#738 – Subthreshold Innovation at Ambiq with Scott Hanson

Intro announcer: This is The Amp Hour podcast, released November 10th 2026. Episode 738, Subthreshold Innovation at Ambiq with Scott Hanson

Chris Gammell: Welcome to the Amp Hour. I'm Chris Gammell of Contextual Electronics.

Scott Hanson: I'm Scott Hanson. I'm founder and CTO at Ambiq.

Chris Gammell: Ambiq is a very interesting company and doing a lot of low power microcontroller type things. And you started it all in your PhD. Is that right? Yeah. Your research kind of led into the company, I believe.

Scott Hanson: Yeah. Yeah. Yeah. So I was at the University of Michigan back in, let's say, the mid 2000s. So we were studying what everybody at the time was calling smart dust. And the idea was, let's build cubic millimeter computers. Let's put battery and sensor and microprocessor and radio all in this tiny little package and go use it in health care use cases and defense use cases. And as you can imagine, when you're building something that's a millimeter on a side, you have these incredible energy constraints. And so we were thinking in nanowatts and picajoules. And that really was kind of the basis for Ambiq. What eventually became Ambiq.

Chris Gammell: Awesome. I mean, I remember, I think I knew the dust term because didn't like linear tech bought someone named something with dust in the name. And I think Google board also did something with dust. It was like a thing. It was like a terminology.

Scott Hanson: Yeah. Yeah. So there was a company. I want to say it was called Dust Networks, spun out of Berkeley. And actually Berkeley was the researchers from Berkeley were the ones that started this whole smart dust concept. And then everyone else kind of jumped on top of it. But actually the core technology for what we did, in fact, went back several decades before that. So what we do is something called subthreshold and near threshold circuit design. So we're running transistors at much lower voltages than your circuits one-on-one textbook says you can run them at. And in fact, the wristwatch developers of the 1970s and early 1980s were also exploring the same technology. But it went quiet for a number of years. And then in the early 2000s, there was this resurgence of interest in that technology. And in particular, there was this smart dust concept that started floating around. And I would say that was probably the precursor to what eventually became called Internet of Things and what today we're calling edge AI. So if you want to trace the lineage of edge AI back, you can go all the way to those smart dust days back in the early 2000s.

Chris Gammell: And so when you were doing this research, were you actually fabbing at the University of Michigan, fabbing out at these subthreshold circuits?

Scott Hanson: So we were actually building at the big foundry. So at the time, that included IBM. IBM was still serving as a foundry, included TSMC. So they would have these shuttle runs and you go and quickly build chips. And now that I build chips here at Ambiq and we have to build commercial-grade chips, I can't believe the speed with which university researchers can build chips. Now, you know, the goal is to get one to work and make it work at room temperature, not necessarily make hundreds of millions of chips work. But yeah, we moved quickly. We worked around the clock. We had a bunch of grad students. Those were fun days for sure. But we were building real chips. We did have a fab at the University of Michigan. They've got a clean room. They can build circuits. But it's, you know, it's a few generations back from what you can build. At least at the time, it was a few generations back. And I expect that they're even further back now. But a lot of the work there was really focused on MEMS, manufacturing MEMS. So Michigan is and was a pioneer in the MEMS space. And so they participated. They would, you know, build and fabricate sensors there that would be part of smart dust type solutions or some of the cubic millimeter computers that we did or that we were building. But we were ultimately manufacturing at the big foundries.

Chris Gammell: Yeah, yeah. I went to Case Western. They had like a MEMS facility as well. But it was like, yeah, it was some big geometries. And it's interesting with, you know, the kind of using leading edge at the time, things like that. So as we've gone to smaller and smaller geometries as well, it feels like we're already dropping to lower and lower threshold voltages. So then how does that interplay now with like, is kind of everything not subthreshold, but like the thresholds moved down. So do you still have to move below that?

Scott Hanson: Yeah. So the, you know, we, as part of scaling, voltages drop, threshold voltages tend to drop. Both of those have saturated as we've gotten to the more leading edge. FinFET transistor gave us a kind of a step function improvement. So these FinFETs are fantastic transistors. One of the benefits of the FinFET is that you have a near ideal subthreshold slope. So as you're, you know, for every millivolt that you raise the gate voltage, you get a substantial increase in current drive capability. And that really helps in driving, driving your circuits at low voltage. And so the FinFET gave us this one, you know, step function advantage. So VTs were able to drop again and VDDs were able to drop again, but we're pretty saturated at this point. What that does mean for us, though, is that, you know, we have, you know, what used to be a near threshold or subthreshold voltage in, say, 0.18 micron technology is no longer a subthreshold or near threshold voltage at some of these advanced nodes. That doesn't mean we stop there. We're, you know, running at much lower voltages, of course, than even we were before. So, you know, it might be that an older node, 0.7 or 0.8 volts is a near threshold or subthreshold operating point. But in these advanced nodes where, you know, VTs can be anywhere from like 0.1 volts or 0.2 volts up to 0.5 volts, the definition of what is a subthreshold or near threshold certainly changes substantially. You know, the good news is we've spent at least at Ambiq and prior to that at the University of Michigan, we spent a lot of years figuring out how to run at extremely low voltages. Actually, our, a lot of the research work that we did was running at voltages as low as say, 0.1 volts, 0.1 volts or, or, you know, we even did some, some work to figure out, hey, how low can you go? And we saw that you could run down to, you know, let's call it 65 millivolts. You know, in practice, you would never want to be down there, but it was fascinating to see just how low we could actually get circuits to run under ideal conditions.

Chris Gammell: Yeah. That's wild. That is really wild. So let's talk about the jump. I mean, so you made the jump from research, you know, tech transfer effectively to full company now on the public markets as well. I mean, that's, that's a wild journey and you're still at the helm in the technical side. So how did that all go? I mean, it must, it must be a lot of stuff, you know, like a lot of stuff.

