Latest / WTR Small-Cap Spotlight / Ambiq Micro (AMBQ): Intelligence Everywhere: Ultra-Low-Power Semiconductors for the Edge AI Era

Transcript

Ambiq Micro (AMBQ): Intelligence Everywhere: Ultra-Low-Power Semiconductors for the Edge AI Era · WTR Small-Cap Spotlight

• 0:01
  SPEAKER_01: Welcome to the WTR Small Cap Spotlight Podcast.
• 0:04
  SPEAKER_01: I'm your host, Tim Gerdeman, Vice Chair and Co-Founder and Chief Marketing Officer of Water Tower Research.
• 0:14
  SPEAKER_01: Today we're joined by Scott Hansen, founder and chief technology officer of Ambic Micro, NYSE ticker symbol AMBQ.
• 0:22
  SPEAKER_01: Ambic develops ultra-low power semiconductors for Edge AI.
• 0:27
  SPEAKER_01: Its proprietary sub-threshold power optimized technology, or SPOT, delivers a two to five-fold reduction in power consumption versus conventional designs depending on the application.
• 0:40
  SPEAKER_01: That helps bring on-device AI to battery-powered products such as smart watches, medical monitors, smart glasses, and industrial sensors.
• 0:48
  SPEAKER_01: IBIC has chipped over 300 million units worldwide, which is quite an impressive number, I must say.
• 1:04
  SPEAKER_01: Scott founded the company in 2010 and has served as CTO since 2013.
• 1:09
  SPEAKER_01: Also joining is my WTR equity research colleague, James Kisner.
• 1:13
  SPEAKER_01: Gentlemen, thanks for joining me this morning.
• 1:15
  SPEAKER_01: Thank you.
• 1:15
  SPEAKER_01: Happy to be here.
• 1:18
  SPEAKER_01: Scott, for listeners new to the name, could you give us a 30,000-foot overview of the company's current business model and perhaps also give us some color on where the company is strategically headed?
• 1:28
  SPEAKER_00: Yeah, so Ambic is uh a fabulous chip company.
• 1:32
  SPEAKER_00: We build the world's most energy efficient chips.
• 1:35
  SPEAKER_00: Uh anywhere you want to add intelligence where there wasn't intelligence before.
• 1:39
  SPEAKER_00: So uh we as you mentioned earlier, we've got this proprietary uh technology platform that we call Spot.
• 1:44
  SPEAKER_00: It allows us to build chips that are two to five times less uh uh or consume two to five times less power than uh traditional approaches.
• 1:51
  SPEAKER_00: And there's four primary markets that we're targeting.
• 1:54
  SPEAKER_00: We're we're targeting what we call personal devices.
• 1:55
  SPEAKER_00: That's any any of the stuff I'm wearing here, smart watches and rings and glasses and everything like that, medical devices, uh, industrial devices, and then smart home and smart building devices.
• 2:05
  SPEAKER_00: And we're present in in all those markets today.
• 2:08
  SPEAKER_00: Um we talked, uh, we we have a flagship family that we call Apollo.
• 2:13
  SPEAKER_00: This is a microprocessor family that has spanned five different generations and will continue to expand.
• 2:18
  SPEAKER_00: We have a new in-development product family called Atomic.
• 2:21
  SPEAKER_00: This is a higher-end system on chip that uh includes a neural processing unit for the acceleration of AI workloads.
• 2:28
  SPEAKER_00: Uh and uh, as you said, we've shipped more than 300 million units today to a broad variety of customers.
• 2:33
  SPEAKER_00: So a lot of excitement at Ambic, and I'm very excited about this this whole AI every everywhere narrative that's uh I think finally uh catching up in the in the public.
• 2:44
  SPEAKER_01: You and me both.
• 2:45
  SPEAKER_01: I'm leaning more into it by the day, and I know James is quite the AI guru, so uh I'll turn it over to James to ask a few questions.
• 2:51
  SPEAKER_02: Thanks.
• 2:52
  SPEAKER_02: Uh thanks, Tim, and welcome again, Scott.
• 2:54
  SPEAKER_02: So um you uh you uh got your PhD and I guess undergrad too at uh University of Michigan.
• 3:00
  SPEAKER_02: I'm uh Big Ten also, Indiana.
