, you know, hearing in sight to keep yourself, your brain healthy because your brain, that's, that's what it's doing. It's constantly processing those signals. When it doesn't have those signals, it starts hallucinating just like an LLM.

• 47:55Craig S. Smith?

  That's right. Yeah. And so you've, you've, you've got these partners that are using the IP in their chips. You're starting to produce your own chips.

• 48:09Craig S. Smith?

  And I would imagine research is ongoing. What, what are, what's around the corner or over the horizon?

• 48:16Speaker 00

  One of the things that was interesting is we got a contract with

• 48:21Speaker 00

  Air Force Research Laboratories to work on radar using these algorithms, right? And the results actually surprised us and they surprised the contracting agency. And now we're expanding that and we think that we can, you know, add capabilities to radar that weren't there before, like for example,

• 48:45Speaker 00

  radar can detect things, right? But it can't tell you what it is. Well, we can classify objects now with radar in addition to detecting them. We can improve the tracking and, and the, the latency of these radars. But we can also make them a lot smaller, right? So it's that size, weight and power. Can I put a radar in a robot? So when its hand has got a radar signal in it and it can, and it can basically navigate, you can paint the scene without a camera. You can use it like a camera to paint the scene and recognize and grasp things that. A drone, you can fly it inside tunnels or buildings indoors, can, can map out where you're going.

• 49:28Speaker 00

  We see this shrinking of the conventional radar technologies to really go into anything moving because it's all weather, it works in the dark. And if it can replicate some of the things in vision, then you know, you don't have to worry about rain and fog and snow and all the, the issues that visual, you know, control of robots.

• 49:51Craig S. Smith?

  Yeah. And are you working with robotic companies or, or is this still in the research realm?

• 49:57Speaker 00

  It's, it's still in the research. We're working with companies that are creating components or solutions that go to the robotics companies. We are in active conversations with robotic companies today and they're in evaluation of this, right? But what we decided was is to create reference platforms that demonstrate these more wholly, rather than having a, you know, here's the algorithm, go figure it out. We will build a little prototype. So we're doing reference designs and radar. We're also going to do this in these wearables. And this is a quite interesting approach. You know, the air tag, right?

• 50:33Speaker 00

  Yep. What about having a brain tag? It's a smart air tag, right? It's got inferencing on it. Say you set it down. It can continuously manage it, but there's a microphone in it and it can say, oh, I can hear the dog, you know, the baby crying or the dog barking or my husband walked in or whatever. You can do all this stuff, but you can do it on the size of an air tag, right? And it can quietly sit there and the battery will last a long time. It'll wake up, it'll Bluetooth, you know, to send the signal that says, hey, somebody just came in the door. You don't need a big fancy, you know, blink camera sitting there. You can just have a brain tag and you can just, you know, use it, you know, in a bunch

• 51:18Speaker 00

  of different use cases like that. So this is one of the ways that we can easily demonstrate to a consumer, hey, this is actually a change in the capabilities, right? And then we talked about large language models and voice. Oh, and one of the other things we're doing is building reference platforms for this voice chats, you know, or what we call it voice assistants, right? And the difference between a voice assistant that you get from Apple or for Google and with the one we're talking about is a private voice assistant. So your voice doesn't leave the device. You do the denoising, the voice recognition, the large language model and output a text to speech in the natural language.

• 52:03Speaker 00

  It goes back into your headset. And we see these use cases for, say, somebody in industrial site, they're out working and they need safety instructions. Don't touch this or do this in the sequence or this is how you disassemble the cooker assembly and the, in the foundry or the, you know, some kind of energy plant or a defense application. You've got somebody out in the field and they're in a situation they don't know. It detects their, their, their blood pressure, their temperature, and it says, hey, you need to relax. You need to do this and you need to move to this location. Talk them through things. So they got somebody helping them solve problems that they've never seen before.

• 52:43Speaker 00

  And there's nobody there to help them. So this has got both the, these defense applications, but it's in enterprise. It's everywhere where you have an employee. What's the company policy when I do this? You know, what's the procedure when I do that? You know, when do I get my training? Your training is in your ear. It's ready to go every instance that you're out there working and it makes you very effective. Right? And so we see the use of what I call enterprise class of voice assistants that have proprietary knowledge of the enterprise that they don't want to share with the cloud provider because it's their secret sauce. You know, you know, this is how we build the special food that you're eating, or this is how we build the special product. You don't want to put that into a, a public chat bot, right?