Scott Hanson: It was an interesting process for sure. So, you know, I, I, um, as a PhD student, there was some point where I realized that the technology we were developing, uh, was going to get out there, whether, whether it was me, Scott Hanson driving that, or it was one of the other universities driving it, or it was, you know, some big corporate research lab. I had this feeling that the technology had a, it was solving a real problem. We had solved a lot of the core, you know, fundamental challenges and it needed to get out there. And so, um, I decided at some point to start a company and I, I, I do remember, I remember riding in a elevator in the computer science building at the university, um, thinking about this and going, okay, if this is going to happen, I'd like to be the one to make it happen. Right. And so, um, started, decided to start the company. Um, I would say that my, my thesis advisors, uh, who have been extremely supportive throughout this all, this whole process, I think they were probably tentative at first, tentatively supportive, right? Because, you know, Hey, we've got a student with some good research, good publications. Let's place him at a, at a, you know, a well-regarded university as a professor or at one of the, you know, well-regarded research, corporate research labs. Um, so this was different. It was different than anybody else in my lab, uh, different from anything anybody else in my lab had done. And so, um, the, the first step was actually that we got, uh, we got some, uh, state and federal research money that supported the commercialization of the technology. Right. So I actually stayed at the university for a one year postdoc plus or minus to continue to develop the technology. Cause you know, having a technology is one thing, but knowing what product to build, what markets to chase, what, what does the company look like? Um, it's a whole other thing. Um, and you know, there were times where I thought maybe we're building a medical device company, right? Maybe the core technology is a low power chip company, but we're going to build medical devices. Uh, and, and, and then the other extreme was we sell IP or we, we, we build chips and ultimately, ultimately we decided to build chips, uh, rather than devices, rather than IP. That was really the best way to get the technology out there in the broadest possible way. Um, but you know, from there it was a, it was a long journey. And in 2010, as I was starting this company, uh, I think there was, there was a, maybe some discontent with semiconductor investing, right? People had, had, uh, done really well in, at certain times they had done very poorly at certain times. And 2010 was kind of a low point where if you look historically at venture funding for chips, it was at a low. And today we're, we're looking at it at historical highs, right? I mean, everybody wants to get in on the AI explosion. There's, um, there's so many AI chip companies out there trying to make a name for themselves in 2010 though. That was not the case. I think a lot of people had, had lost money on wifi companies. I, someone told me that at some point, I don't know what the number was, but let's say it was, there were 50 wifi companies at one point. And it's, you know, we understand now the world cannot sustain 50 unique wifi companies. That doesn't make sense. And so I think there was, I, there were a lot of people telling me, Hey Scott, don't, don't start a company right now. The world doesn't need another one of those. And, you know, at the time what I was pitching them on was, Hey, this is not like the chip companies that need to raise a ton of money to build huge digital processors. We're not building, um, you know, giant networking chips or, or, or, uh, giant, uh, MPUs. We're building lightweight processors that are going to lag behind Moore's law. And so it's going to be much less expensive. What I kind of, what I, what I was wrong about was we do need to play the Moore's law game. We did need to jump ahead to, to build bigger and better processors. And we can talk more about that later, but, um, I was wrong about that. Uh, but I was fortunate in that the investors jumped on board. They believed in me and were able to go, go build this company.

Chris Gammell: 20, 10 is, that's still in the shadow of a financial crisis and all that other stuff too. So like, that is just, everybody's kind of titchy about any, any kind of money stuff. So I think probably being, I mean, approaching them and being like, Hey, I, I live on ramen. I'm a student still that sort of thing. And then being able to say that and still, you know, working through all the, the long hours of simulations and similar probably really helped. Yeah.

Scott Hanson: And it was, look, it was a really fun time. I'm in, in hindsight, it's, it's sometimes I'm surprised that I did it right. Because yeah, because we're sitting in the shadows of, of, uh, uh, market, uh, meltdown. Um, uh, all the people around me, all my friends were going and making money, good money. You can make great money as a, as a PhD grad out of the university of Michigan, right. And go work at a research lab, go be a professor or whatever. Right. So I, I had my moments where I wondered if I was doing the right thing, but, um, it, it worked out and, um, I chose Austin as the place to start Ambiq in part because money was, was hard to come by. Right. It was, it was going to be easier to raise or to, to, uh, uh, pay the bills here in Austin. Right. Um, salaries are generally lower because of the cost of living difference compared to the Bay area. Uh, and it worked out for us, right. I mean, we, we, there's a fantastic pool of talent here. I think in COVID, everybody realized that. And suddenly you had a bunch of Californians coming, uh, to Austin. And so now every, you know, name a chip company and they've got an office here, uh, building.

Chris Gammell: I mean, we missed each other by a couple of years. I think, I mean, I was down there to 06 to 08 at Samsung and, uh, Oh, no kidding. Cool. Yeah. Yeah. I, in the fab, I, I, I don't, uh, I don't know how to design chips. So that's great though. So, so then let's, let's start from that. So that launch point 2010, you're, you're kind of spooling up from Michigan down to Austin. How did you, I mean, I guess at that level too, how do you get access to PDKs and the tooling and all the things that like, just everything is so expensive in the semiconductor world. Yeah. How do you, how do you actually like make your case for them? Is it like a co-investing kind of perspective?

Scott Hanson: So we started at the university, right? And this is what I tell entrepreneurs all the time that come to me with an idea. I say, Hey, don't be in a rush to leave the university. Cause still the university of Michigan or probably UT or wherever their access to CAD tools dwarfs even my company's access, right? That you, we never had any thought about how many licenses can I pull down? How many seats can I pull down of XYZ tool? We had infinite licenses basically. And so, um, if you want to run place around, you got place around licenses. You want to run synthesis, you got those licenses. We, we even had a special room at Michigan, uh, arm sponsored research there. And we had a special room, a secure room where you could go and you could access, I think pretty much any arm IP that you wanted access to. That's why arm change your business model where today you can, you can get some taste of that. Right. But, um, it was an incredible thing at the university to gain access to that. So we're, we were able to early on do, um, exploratory development, not real product development, but exploratory development, demo development, et cetera. And then when I spun out though, you're right. That was, I was in for a shock there because suddenly I got to pay what for what? Yeah, exactly. You got to see how much these things cost. Now at the time, um, Cadence gave us a great deal, basically a startup deal, um, which is, was crazy low at the time. Um, of course, you know, terms change as soon as you, you raise a reasonable amount of money. Um, they're like, they come knocking and they're like, Hey, we saw that announcement. Yeah, yeah, exactly. And look, there's, there's three companies that make the, with all these startups, there's, there's three groups that make money. There's the arms of the world that are making money because, you know, they're, they're sucking, soaking up a lot of license fees and the, the CAD tool vendors are making a lot of money. And then the attorneys are making a lot of money. And so in the early days of a startup, it's those three groups making money after that, eventually, you know, the, the successful few companies like ours, you know, emerge and, and, and, you know, break away from that. But yeah, it's hard. It's hard in the early days to do that. And part of it is, yeah, you have to convince these companies why they should pay attention to you, right? Because they're getting inundated with, with students, you know, kids like me at the time trying to, um, raise money for a company and they're trying to figure out where do I spend my time? So yeah, I had to pitch to them. Um, TSMC was maybe a harder nut to crack, right? Um, look, TSMC is a fantastic fab. They, they have really built really high quality, uh, product. And when you're doing something like we do, which is to, to push the boundaries of, of transistor operation, when you're running the transistor in a region that is, is not really, uh, the correct place to run, you've got to be working with the highest quality transistors. And so, you know, TSMC was at the top of our list and, you know, early on, um, they accommodated us. Um, uh, we, we had some, uh, through our network, we were able to access some people who, um, who were, um, you know, well-placed at TSMC. Um, eventually for us, the real turning point was a few years into the company. We took investment from Kleiner Perkins, uh, one of the big, you know, uh, best, most successful venture capital firms ever. Um, and the guy who actually continues to be our chairman at the company, Wynn Shea, he came on board as an investor and he was really, he had a fantastic network at, at TSMC going all the way up to the top and he said from the start, look, we have to build a close relationship, show them why strategically they should care about what we do. Uh, and, you know, show them that, Hey, we're, we're going to build the world's lowest power microprocessors. You want them built on TSMC fab. Let's show them, let's maybe share information with them, right? There's learnings that we can, we can provide that allow them to build better manufacturing processes. So, um, you know, we spent a lot of time there. There were times where I was flying to Taiwan, uh, every other week, um, just to meet with people there and meet with all the various parties. Um, that's a fantastic company, right? They, they, uh, very high quality, uh, engineering culture, um, uh, quality, somehow they manage quality and performance in a really impressive way. I think there's a lot. Scale too. And scale.