• 3:02
  SPEAKER_02: But um, you know, you worked on tiny and pentable systems with severe power constraints, and then you started MBIC with your thesis advisors in 2010.
• 3:11
  SPEAKER_02: So, you know, what problem were you trying to solve and how is that vision held up?
• 3:15
  SPEAKER_00: So the the the vision at the time when when we were doing research was uh smart dust.
• 3:20
  SPEAKER_00: Okay, everyone was trying to build tiny little computers that would fit in uh in a dust mode or be uh a millimeter on a side or something like that.
• 3:29
  SPEAKER_00: Uh and it was for all kinds of use cases.
• 3:31
  SPEAKER_00: So think um think medical use cases, think defense use cases.
• 3:35
  SPEAKER_00: And so we were building cubic millimeter computers with a sensor, with a radio, a microprocessor, uh, and and they were we were uh using them for intraocular implants, so eye implants for glaucoma patients.
• 3:48
  SPEAKER_00: And um, you know, when you when you got a cubic millimeter computer, you got a tiny battery, you can't consume much energy, right?
• 3:54
  SPEAKER_00: And so we were really looking at the limits of how low could you take energy.
• 3:59
  SPEAKER_00: Um and the way we chose to do that was using something called subthreshold computing.
• 4:03
  SPEAKER_00: And this is a it's a form of computing that's existed for decades.
• 4:07
  SPEAKER_00: Okay, so it goes back to the late 70s with the uh wristwatch uh developers back then, but it never really widely took off and it had all kinds of problems.
• 4:16
  SPEAKER_00: Basic idea is that we run the underlying transistors at much lower voltages than would otherwise be possible, and you save a lot of energy.
• 4:23
  SPEAKER_00: So we explored that.
• 4:24
  SPEAKER_00: Uh, we got to the point where uh where it was commercially viable, and then in 2010 I spun it out of the university and have been uh driving that ever since.
• 4:33
  SPEAKER_02: Let's let's dive into that a little more um you know about spot and kind of this sudden near threshold operation that saves power.
• 4:40
  SPEAKER_02: Like what makes that approach so difficult to commercialize at scale?
• 4:44
  SPEAKER_00: Yeah.
• 4:45
  SPEAKER_00: Yeah, that's the right question because it sounds simple, right?
• 4:48
  SPEAKER_00: You just dial down the voltage and magic happens.
• 4:50
  SPEAKER_00: Um well, you become exponentially sensitive to a whole bunch of quantities.
• 4:54
  SPEAKER_00: You as the temperature changes by a few degrees, you're exponentially sensitive to that.
• 4:58
  SPEAKER_00: Uh as the voltage that is coming in uh uh from a battery or a regulator gets applied, you're exponentially sensitive to changes in that.
• 5:06
  SPEAKER_00: Uh even little manufacturing changes uh have exponential changes.
• 5:10
  SPEAKER_00: So, you know, we're we're building things at an atomic scale at this point with these chips.
• 5:13
  SPEAKER_00: And so just these tiny changes lead to enormous performance and power and functional changes in these chips.
• 5:20
  SPEAKER_00: Um, you know, imagine that you're trying to build a chip that runs at a hundred megahertz uh clock rate.
• 5:24
  SPEAKER_00: Well, it might run at one megahertz in some in some cases, it might run at a gigahertz in others, right?
• 5:29
  SPEAKER_00: So it's really difficult to build a chip under those kinds of conditions.
• 5:32
  SPEAKER_00: And it's that problem that we've solved at Ambix.
• 5:34
  SPEAKER_00: So we have uh built a whole circuit platform, what we call spot, that allows us to fight those sources of variation.
• 5:40
  SPEAKER_00: And it's a it's a what's cool about the platform is it's diverse.
• 5:43
  SPEAKER_00: It's dozens of techniques, uh, some of them digital, some of them analog, some of them memory, um, some of them just related to how we production test and trim all these transistors.
• 5:54
  SPEAKER_00: But the net result is we can run with two to five times less energy as compared to the incumbents.
• 6:00
  SPEAKER_00: And what's also cool about it, in addition to being kind of a broad-based uh uh solution, is it's applicable to all kinds of chip types.
• 6:08
  SPEAKER_00: So, you know, we're building today the Apollo and Atomic Series families, but the truth is the transistor doesn't care what type of market or what type of product it's used in.