• 53:28Speaker 00

  Because that's going to be trained on that data and it's going to teach your competitor how to do what you're doing. So how do you protect your company secrets? You keep them inside and if you can create a captive L L M that your employees can use to rely on to, for the, the company manual and for the guys to do this and guys and what not to do, you know, it's going to improve the productivity of a human. And consumers are going to want to use this to plan their, their schedules and what they're doing without having to, you know, type it into a screen and hope the cloud connection, hope A W S doesn't crash and all these type of things to make sure it works.

• 54:13Speaker 00

  I mean, uh, uh, I was working, I think yesterday and A W S crashed and suddenly half the things I was using didn't work. And I was like, ah, that's the cloud. A W S is so reliable that you're not used to it. But when it happens, you're like, okay, I, that, I, that's the gotcha there. Right? So.

• 54:30Craig S. Smith?

  Yeah. So neu- neuromorphic computing has been around for a while. Why has it taken such a long time to find commercial application?

• 54:40Speaker 00

  Well, if you look in the scientific journals, there's lots of reports demonstrating and studies and research efforts reporting the benefits, right?

• 54:50Speaker 00

  It was always, how do I put it into an everyday product? And there wasn't a way to do that. And one of the things is, you need a chip to do it. And you needed a software tool that could easily take processes they're used to, and then convert it to that. So one of the key things we did was we created a tool that converts a conventional CNN or convolutional network to a spiking neural network. Our digital version of it, right? That tool has been crucial for us. And we're coming out with new tools, you know, next month that'll take any model from a format called Onyx, which is used by the whole industry,

• 55:33Speaker 00

  and push button and convert it to a spiking network. This is huge advance for us. And it's, you know, people get too focused on the hardware, right? The neuromorphic chip. But the question is, the software. Nvidia told everybody that's how it works. We know that's what's done. So we've set up a new website called the Developer Hub for Brainship, or developer.brainship.com. And it's all for the engineers. Here's the source code. Here is the tools. Here is a training modules. And here is a forum to meet other people doing this kind of work. We got a university program that university students can get a discount on our boards and get free software.

• 56:15Speaker 00

  In fact, we've just signed with a major manufacturer, a defense manufacturer, for a university contest where they're going to fly drones against each other over an obstacle course, and use the brain chip for the guidance for that. And I'd love to tell you more about that when we're able to go public with that. But it shows you that we've got to seed the market with programmers that know how to use our technologies and provide them the right tools to convert it from the classical formats into the neuromorphic formats.

• 56:51Craig S. Smith?

  Right. And this tool, I mean, you've got...

• 56:56Craig S. Smith?

  Temporal event based neural networks, that's, that's the, the, the models that work on brain chip, right? And then you've got tools to convert other models into temporal event based neural networks. Is that right? And, and are those, those tools, I'm, I'm just looking at some notes, Meta TF tool chain, is that what you're talking about?

• 57:25Speaker 00

  Yeah, well, the, there, there's, is, is that entirely accurate? What the Meta TF tools take the conventional CNNs and convert them to spiking neural networks. Okay. So that was the invention of brain chip. Since then, our temporal event networks are actually separate from the neuromorphics and they're the state space innovations where we were, we were actually, it's an algorithm advance where we change the, the, the basis functions that we use for the, the state space models so that it works really well on this time series data. So we

• 58:06Speaker 00

  can put a streaming time series data in there and get state of the art accuracy with a tenth of the computational power. So we're actually looking to open source some of our advanced neural network models based on tens so that the industry can go, oh, I see how I, this is really powerful and we get adoption. So we're gonna put some open source out there and we're gonna also make a demonstrator that you can see these models working on a mobile phone. So you can like download a, a Google app and you're going, wow, I can run this language model on my phone or I can do this denoising on my phone. I'm interested in putting in my end product. So it's awareness and it's also.

• 58:51Speaker 00

  Education. So that's why we went back to this development network and the university program to educate people on the algorithms.

• 58:57Craig S. Smith?

  If people want to explore this further, what's the URL for that they should go to?

• 59:05Speaker 00

  Well, the, for basically a non-technical person, just brainship.com and we have use case tabs and you can click on there and see videos of eye tracking, people detection, radar, medical, all kinds of, you know, these different types of ways that you can use it. And if you're a technical person, you go to developer.brainship.com and you can look at the source code of what we're doing and you can download the tools for free. And we even have a store where for a few hundred bucks, you can buy your own neuromorphic board and put it in a Raspberry Pi or a PC and run the programs and hook a camera up

• 59:48Speaker 00

  to it and away you go. So the whole idea is to simplify for people to take a look at it and understand the benefits of it.