Chris Gammell: Like, oh my God. Yeah.

Scott Hanson: I think there's a lot that, that a lot of companies could learn from the way they, they engineer and build products. So anyway, that was a really smart investment of time that we made early on was go, let's be close with TSMC. Um, and, and look, you have to make sure that in this world, no matter how great your technology is, you still have to be at the bleeding edge of, of the manufacturing. You can't, I can't be building 40 nanometer products while my competitors are building six nanometer products, right? There's only so much I can overcome with, with low voltage and great technology design. Um, you know, we can, we can make up for a couple noted advantage. Like we are able to today successfully, if I had 22 nanometer silicon, I feel confident I could beat, you know, competitors, six nanometer silicon, but there's only so big that gap can grow before you go, okay, you know, we got to catch up. So you need a close relationship with the fab. Uh, you need to invest in that.

Chris Gammell: Can you tell me a little bit about that? Like the, okay. So you, you were flying out to TSMC and you said you're, you're working within their process kit and you're working with their engineers. But how much of it requires actual like custom, custom tooling, custom transistors on your, your behalf that you're designing versus, versus using the standard off the shelf, you know, from the, from the process kit.

Scott Hanson: Yeah. The short answer is that we use a majority off the standard off the shelf, whether transistors or CAD tools or whatever. Um, certainly the transistors themselves, they must be the general purpose transistors that TSMC offers to the rest of the world. This was important from the start, right? If I had, if I had the ability to tweak the transistor structure, that would absolutely help us, right? There are benefits that we gain there.

Chris Gammell: Um, but yeah, then you lose yields and everything else too. Like all this stuff, all of that scale that you're getting, then you lose all that, right?

Scott Hanson: You lose all that. Yeah. And, and yeah, especially today, look at, look at how everybody's fighting for that last wafer of capacity, right? As the, as the data centers are soaking up all available capacity, um, everybody else is left kind of fighting for, for what's left.

Chris Gammell: And if you're in it, if you're in this narrow, some for the rest of us. Yeah.

Scott Hanson: Yeah, exactly. And if you're in this narrow, um, technology node that nobody else is using, there's just not that much available. So that was another good decision that we made is let's stay with, uh, general purpose transistors. And so then it's about, well, let's make sure that the models are as good as they can be. So we worked closely with TSMC to make sure that the, the spice models matched, uh, reality. You know, when you run, we're running at such low voltages that, you know, they might not care that much about the matching of spice to, to silicon, right? Whereas what they really care about is, Hey, when the transistor saturated and you're, um, you know, you're, you've got a typical CMOS gate that's all the way on or all the way off. How good are we getting, or how, how good is the correlation? We had to care about, you know, a voltage region that maybe nobody else did. And so we, we worked together to figure out what, which areas mattered. Um, and then, you know, at the time too, we're, we're using completely standard CAD flows, but there's, there's, um, we have to build our own, uh, uh, libraries and characterize them. And, um, fortunately the, the library characterization world was moving towards, uh, more low voltage tolerant, um, uh, structures. So there's something called LVF, which is a, which is a way we can basically represent, uh, uh, static timing analysis data in a distributed way that, that accounts for, for natural distributions of the Silicon. It turns out that the bleeding edge, the bleeding edge, uh, chip manufacturers were kind of pushing the boundaries of, of, uh, these models and they needed this LVF innovation that benefited us back at the older nodes. And so, you know, yes, at times it looked like we had needs that might be need to be custom, but, but some of the innovations that were happening at the bleeding edge, uh, trickled down to us and we were able to use it. So the CAD flow is completely standard. Of course we wrap a lot of our own scripts around that. We have to do a lot of our own library characterization. We, we have a lot of our own low power oriented flows where the magic really starts to happen is that an architectural level. Um, you know why? Okay. So here's, let me back up. I said, we run at low voltage. We run sub threshold or near threshold. The real problem with that is that you become exponentially sensitive to everything, the temperature, the voltage, to, to, uh, uh, manufacturing variations. And so you have to build circuit structures, circuit architectures that are resilient against that kind of variation. So yeah, let's work to make the transistor models as, as accurate as possible. Let's make sure that the CAD tool flows accommodate us. But the real magic is how do you build resilient architectures that can handle that kind of variation. And so that's actually where we've spent a lot of our time.

Chris Gammell: Can, can we do a little bit of magic? I, you know, I read kid books with my kids all the time. Can we, can we do a little bit of magic school bus and shrink us down and go down to the transistor level? And so like, okay, so now we're only going to be switching like sub threshold, sorry, sub threshold, you know, really small voltage changes have these effects. So what is actually happening down to that transistor level? And then do you have to have like additional, so the architecture stuff is like having kind of test off flows to make sure that it's actually at the level you think it's at, or how does that all work down, down, down, down, all the bottom. Yeah.

Scott Hanson: So we'll, well, let's, let's go down to the transistor level and then we can jump up to the, you know, test, test flows and all that. So, you know, at a transistor level, when we're running sub, let's go all the way sub threshold. So the, the, the gate voltage on the transistor is below the threshold voltage below the turn on voltage. But we do have a drain volt. We've got a drain source voltage across there. You're, you're not all the way on. So you don't, you know, when, when you're all the way on, you have a, you have a channel that has formed in that transistor. You get a bunch of conduction. We're, we're in the so-called weak inversion region, meaning there's, there's just a very small flow of, of electrons. It's a diffusion current. It's not drift current. So it's not the, the normal drift current that we get in, in saturated mode. And so the current is very, very small, but it's still there, right? It's, it's, it's the leakage that everybody talks about, right? We're so worried about the leakage of transistors. You just kind of leave your phone there and it does nothing. And yet there's a ton of power being consumed because of leakage of the transistor. So we're, we're using that current. Now, you know, the, the, the way we normally think about that is that when you're in that weak inversion mode, when you're in that diffusion current mode, the transistor's off, the current is zero. It's of course not zero and it's not, it's non-zero, but in fact it is, it's voltage sensitive. And so the transistor still operates like a transistor, meaning you apply higher voltages and the current increases, right? Because, because of that, we can use it like a transistor in the, in the deep subthreshold region. And so we, on the analog side, we will make extensive use of deep subthreshold operation, right? It transistor still works like a transistor. It's totally usable. We're sensitive, hypersensitive to everything. But, you know, we're building amplifiers with tail currents that might be pico amps, right? It's, or, or nano amps.

Chris Gammell: Just to make a analogy here. So if, if a CMOS transistor is basically like, I have a bunch of water behind a door and I'm holding the door shut, that's all the way off. And you're saying in subthreshold, I let the door open a little bit and a little bit of that water flows underneath. But you're saying is, is the idea that the, because there's variable amounts of water, depending how hard or soft I am pushing that door open or closed, even just to let that trickle out. So you're using that as the on-off state then? Is that, is that the right amount? Yeah.