• 6:17
  SPEAKER_00: We we just know how to operate it more efficiently.
• 6:18
  SPEAKER_00: And so I can take it to uh automotive, I can take it to data center, and I could take it to all kinds of interesting use cases that uh that care about power.
• 6:28
  SPEAKER_02: That's helpful.
• 6:29
  SPEAKER_02: So um, you know, just maybe stepping back a little bit, like why does running AI at the edge matter?
• 6:35
  SPEAKER_02: Like what really becomes practical when a device can do more inference within a fixed battery budget?
• 6:41
  SPEAKER_00: Yeah, so there's a whole variety of reasons, and it it depends from market to market.
• 6:45
  SPEAKER_00: Um, one of them is latency, right?
• 6:46
  SPEAKER_00: I might really need to do the AI right there at the edge.
• 6:49
  SPEAKER_00: Um a good example would be like a hearing aid, right?
• 6:52
  SPEAKER_00: Um, sound hits my hearing aid.
• 6:54
  SPEAKER_00: I I don't have time to send that sound to the phone or to the cloud for processing before it comes back.
• 6:58
  SPEAKER_00: I need to do the processing right there and in line.
• 7:00
  SPEAKER_00: So latency matters in a lot of use cases.
• 7:02
  SPEAKER_00: Um, sometimes it's a uh it's a privacy or security thing.
• 7:06
  SPEAKER_00: Um there's a lot of applications where I want my data to be my data.
• 7:09
  SPEAKER_00: I'm a health device, right?
• 7:11
  SPEAKER_00: You want that data to be yours right there.
• 7:13
  SPEAKER_00: Maybe you don't you don't want it being owned by the cloud company.
• 7:16
  SPEAKER_00: And so you send only anonymized data up if possible.
• 7:19
  SPEAKER_00: Um sometimes it's a cost thing.
• 7:21
  SPEAKER_00: We had one customer who actually resisted doing AI at the edge because uh of a variety of reasons.
• 7:29
  SPEAKER_00: It was hard, right?
• 7:29
  SPEAKER_00: It's it's hard to do.
• 7:30
  SPEAKER_00: And then once they saw their cloud bill, they saw their AWS bill, they they you know got quickly uh got behind edge AI.
• 7:37
  SPEAKER_00: Um, I mean, I know, I don't know about you guys, but I know for us at least, those AWS bills can get really large really quickly if you're not careful and they're you're paying that toll bill every time.
• 7:45
  SPEAKER_00: So if you can do the AI at the edge and just send a result instead of all the raw data, that's a big deal.
• 7:50
  SPEAKER_00: And then a big piece of it is just energy.
• 7:52
  SPEAKER_00: Um, you know, imagine all these devices that I'm wearing here, they're collecting thousands of data points per second about me.
• 7:58
  SPEAKER_00: And and if you in in aggregate, if you're if you're constantly streaming those over the radio up to the cloud, that's gonna consume a lot of energy.
• 8:04
  SPEAKER_00: So if you can just process it right there locally, that's a big deal.
• 8:07
  SPEAKER_00: And then and then there's availability.
• 8:09
  SPEAKER_00: Sometimes you don't have a radio nearby.
• 8:11
  SPEAKER_00: I mean, uh again, to use my watch example, I go out for a run, I don't take my phone with me.
• 8:16
  SPEAKER_00: This thing still needs to do its job and and serve as my you know, my health and training companion.
• 8:20
  SPEAKER_00: So you need to do the AI right there at the edge.
• 8:23
  SPEAKER_02: Yeah, that makes a ton of sense.
• 8:24
  SPEAKER_02: Um, you know, maybe you can just walk us through the roadmap.
• 8:27
  SPEAKER_02: So there's there's Apollo 340, I believe it's aimed at the mass market edge, while Atomic is designed to add a full uh NPU for more demanding workloads.
• 8:36
  SPEAKER_02: So just how do those products expand Ambix opportunity?
• 8:40
  SPEAKER_00: Yeah, I mean, first of all, the big picture is that I want to be the one-stop shop for for edge compute, right?
• 8:46
  SPEAKER_00: From you know, long-term from a 50 cent processor or a low-end processor to a$50 application processor, I want to cover that full spectrum.
• 8:53
  SPEAKER_00: And so we're building out our product portfolio to do that.
• 8:56
  SPEAKER_00: The Apollo family has all these different generations, Apollo, Apollo 2, 3, 4, 5.
• 9:01
  SPEAKER_00: And what we've been doing over time, think think broadly about with each new family, as the number goes up, the capabilities go up, the memory goes up, the processor speed goes up, the capabilities go up, uh, the price goes up.
• 9:14
  SPEAKER_00: Um, and and so what we're doing with the Apollo 340 is we're actually migrating some of the Apollo 5 features back to the Apollo 3 family, which has been a big runner for us.
• 9:23
  SPEAKER_00: Um, it there's there's some prominent products that have actually only recently designed an Apollo 3, which is probably a nine-year-old uh uh product at this point.
• 9:31
  SPEAKER_00: So, you know, we still our our whole portfolio still has good legs, but we're finding that the companies using Apollo 3, they want AI capabilities that maybe they they they couldn't can't get there.
• 9:41
  SPEAKER_00: So we're migrating some of the latest technology back to Apollo 3.
• 9:44
  SPEAKER_00: It's meant to be a broad-based uh um uh product.
• 9:48
  SPEAKER_00: So it's gonna it's gonna focus on medical and smart buildings and and uh industrial and and wearables and personal devices.
• 9:55
  SPEAKER_00: So it's got a broad uh set of customers that we're targeting at a compelling price point.
• 10:00
  SPEAKER_00: The atomic sits above the atomic product sits above the Apollo family.
• 10:04
  SPEAKER_00: This is our first product that was designed from the ground up for AI.
• 10:08
  SPEAKER_00: So you can run AI on any of our products, but this is one where there's a neural processing unit on board.
• 10:13
  SPEAKER_00: There is uh uh tons of memory uh uh on board as well to support bigger models.
• 10:20
  SPEAKER_00: Uh, and it's by far the most energy efficient and most performant product we've ever built.
• 10:24
  SPEAKER_00: So it's actually built in a 12 nanometer product process where previous products were built at 22 and 40.
• 10:29
  SPEAKER_00: Uh it there's a new version of our spot platform that's much lower power, running at much lower voltages.
• 10:35
  SPEAKER_00: So this is one that gets us access to higher-end applications where you've got big AI models that could be a couple hundred uh million uh parameters.
• 10:43
  SPEAKER_00: Um, it it takes on use cases where maybe there might be a camera.
• 10:47
  SPEAKER_00: So today you cannot uh actually bring camera data on chip, or at least it's it's it's somewhat difficult with Apollo because we don't have the right plumbing to handle that.
• 10:55
  SPEAKER_00: So that opens up a whole new set of markets.
• 10:57
  SPEAKER_00: Think smart glasses that have cameras built in, think all manner of cameras from security cameras to doorbell cams to hunting cameras, et cetera.
• 11:05
  SPEAKER_00: So um, you know, think of that being that atomic family being a really high-end uh product family that will grow much like Apollo.
• 11:12
  SPEAKER_00: Uh and then again, over time, the goal is to continue to grow at either end of those portfolios and to kind of grow them out.
• 11:18
  SPEAKER_00: We're up against all kinds of companies that have just these enormous product portfolios.
• 11:21
  SPEAKER_00: They've got thousands of part numbers, and it takes some time for us to build that up.
• 11:26
  SPEAKER_02: That makes sense.
• 11:26
  SPEAKER_02: It sounds like expansion opportunity and a lot more to come.
• 11:30
  SPEAKER_02: Um so um, you know, I guess uh, you know, like some other companies that you kind of alluded to, that maybe one that reported this week, you're not just a silicon company.
• 11:39
  SPEAKER_02: Um, you know, how how uh do uh Helia and NeuralSpot and the rest of the software stack make edge AI easier to deploy and how does that help Ambit compete?
• 11:49
  SPEAKER_00: Yeah, you know, we we realized at some point that we're not just a hardware company, right?
• 11:54
  SPEAKER_00: We had to be a software company.
• 11:55
  SPEAKER_00: If I'm gonna get to have the tagline that I'm enabling intelligence everywhere and I'm giving my customers new intelligence capabilities, I've got to give them the software that makes that possible too.
• 12:04
  SPEAKER_00: And so we've invested heavily in all kinds of software from graphics to security to uh to connectivity, but most recently also on the AI side.