Scott Hanson: You could, you could argue that that's a, that's a pretty good analogy.

Chris Gammell: Yeah.

Scott Hanson: Yeah. So there's, yeah, there's water leaking out from under that door naturally. You kind of just slightly let a little bit more out and starts coming up more and then more. And then you fully, you fully open it and the water rushes out. We're making use of that water trickling out of the, out of the doorframe. Like you just said.

Chris Gammell: And that's the 0.1 volt difference though. You're saying that that that's enough to, to have multiple, like one on-off state.

Scott Hanson: Yeah. Yeah. And, and, and, and, in, in the academic work, we would regularly run in the between 200 and 300 millivolts of 0.2 and 0.3 volts. We found that was a really good trade off. In fact, there was a lot of work. Okay. So we, we proved that you could work down to 0.1 volt or even lower, but then the question is, should you operate there? Right. And it turns out we did a lot of work in, in, in academia as did some other universities. And what we found is no, you want to back off that somewhat. Right. And I won't get into the specifics of why it's, we'll get into a boring academic argument, but.

Chris Gammell: I mean, there's papers. I'm sure people could go read, right? They can, they can go read your research.

Scott Hanson: Right. Go read, go read some papers. Yeah. But you want to, it turns out that, that for a lot of circuits, somewhere in the 200 to 300 millivolt range is the ideal spot. That's going to give you the best energy efficiency trade off. And that's still subthreshold for most circuits where we operate at Ambiq on the digital side, at least is we tend to be what I would call near threshold, meaning just above the threshold voltage. And you have all the same challenges of subthreshold, but they're, they're lessened, right? It's just that your sensitivities are not quite as great. You still get close to the same energy benefit, but, but a lot of your sensitivities are improved. And there, your analogy that you get, you use still holds up, right? Where, yeah, the, the technically the transistor's on, but there's still a long way to go between something that's near threshold and, and, you know, fully super threshold as we like to call it. So we, we've tended to operate there. It makes things a little bit easier though. You know, with that said, as transistors have gotten better, meaning we've gone from planar transistors to FinFET transistors, it allows us to push limits where suddenly subthreshold might make sense again on the digital side. We're making heavy use of it on the analog side, but on the digital side, it may be time to exploit that a bit. Got it.

Chris Gammell: And so you've, you've said digital analog. So is there even within those classes and are there like, do you pick and choose where these are working within the circuit? Like the, in the, the arm cortex part of the chip versus the measurement part of the chip versus, and I definitely get to the measurement. Cause I know there's some new IP in that space too, but then versus the radio and you guys kind of put all the, put all this stuff on the same, on the same die, right? It's not, it's not multi-die.

Scott Hanson: You know, a lot of the talk in the data center about AI, et cetera, you, you have AI accelerator chips and you have, you have a separate memory chip and you, you have a main, you know, host processor chip. Um, at the edge when we're building devices, whether it's smart dust or it's, or, or some other edge based device, some like the wearables I'm wearing right now. Um, everything has to be on a single chip or most functions have to be on a single chip. So if you were to break open one of our Apollo microprocessors, you'd find a radio, a microcontroller host core, uh, like a cortex M55. Uh, you'd find a GPU for, for managing display. You'd find all kinds of ADCs. You'd find power management to take the, you know, much higher voltage and convert it down to the low voltages that we, we need. Uh, you'd find all kinds of memories, right? You find caches and SRAMs and non-volatile memory. Uh, so these are very complex systems. And yet to your point, you have to make judicious use of these technologies of, of subthreshold and near threshold and other low power technologies, uh, all over. So something like a low power oscillator or a voltage regulator, uh, or a voltage reference generator that these are all analog functions that are probably using subthreshold operation. Something like a microprocessor core that consumes a lot of energy is going to run near threshold because I want, I want low voltage. Um, but it's, I don't necessarily want to go subthreshold because maybe I need performance out of it, right? So one of the downsides of running a low voltages, uh, is that my frequencies drop, right? The speed at which the transistor can switch is, is reduced. So, uh, we, we make a trade off there in digital where you might run at near threshold. And then there might be something like a USB peripheral where I go, Hey, my customer's never going to use that during normal operation. That's something you use when you hook up to a computer and you got basically infinite power available. So I'm not going to worry about the power of that block. I'm going to run it at a high voltage and just kind of be done with it. Right. And, and, um, and that thing makes things easier. So we have, if I was to break open our chip and show you, we have a whole bunch of voltage levels. We have, um, 50 plus five, zero plus power domains, um, which are all unique. So it's really complex. We, we have people who come from the big processor companies and they look at the complexity of this and go, this is wild, right? There's, there's so used to huge processors and having a relatively few power domains. They come here and go, Oh, it's simple. It's, you know, this is just an MCU. It's not just an MCU. When you open up the, the, the, the lid on this thing, right? It's pretty complex. So it's, that's what makes it fun, right? That's why our engineers are out there having a great time right now, designing, designing these, these cool processors.

Chris Gammell: That actually helped clear it up too, because I, I can imagine like, you know, just thinking about my limited microcontroller experience overall, but just like the, the things that I'm doing when I'm trying to conserve power, when I'm trying to just put it into sleepier and sleepier modes, that's when the sub threshold probably takes much more, uh, front seat. Like you said, the analog stuff, controlling the clocks, things like that, like that, that all really starts to make sense for it's not, I'm not trying to crunch numbers. I'm not using the floating point unit. I'm just sleepy, sleepy, sleepy, do something, go back to sleep. Right?

Scott Hanson: Yeah. And like I said earlier, the, the cool thing that's happened though, is that, is that, um, as we scale to more advanced technology nodes, transistors get naturally faster, smaller transistor switches faster. Uh, and so, especially when you get to the FinFET devices, we can actually hit some pretty reasonable, some really high speeds running at, uh, extremely low voltages. So, you know, you're running near threshold, even in high performance modes or in, in, in high speed modes. Um, this becomes especially true for something like, uh, the minute you add an NPU, a neural processor unit, which we've got a product that is launching, uh, soon that has a really big NPU on it. And that particular NPU can do a couple hundred multiply accumulates in a single cycle. Okay. So I can do big, I can run big AI models on that thing. If you can run a couple hundred max in parallel, you don't need to run fast, right? You could run at a hundred megahertz and achieve an enormous throughput, right? And, and handle really big models and a hundred megahertz at a 12 nanometer node. I can run, I can run at extremely low voltages. So, um, Moore's law is our friend. It helps us run at much lower voltages, no matter the mode.

Chris Gammell: Still, still here. Still a thing.

Scott Hanson: Still here. It's still alive. Right. AI is opening some pretty fantastic opportunities for the parallelization of workloads. And that has an impact on how we architect our chips. So that's all really exciting and it's, it's all playing really well with our technology. You know, we, sometimes I wonder if I started Ambiq like five years too early because we, it just would have been so much easier to raise money in the early years if AI was, was known and understood because suddenly it's, it's made, there's this incredible demand for what we do, right? People get, Oh, low power matters. Oh, energy matters even in the data center. Right. Uh, definitely. It's an exciting time to, to be in the chip industry.