• 12:12
  SPEAKER_00: So we have this Helia platform that you mentioned.
• 12:14
  SPEAKER_00: Um, Helia is a collection of um libraries and runtimes and utilities that allow our customers to quickly develop their own AI.
• 12:22
  SPEAKER_00: So, for example, it includes a whole library of AI kernels.
• 12:25
  SPEAKER_00: This is all the little functions that implement the various uh layers and operators that that would pop up in a normal uh neural network.
• 12:33
  SPEAKER_00: And we're we're making them more efficient specifically for our for our chip.
• 12:38
  SPEAKER_00: Uh we include uh runtime multiple runtimes, which is uh an AI runtime is the thing that executes the AI model itself.
• 12:45
  SPEAKER_00: And it plays a pretty profound role in the energy efficiency and the memory footprint and all that good stuff.
• 12:51
  SPEAKER_00: And then we have utilities uh uh um that that will monitor things like uh how many uh how much memory are you using and where are the bottlenecks, um, where how much energy are you using and where are the bottlenecks there.
• 13:05
  SPEAKER_00: Uh and so um that's that's a tool called Helia Profiler.
• 13:09
  SPEAKER_00: Uh we want to provide that whole stack up of solutions so that our our customers can can go build models.
• 13:16
  SPEAKER_00: And I like to say that we're providing the picks and the shovels, um, and we're letting our customers mine for gold.
• 13:21
  SPEAKER_00: And and for the most part, that's true.
• 13:23
  SPEAKER_00: Um, that said, some of our customers are asking us to provide models, pre-trained models.
• 13:29
  SPEAKER_00: And so we provide uh AI models.
• 13:31
  SPEAKER_00: We we have got a smart team here that's building AI models, and some of them they provide as templates that can be used by our customers to uh develop their own models.
• 13:42
  SPEAKER_00: In some cases, it's actually production-worthy models that we're providing to them.
• 13:45
  SPEAKER_00: So um that'll be an area where we we're investing heavily in the next few years, and and you can expect to see a fair bit of innovation on that front.
• 13:53
  SPEAKER_02: Great.
• 13:54
  SPEAKER_02: So um longer term, you've you've uh discussed licensing spot and taking into new ship classes.
• 14:00
  SPEAKER_02: Um just how far can the technology travel beyond your current markets?
• 14:05
  SPEAKER_00: Yeah, uh as I said earlier, the transistor doesn't care what market it's used in, and so we've got a pretty wide latitude on where we can take this.
• 14:14
  SPEAKER_00: Um we've got those four markets I talked about earlier that we're focused on today.
• 14:18
  SPEAKER_00: Um, in the midterm, I see us building chips for robotics as one of the big opportunities to think robots and drones.
• 14:25
  SPEAKER_00: Um and then there's a handful of markets where we're looking at licensing as maybe a better opportunity.
• 14:30
  SPEAKER_00: Um, automotive being a great example where there's a big barrier to entry in that space.
• 14:34
  SPEAKER_00: And um, you know, frankly, there are a lot of companies that build their own chips that will never they'll never be competitive against us.
• 14:40
  SPEAKER_00: Meaning they build chips for automotive modules and that's all they do.
• 14:43
  SPEAKER_00: They're never gonna pop up at one of our customers uh competing against us.
• 14:46
  SPEAKER_00: So I could license to them.
• 14:48
  SPEAKER_00: And so we're exploring that in automotive, in data center, uh, as well as a handful of other markets.
• 14:53
  SPEAKER_00: Um, data center being the extreme, right?
• 14:55
  SPEAKER_00: Where you go, wow, that's pretty far afield from the edge stuff that you're doing.
• 14:58
  SPEAKER_00: But look, they care immensely about power.
• 15:01
  SPEAKER_00: Uh and and so we're looking at what the opportunity might be there.
• 15:05
  SPEAKER_00: Um, I would say the key difference there is that they're operating at the most bleeding edge technology nodes.
• 15:11
  SPEAKER_00: Um, we're operating several generations back.
• 15:14
  SPEAKER_00: As I said, 12 nanometers is the most advanced we're doing right now, and six nanometers is the next up.
• 15:19
  SPEAKER_00: So there's a bit more of an investment to get there.
• 15:21
  SPEAKER_00: Um, but I would say, you know, it's it's a few years out still.
• 15:24
  SPEAKER_00: So we at IPO time last year, we talked about a three to five year time horizon for licensing the technology, and um we're still working toward that.
• 15:33
  SPEAKER_02: Great.
• 15:34
  SPEAKER_02: Right.
• 15:34
  SPEAKER_02: Last one to kind of wrap up here.
• 15:36
  SPEAKER_02: Um just over the next 12 to 18 months, like what are the technical or commercial milestones that investors or listeners should watch most closely?
• 15:44
  SPEAKER_00: Well, uh, you know, on the commercial side, we're just we we've shown continuous growth, we've had five consecutive quarters of growth, we've we've um guided a sixth consecutive quarter of growth.
• 15:53
  SPEAKER_00: And so uh we want to keep that, we want to keep that going.
• 15:56
  SPEAKER_00: Um but but more specifically, we have um two new products in development Apollo 340, which we talked about, as well as the first generation Atomic.
• 16:04
  SPEAKER_00: Uh it's the first time, by the way, that in the history of Ambic that we've had multiple products being developed in parallel.
• 16:10
  SPEAKER_00: This is one of the great benefits of going public and having these follow-on offerings, is we've been able to build out our team and build out our IP uh portfolio to support that.
• 16:19
  SPEAKER_00: So we've got to get those products out, we've got to get them sampled to customers uh and then start the production ramp uh toward the end of next year and then more appreciable revenue in 2028.
• 16:28
  SPEAKER_00: So um we'll be reporting on more specifics related to those two products, you know, when we get first samples to customers and and and and as we start getting customers moving into production.
• 16:38
  SPEAKER_00: But those are definitely things that we want to track.
• 16:40
  SPEAKER_00: Um now that said, we've got a bunch of products that we've launched, and they are getting design wins every every week and every month.
• 16:46
  SPEAKER_00: And and so I think that's also the near-term thing that we want to track is let's show that we've got that continued growth, that we've got a strong pipeline for the future.
• 16:54
  SPEAKER_00: Uh and um look, we're excited about how all that stacks up.
• 16:57
  SPEAKER_00: We're stacking up wins with the existing products, we're stacking up great relationships on the new products.
• 17:01
  SPEAKER_00: So I'm very optimistic about where things go from here.
• 17:05
  SPEAKER_02: Oscott, this is a really exciting story, and you're right, right in the thick of uh of AI and all the action is happening.
• 17:10
  SPEAKER_02: So uh thank you so much for uh for joining us.
• 17:13
  SPEAKER_02: I'm gonna pass it back to Tim for uh final comments.
• 17:16
  SPEAKER_01: Thanks, James.
• 17:17
  SPEAKER_01: And I I agree with everything James has said.
• 17:18
  SPEAKER_01: So thanks for joining us, Scott.
• 17:20
  SPEAKER_01: Um, just intuitively, I would tend to think that given the proliferation of small electronic devices, your total addressable market has to be massive.
• 17:28
  SPEAKER_01: It's huge.
• 17:29
  SPEAKER_00: Yeah, and it's growing all the time, right?
• 17:30
  SPEAKER_00: There are billions upon billions of devices that we can serve.
• 17:32
  SPEAKER_00: And look, I'm I'm really proud of the 300 million plus devices we've shipped to date, but it's nothing compared to the big opportunity that's ahead of us.
• 17:39
  SPEAKER_00: So uh yeah, very, very excited about we're we're at the the front edge of what I think is the biggest technology revolution that any of us has ever seen.
• 17:46
  SPEAKER_00: So excited to be part of that.
• 17:49
  SPEAKER_01: 100% agree.
• 17:49
  SPEAKER_01: Uh, thanks so much for your time.
• 17:51
  SPEAKER_01: Thank you for listening.
• 17:52
  SPEAKER_01: And don't forget to subscribe, as well as visiting www.watertowerresearch.com to stay up to speed on the company's small cap written research reports, podcasts, fireside chats, industry specific symposiums, and conference schedules.
• 18:07
  SPEAKER_01: We will see you next time for another edition of the WTR Small Cap Spotlight Podcast.
• 18:13
  SPEAKER_01: Finally, a special thanks to the producer and editor of the podcast, Christoph Fitzpatrick.