Chris Gammell: Yeah. Yeah. Actually, this is a good, good opportunity to, to kind of talk about portfolio and the scale difference, right? We have people, some people that are designing for data center, they're probably listening. Some people that are doing super low powered, Preston company included on that side of things. Like how, how much variety would I see, you know, wearables on the way up to, you'd said, you know, displays, GPUs, NPUs, things like it. So like where, where in the market does, does, does Ambiq sit for this sort?

Scott Hanson: So there's, there's kind of two answers for that. One is, is what are we capable of attacking? And let me start there and say, it's very broad. Our, our, we have a technology platform. We can, we know how to run transistors at extremely low voltages in high volumes. That transistor doesn't care if it's in a little MCU that goes in a wearable or a big processor that goes in a data center. We've started in a set of markets that just cared so deeply about power that, that, that it made sense to, to launch there. So, you know, long-term big picture, we could, this technology is useful in data centers down to cars and robots down to, you know, the, the smallest, most energy efficient wearables. The, the slightly, you know, smaller answer is we started building timing chips and microprocessors that were serving, you know, so-called edge devices. So if you look at the bulk of our volume today, our processors, the, the Apollo processors are being used in a wide variety of wearable products. Um, so think, uh, watches and, and, uh, fitness bands and smart rings, um, smart glasses were being used in all, all these types of products today. Uh, and they were the, they were the customers that were willing to take a risk on a young company that hadn't been proven. And it's because they had some really aggressive things they wanted to do. They were the, the first times I saw AI models being used at the edge was from wearable companies that were doing heart rate monitoring and sleep health monitoring. And this was, it was probably almost 10 years ago now that I first saw that, but they, that they were starting to use AI models.

Chris Gammell: Yeah. They were probably calling them machine learning instead of AI, right? Yeah, exactly. This is like, yeah. Okay. We're, you know, neural nets and we're training it, but it's not like AI. And, but now it is AI. Yeah.

Scott Hanson: Yeah. So we're, we started there and, but we're also being used in industrial settings. So we have companies building, um, machine health monitor devices that, you know, little vibration and sound sensors that could put on industrial electronics. We're in utility meters, um, uh, power grid sensors, um, on the medical side, all kinds of medical use cases. We have customers building ECG, um, patches and, and, uh, blood glucose monitors and, um, EEG monitors. So brainwave monitors. Um, we have companies on the smart home side building remote controls and, and security systems and, and all sorts of interesting things. So, um, what's cool, again, what's cool about the technology even today is that it's broad based. It could be used anywhere. And so, you know, we went, we went public last year and the goal for that was to raise money for us to, to just really go broad. Right? So we had a really big customer base in the wearable space. Uh, at the time of our IPO, we had two customers that were above 10% of our revenue and they were, I think at the time, you know, 30 plus percent of our revenue each. Um, we've added a couple more, you know, customers that are of, of big magnitude. And then there's, you know, there's a couple of hundred other smaller customers in other spaces. But the goal with that IPO is to significantly broaden the company's product offerings, um, and, and serve more markets. And, and, you know, unfortunately for me as the, as the chip guy, um, what seems like the easy part is actually the hard part, the easy, what seems easy is, Hey, we should be able to build a bunch of documentation and software and all this stuff that scales really quickly to attack these other markets. You know, the, in the end of the day, the chips that we build, the Apollo processors that we build are general purpose parts. You can take into any market if you want, but all that stuff is really hard to build, right? It's hard to build easy to understand, um, knowledge based articles and, and software that is scalable and and, and, and all that. So, um, the good news is we've got a great team. They're growing quickly. They're, they're attacking a lot of these markets and working closely with customers. The other bit of good news is that AI is a fantastic productivity tool, right? So, you know, it, what, what used to be hard in terms of code development or, or PowerPoint development or, or document development is getting a whole lot easier with, with the help of AI agents on our side. So I, I think we'll be able to scale up, um, pretty quickly to attack a lot of these other markets.

Chris Gammell: You know, I've talked to a lot of chip companies in my, in my years so far. And, uh, I think, you know, I talked to a lot of them. They're all Silicon designers, hardware designers, whatever. They're like, yeah, we're software companies now too. Cause we have to be, you know, it's just like, like the, I think, I think, uh, there's a lot of longing for the old days when there used to be like standalone, like support software companies that would design against the Silicon. And it's like, no, that we're, we're past those days. Unfortunately, it's like, no, the software's free. Right.

Scott Hanson: When, so that's the, Hey, that's the other problem is yes, we're increasingly a software company more on any given day. There are more people writing code at Ambit than there are building hardware. Right. If you include the AI team, the software team, the field apps engineers. Um, and, and yeah, the challenge is that the, the typical model is that customers don't pay for software. Right. And so, um, you know, we're exploring a couple of different paths in that front. One is can our software be so good that our customers are willing to pay a higher chip price, or can the software be so good that they are willing to pay a license fee and royalties for the, for the software. So, um, you know, the burden is on us to build software. That's good enough to justify that. Right. And so that's what we're, that's what we're working on now. Um, we have, uh, uh, AI team that's building all kinds of interesting software as increasingly the workloads that our customers are building are our AI models.

Chris Gammell: And just to, just to clarify on the AI model. So they're, they're developing like AI models that are going directly on the chip, like on the NPU and stuff like that.

Scott Hanson: Yeah, exactly. Yeah. Running on either the, the, the M Cortex M class core or the NPU. And, um, you know, to do everything from anomaly detection, like imagine a piece of vibrating, um, machinery and you want to know when does that vibration signature change? Right. I don't care what the signature is, but when it changes that suggests something might be broken. Right. So let's go out and fix it. Um, or you have, uh, security sensors that are listening for glass break. Um, or increasingly we have customers that want to do things like add conversational user interfaces, you know, voice user interfaces to devices. So we've gotten so used to talking to chat GPT or Gemini in a natural way that you want to interface with everything. Like I'd like to all these wearables. I'd like to be, uh, talking to them naturally. I don't want to have to remember one key phrase. Maybe I don't want to remember the esoteric set of button clicks that's required to, to, to open a certain menu on the, on the watch. Yeah. A lot of use of AI is happening. And so the AI software team is figuring out how do we make that easier for our customers? Some of our customers are super sophisticated and they, they don't want our help. They want us to give, give, give the, give the hardware in a thin layer of software. Yeah. APIs, please. That's all I'll take is APIs and I'm good. And then other customers want more than that, right? Maybe they even want a whole network delivered, pre-trained, and maybe some ability to customize for their machine. So we're, we're working on solving a lot of that. And, and, you know, it's funny in engaging with customers. What I'm seeing even at the most sophisticated companies is that there is friction in the AI model development flow. So, um, you know, it might be that it's some of the biggest customer companies. One group builds the AI model, builds and trains the AI model. There's a group of data scientists that's really smart and they, they kind of, you know, sit in their ivory tower and do that. Then they toss that train model over to a firmware team. Who's got to get it running on a piece of hardware. And it may be that these teams don't talk. Right. And so, you know, you have to enable both those teams, figure out how to enable both those teams. So again, back to the stuff that seems like it should be easy as a hardware guy. I'm going, look, man, we're solving physics problems down here and you guys can't figure out communications problems, but it is a real software.

Chris Gammell: It's just software, right? It's just software, right? It's just software, right? It's just software, right?

Scott Hanson: I know. I know. Yeah. Well, I'm getting a good education on why that stuff is actually really, really hard. And, you know, I have appreciation that, hey, one bad line of code can completely mess up the energy budget or performance budget of the system. So it's, it's on us as a hardware vendor to provide the right tools, right? So whether it's a simulator, maybe that data science team, they don't have access to our board, right? That's a completely different team. And so we've got to give them simulator capabilities or we have to give them profiling capabilities so that they can profile their model and spit out, okay, you know, on the atomic processor or Ambix Apollo processor, you're going to take X cycles for this whole network. And this layer right here happens to be the bottleneck. We need to give that information to our customers. So, you know, getting back to us being a software company, we're investing a lot in making that stuff easier for our customers.

Chris Gammell: One thing that's interesting about what you said about the models development and kind of the target spaces is it kind of sounded like those were going to be kind of sitting there doing intelligence. You mentioned like the vibration monitoring on like a, in a factory, but you guys are doing mostly Bluetooth. And I saw you also have 802.15.4, like that kind of thread thread and similar type things there. So is that, is that the mode you're expecting that these more industrial, less like consumer wearable type things are working through?

Scott Hanson: A lot of the customers certainly will be using Bluetooth classic, Bluetooth low energy, the various 15.4 standards. There are others, of course, Wi-Fi is another one that, you know, pops up a lot. And so when our customers want to use Wi-Fi, there's another Wi-Fi chip that gets thrown down next to our processor. And cellular, we see, or let's, let's say, wide area communication standards, whether we're talking about cellular or, you know, NB-IoT or NTN now is the big one, right?

Chris Gammell: Yeah, totally.

Scott Hanson: So those, again, tend to be extra chips that get thrown down next to us. And long term, you could imagine that we would want to integrate a lot more of that functionality. Certainly, you know, wireless is interesting because I would say in, in, when the Internet of Things first came along, we saw a lot of companies sprinting towards, let's, let's check all the boxes of the wireless standards. And I would say that for a while there, wireless was a major differentiator in terms of how, what was the performance and power and sensitivity of your wireless. And that's becoming, I think that's becoming maybe less true in some cases. You know, you end up with, with, you need, at a minimum, you need to check the box of, do you have Bluetooth? Do you have 15.4? Do you have Wi-Fi? And so there's this race for everybody to check those boxes. I will say that it's this fight between how much compute do I do on device versus how much do I send over the radio up to the device? Do I send over the radio up to the cloud? And I think there's, what we're going to see emerge is a healthy balance where we're doing some level of compute on device, some level of AI model, AI model execution on device, and then some, some amount of sending that raw data or compressed data up to the cloud for further processing. So the radio and processor kind of go back and forth a bit. And, you know, to your point, it depends on the market, right? Consumer electronics love Wi-Fi and Bluetooth. The industrial electronics might prefer 15.4 or, or one of the long range standards. So it's interesting. More cable.

Chris Gammell: Got to love it. Have a little can, you know, can action. But I think what really, what I'm hearing here is like, you know, you guys are so efficient and the, on the processing side, the, the low power side, that it makes sense to just throw, throw a modem outside of it, power that up and down as you need it. But still the, the corpus of the work is done by this ultra efficient. Yeah.

Scott Hanson: I, I, yes. I mean, I'm going to say that. Yes. That has generally been our pitch to customers is, is we're so efficient. It's worth you buying that separate chip, you know, for sure, relative to a complete SOC. Absolutely. And, and I, a hundred percent that's true. But this trade off again, between how much compute do you do locally versus how much do you do over by sending over the radio? That's kind of a dynamic thing. It changes from customer to customer. And, you know, we, we have radios in house today. We have ideas on what we could do to apply our technology to radios. And so I would, that's one thing I'd really love to see happen is I'd love to see Ambiq figure out how to apply the spot technology, our subthreshold and near threshold technology to radios. Because I think then if you truly have an ultra low power radio and ultra low power compute, then you maybe don't have to make those trade offs. Maybe you can do both the local compute and you can, you can, you know, send the data wirelessly elsewhere for, for processing in the data center.

Chris Gammell: That's great. I mean, one of the things you had mentioned as well, like the, the idea of like best chip in the world. If you leave, if you leave that USB port on, that's like full power, who cares? You know, you're still burning all the wants. So what about, you know, firmware ecosystem side of things? I had seen, I think the first time I looked at Ambiq stuff, I was looking at it from a Zephyr perspective. I don't think I've seen Zephyr support. Yeah. Has that changed or is that a plan?

Scott Hanson: No, we do have, we do have Zephyr support. I don't know, frankly, what's reflected on our website. We do have Zephyr support and it varies from product to product, but yeah, we increasingly are seeing customers request Zephyr and it's, it's becoming, let's say a higher priority for us to make sure that, that it's, you know, it used to be that we, we had alpha customers that would sign up with us early on. They would get first access to Silicon and we would prioritize software, a software development environment that matched their needs. And increasingly those customers are asking for Zephyr. So it'll be, I would say increasingly Zephyr will be part of the early release of, of future Ambiq products. That makes sense.

Chris Gammell: Yeah. That's cool. I mean, what is the, you know, the, the, everything else then? So like I'm, I come from the Zephyr space. I always ask about that, but like, it just kind of bare metal SDK, are you doing free RTOS ports, that sort of thing as well? Like how do you find that most of your cut, that, that the thing that you were offering in those early customers? What, what does that look like?

Scott Hanson: Yeah. I mean, it's a mix. Usually we're, we're offering a hardware abstraction layer. Our SDK includes hardware abstraction layer, board support package. There is RTOS bundled in, but there's also a RTOS abstraction layer. So you can bring your own, a lot of these companies have different RTOSes that they want to bring. Increasingly, there's also a whole bunch of AI development tools, right? So, you know, on top of that, you have to, you have to have kernels and, and, and operator libraries for all the AI models. You have to have AI runtime. So, you know, think about the runtime. The AI runtime is very similar to the RTOS in that it's, you know, it's the thing that's, that's interpreting the, or sometimes interpreting the model, but always executing the model. And so that's also an integral part of, of that whole release. But we're seeing, we're seeing a whole bunch of things. We're seeing some companies really buy into the whole Zephyr experience. We have other companies that mostly want bare metal and they want to build on top of that.

Chris Gammell: Yeah. I mean, I think we talked about this on the show, not you and I, but sorry, on the show, we talked about, you know, early Silicon, a lot of the fabulous startups. Obviously you guys are, are much past that space, but just like you have to do what customers will tell you, right? It's not, you can't sell 10 parts at a time. You have to sell millions of parts at a time to survive. And so it totally makes sense to, to do that sort of thing.

Scott Hanson: When, you know, in the early days, that was exactly what it was for us is, is we had a relatively few customers. You know, I, I always like to say that we, we did two things really well. One was we were the best low power designers, but two was we did whatever our customers needed us to do. And we were willing to, to, you know, bend over backwards for them and do whatever they want. And that's a great policy, but it's not a sustainable long-term methodology. Right. So, you know, for example, it, it caused us not to be necessarily early investors in the Zephyr, in the Zephyr world. Right. We were, we were kind of behind in, in getting into Zephyr in part because our business, Zephyr in part because our biggest customers didn't want it. Right. And then we started hearing about it from them and it started to trickle its way there. Cause you know, these, these huge custom companies, they maybe don't care about that. Right. They, they have their own enormous code base. They don't need the usability of all that. But as we've broadened out and we have many more customers than just those few, these things have started to become important. So I think we're going through, you know, an evolution that every company has to do, which is you, you start with a few customers, you do whatever they want. And then you, and then you figure out how to serve hundreds or thousands or tens of thousands of customers. And we're in that, that, you know, latter part of the curve right now.

Chris Gammell: Right. What you do is you take a couple of product managers, you lock them in a cubicle and you just see who wins, you know?

Scott Hanson: Yeah. I like that. I like that. I like that. Yeah. Yeah. Yeah. Yeah.

Chris Gammell: On the AI side of things. So compression kit. So first off, let's just define the term. I know that's like a new wish kind of thing, but like, how does it work? How, how would I use it? How would listeners use it for making better?

Scott Hanson: Okay. So, so compression kit is it's basically a kit that allows you to build a neural codec. And what I mean by that is that it will do a compression or decompression using an AI model, using a neural network. And, you know, usually we're, we're using, we're to do compression. We're usually using traditional algorithms, right? MP3 has an algorithm that it uses to, to compress audio data. And it's an algorithm that is tuned to, to compressing audio. We can do the same thing with neural networks, right? In fact, a neural network could emulate MP3 if we wanted it to, that wouldn't necessarily be the right thing to do, but we could do that. And what's cool about the neural networks is I can quickly swap to different types of data, right? I can, I can compress heart rate data. I can compress IMU data. I can compress audio data. And each of those would be a slightly different AI model that, that knows the structure of that data, can compress it efficiently, and then can reconstruct it later. And so, you know, it turns out that these, these AI models can be very, let's say adaptive to the type of data that you're using. So we can quickly adapt to new data types. You don't necessarily need an expert who knows everything there is to know about sound and music and everything else. In order to build this, you can, you can build, as long as you have good training data, you can build a reasonably good neural codec. So what this is, is a software development kit that allows you to build your own neural codec. You can select which of, which of the various models you want to use. And then, you know, you can have a lightweight codec running on your, on your machine. And so we're seeing customers use this for a variety of reasons. One would be going back to saving radio power, right? So, you know, your typical wearable, for example, is collecting a ton of data about you every second, right? You've got a heart rate monitor that's collecting multiple channels of data constantly. You have a inertial measurement unit that's, that's constantly measuring your, your movement. You have maybe a microphone that's collecting data, actually a very high sampling rate. And you have a variety of other sensors. And so if you were to just take all that raw data and ship it out over the, over the radio, you're going to, you know, kill your battery very quickly. With compression kit, I can run a neural codec to compress down that data, maybe 10X, maybe 16X, maybe more. So that I'm sending a much smaller amount of data over the radio. So at a minimum, a lot of our customers are, are looking at doing that. But, you know, going back to data center, guys soaking up all the available capacity out there. That's true on the memory side as well, too, right? So memory is hard to come by. And if we can find a way to take our customers data and shrink it down by a factor of 10, that's really valuable. They can have less memory in the system. That's a cost benefit. That's a size benefit. It's also still an energy benefit, right? Because it takes energy to retain data and memory. So compression kit is really cool. We're expanding it. We have a good AI team, including, including a guy whose PhD thesis was on neural codecs, and he's out there, you know, cranking out a bunch of great work every day. So that's, that's one product that we're building. We've got a bunch of other AI products like that. But the hope would be that, you know, Ambiq eventually will have a big library full of stuff like compression kit that our customers can use to ease their development.

Chris Gammell: Just to, again, just to plan, make sure I'm not confusing the name versus the actual function. So like you mentioned, like, so just like a glass break style, like the detector, right? You have a microphone on board. You want to know if glass is broken. That's like a common one that I've seen. Is the actual like data that's getting transmitted or stored or whatever, like break has occurred or, you know, one or zero, or is it actually like recoverable? Here's an audio file. Here's a smaller audio file. Like what's the difference?

Scott Hanson: So it's really the latter. It's the latter. And the way a system like that would work is you would have a classifier locally that runs on our chip that detects the glass break. And at the same time, though, you'd be, you'd be accumulating a buffer full of all the raw audio data. And then when you detect that, when you detect that glass break sound using one AI model, a compression kit model would then compress up that raw data and send it over the radio link. It could then be reconstructed on the phone side or the data center side so that you could later listen to that sound and go, yeah, maybe, maybe, for example, you have a second classifier that goes, hey, the edge device thinks there was glass broken. I'm going to confirm with a much bigger AI model that that's true. And so it's helpful to have that compression capability.

Chris Gammell: Got it. So just to play on that example a little bit more. So, so like PCM coming out of a microphone, I'm so behind on microphone technology. So just let's let it, let it go, please. Uh, so PCM coming out of there, something like raw data that's going through this classifier on board. And then the idea is that's so low power that when you detect something that might be, that's when you decide to do a energy expensive, like conversion to MP3 and an energy expensive, like transmit over Bluetooth.

Scott Hanson: Yeah. Yeah. You might do something exactly like that. And, you know, in a sense that classifier that detects the glass break is a compression methodology, right? You're, you're compressing raw data down to one bit that says glass breaker. Exactly. Exactly.

Chris Gammell: But that's not recoverable then. That's just, that's just like, that's a, that's a trust me, bro. Right. That's a, yeah, exactly.

Scott Hanson: And Hey, where this, where this also matters too, is, um, you know, increasingly we're seeing companies talk about like life logging where, uh, whether it's a glasses or it's a badge you wear, it's a, it's your watch. It's constantly listening to all the sounds around you. Um, look at the, the Apple watch, uh, recently just launched in there. Um, they're listening for certain sounds, whether sirens or something else, and they can kind of log an event. And so, um, that's useful. Um, that's useful. It's useful both to have that classifier locally to go, Hey, there was a car crash sound or, um, or, Hey, I heard, uh, you had a 15 minute discussion about, um, the approaching tape out deadline. Um, you can use a classifier. You can use a classifier to identify that and build kind of a log of your day. And then you can use compression to, to, to compress that raw data down, whether for local storage or transmission later, we're finding that customers want both. Hmm.

Chris Gammell: Yeah. Uh, I'm not sure if I want that personally, I guess I'm not, I'm not a customer, right? I'm not, I'm not a customer, right? I'm a user of your customers probably, but yeah, yeah, yeah. Yeah.

Scott Hanson: That's a valuable, it's a valuable point. Right. And, and, um, yeah, the first iteration of smart glasses 10 years ago was, no, I had, I was a glass hole.

Chris Gammell: Yes.

Scott Hanson: Yeah. And the newest versions, um, I think are, are substantially better though. We're still facing some, some privacy concerns there. Yeah. Um, and so, yeah, we got to fight all this out, but, but let me make a diversion here and say that that's, that's one of the benefits of edge AI. If I can, and this is Apple's argument, right? Is, Hey, I'm going to do all the processing locally so that your private data is not compromised and it's going to be your data. You own it. Um, we're not sending it anywhere else. And so that's the argument for edge, uh, processing and we're fully on board with that. Um, but yeah, I think that the, the whole, our customers and the whole industry has some ethical and, and other issues that have to be sorted out in the next couple of years.

Chris Gammell: Right. I think, uh, it was said best by, uh, Jeff Goldblum, uh, just, just cause scientists say you could, doesn't mean you should, right? Yeah. Yeah. Fair enough. Fair enough. Well, this is, I mean, this is a very interesting stuff. I mean, where, where do you usually, so if, if, uh, you know, I, I told you before the show, like my, my vision of, uh, you know, uh, uh, amp hour listener is like someone two years, three years out of college, you know, new professional. With all apologies to all the people that aren't that, but where do you tell that person to like, go get started with Ambit products to like, you know, really get booted up and, and, and building things quickly. Where, where's you, where's you, where do you point them?

Scott Hanson: First of all, say it's never been easier, whether it's our chip or it's any other company's chip, it's never been easier to start development. Um, whether you're building an AI model or you're, you're, you know, building a simple sensor device for sure. For sure. Go to our website. We sell a bunch of evaluation kits that have, uh, our Apollo processors on board, um, that have all kinds of sensors on board, whether it's a display and you want to run, you know, something that's driving a rich cell phone like display or smartwatch like display, or you want to do radio communications. You want to add microphones. We've got all that data. We've got, uh, reference designs that, that, uh, can be used. So, uh, go to our website and, and, and start there a bunch of, um, great kits that can get you started. And then there's a bunch of modules that you can buy for these boards that you can throw down. So any sensor imaginable, um, uh, you can, you can buy, and we, we have links to that on, on the website. Um, and these days with, with, uh, agentic coding, it's never been easier to start developing code, right? You don't, you know, it used to be that writing C code for an embedded processor was a very specialized talent. And to some extent it still is, but the AI agents are getting a lot better at understanding memory constraints and processing conditions. And so forth. So, um, I think it's a really fun time to get involved in, in coding. I mean, I've got a 14 year old son who's, who's learning coding at school right now. And I'm just itching to get him started on an Apollo processor. Cause he could do a lot, right? There's an amazing amount that we have. He can do. We have interns that come in and are, it's amazing what they can produce. So my advice is go play around with AI coding agents, um, uh, get them to help you going on, on Apollo as quickly as possible. We'll supply the hardware.

Chris Gammell: And, and just, I mean, uh, you mentioned you're a parent to the child, but maybe be a proud parent of, uh, the Apollo line and the, uh, the latest. So like the latest part, could you give us like a little like pitch on the, was it the 510 is the latest? Is that right?

Scott Hanson: Yeah. We've got a couple of different, um, Apollo five variants. We have the Apollo 510, which is our flagship product that has a cortex M 55 processor. It's got, uh, voice seven or eight megabytes of, of total memory on board between non-volatile memory and SRAM, uh, has a, a GPU. It can run up to 250 megahertz, um, a bunch of cool analog peripherals, uh, and at extreme low power. So, um, it's the M 55 is at the higher end of arms embedded, uh, processor lineup, and it has a, a vector, um, accelerator inside. So you can run as many as eight multiply accumulate per, per cycle. You got a ton of memory there. So it's actually great for, for running AI models. Um, and so that one is out there. It's shipping in, in, in a bunch of volume. You could open up a lot of wearables today and find it inside and, and driving big displays and, and running AI models. Um, so that's a cool one. We have a lighter weight version of that called the Apollo five light. And we have another version called Apollo three 30. These are, think of those as, as, uh, smaller versions that fit into smaller footprints. Um, maybe a slightly smaller price point. Um, and then we have a, the, the thing I'm really excited about is we have a new product coming that we've, we've announced called atomic. And that, that is our first 12 nanometer product. It runs, uh, at much higher speeds. I can't recall if we've, we've actually published the, the speed that it runs at, but let, let's say it's a, it's an integer multiple of what, uh, Apollo was capable of. And it includes a neural processing unit and a ton of memory, far more memory than we've ever had on a chip. So if you're thinking about building enormous AI models that just won't fit in a few megabytes, we've got you covered with this next gen atomic. Um, I'm really excited with what our customers are going to be able to do with that. It's, uh, you know, I can see the future of, of what's going to be possible in terms of the complexity of AI models that can run on tiny little battery powered devices. And I'm really excited and the roadmaps that our customers have, have put out there for products like atomic are really exciting to me. They, uh, I, I'm not going to disclose it here, but the, the ambition that these companies have is, is pretty staggering. And if only a few years ago we were, you know, we, we were really trying to pull use cases from, Hey, I know you want to use AI, but what are you going to use it for? And they couldn't articulate it. Now they're driving us, right? They've got these crazy use cases that, that, uh, are, are pretty exciting. So, um, I would say buckle up in the next few years. If you're a developer, if you're developing in the embedded world, figure out how to use AI, both to help your accelerate your coding experiences, but also use AI to enable features in your products. Right. These chips are, are so capable of running complex AI models today. It's, it's a very exciting time.

Chris Gammell: Yeah. That's really great. Do you guys do a lot of trade shows? Like people might stop by and see you at a trade show?

Scott Hanson: Yeah. Yeah. And we've got a bunch of smaller trade shows. We're doing the next big one is going to be CES. So if you're at, if you're at C, if you're walking around CES, we'd love to see you. We're going to be there. Look us up.

Chris Gammell: Okay, great, great. Oh, and I'll say, I guess you mentioned, uh, are you guys hiring as well?

Scott Hanson: We always pitch that too in case we are hiring. Uh, we cannot hire fast enough. Okay. So we have openings in our AI team, in our, across our chip design team, across our, our firmware and software development teams, operations, um, production tests, marketing sales. We're hiring everywhere. So please reach out and, um, uh, yeah, it's a, it's a good, good time to be at edge for sure.

Chris Gammell: I was going to say I'll fully throw a full throat endorsement of Austin, Austin, Texas as well. So, uh, yeah. And come live in Austin. It's beautiful. Beautiful. Yeah. Just watch out for I-35. It's a, it is a treacherous place, but I-35 is a terrible highway. It's my least favorite highway in the entire United States to drive on. Yes. Everywhere else is great. Well, I-5 is also, you know, runner up, you know, you go out to California. Yeah. Okay. I don't need to drive on I-5, but yeah. Yeah. All these Europeans are like, what are you talking about? All right. Well, Scott, thank you, thank you for being here. It's been really great to, you know, boy, all the way up and down the stack. Uh, really cool and, uh, excited to see these new parts and, uh, hopefully see it at industry events soon.

Scott Hanson: Perfect. Thanks for having me. Enjoy the conversation.

Chris Gammell: Bye.