Smart Air, Smarter Cooling: Actasys's Innovation in Data Center Technology

Speaker A: Budget overruns, brick devices, data breaches. Building connected products is hard. Welcome to over the Air. Sharp unfiltered conversations with executives about their IoT journeys, the mistakes they made, the lessons they learned and what they wish they'd known when they started. I'm your host Ryan Prosser and, and

Speaker B: this is Bill Brock. Welcome to over the Air. Are you struggling with in house fleet management? Have your ansible scripts grown out of control? Has a docker first deployment strategy left you needing more? Oh, it's not just you. At parideo, we work with software teams at hardware companies so they can focus on their products and relax on the infrastructure. Are you interested in learning more about our device centric CI CD solutions? Talk with an expert today pretio.com hey everybody. Welcome to the show today. My name is Bill Brock, CEO of Prideo and we have a very special episode for you. I have David Menu, CEO and co founder of Actasys. Uh, David, welcome to the show.

Speaker C: Thank you so much for inviting me. It's a pleasure to be here.

Speaker B: Absolutely. Um, really excited to learn more about this product. Looking at the website, super cool. Do you actually have a, did you say you had one of these on your desk? Could you just hold it up really quick?

Speaker C: They do, yeah.

Speaker B: So what do you guys do? What is this device?

Speaker C: This is an actuator. It's basically very similar to a fire barrel in concept. It's just electronic. It's an obviously an IoT device. It's based on piezoelectric technology where basically we have a cavity uh, with an orifice, two piezoelectric membranes, ah, that basically oscillate and uh, using a fine wave, generally speaking we generate at air in frequency in different frequencies. To me the best analogy is probably we, we are to air what laser is to light. We, we operate, we analyze the problems in terms of their fluid structures and we use uh, very specific frequencies in order to provide very precise impact or the type of tasks that is at hand. Different frequencies, not very different than for example for uh, noise canceling. Uh, headphones have different um, levels of encrement with the surrounding ear and they will deliver a different impact time as uh, laser at different wavelengths will have uh, a different impact uh on uh, the light.

Speaker B: Simple, simple, simple terms. It looks uh, similar to a diaphragm and you're pushing air out. But it's far more sophisticated than that that um, given the uh, the conversation. So where did you start? Like where, where did you even come up with this idea?

Speaker C: So first of all we Are originally a spin out from Brain Cleo Polytechnic Institute. And the technology, the technology is not new in academia it's been uh, different types of uh, actuators that are part of what's called active flow control. Which is a branch branch of aerodynamics, has been in academia for almost for decades and they've been used for mainly for aerospace applications. So usually what's called separation and circulation control around the airfoil and later on also on reduction of structural, mitigating structural vibrations on uh, other types of airfoil such as, you know, wind, wind, uh, turbine blades. We started a company actually looking at aerodynamic aspect but more related to vehicles. So basically reduction of aerodynamic drag.

Speaker B: Yeah. So you got you these on the vehicle and they're adjusting the flow as it's going around the car to smoothen.

Speaker C: There you go. It's basically a boundary layer manipulator. Every, every body, every bluff body, when it moves through air, it pays a price for pushing the air around. Okay. Uh, a class A truck is pretty much a brick that you push forward. And more than 50% of the, the fuel actually is spent on dealing with the uh, dynamic drag. Uh, a lot is also goes into the frictions with the ground with tires. And what you're trying to do with manipulating the boundary layer is basically to attach it reduce pressure fluctuation on both sides. Many things that at the end of the day you pay a uh, toll. You pay for that friction, you pay for that mess, for turbulence, for places where the air doesn't go where you want it to go or reduce or increasing the wake, you pay a price and that comes down ah, fairly quickly to the fuel consumption when it comes to vehicles. That's where we started.

Speaker B: This is fascinating. I had no idea. Is your background aerospace, cfd, anything in that area or.

Speaker C: I am an architect. Oh wow. Um, I used to build buildings, uh, until my PhD at FBI. In that PhD actually I focused on two areas that are related to buildings. The first of all is the sway of super tall buildings such as the ones that you see in New York. Very high aspect ratio between the base and the side or the height. And in these cases what you see is because of it called crosswind response due to the wind passing in brushing against the building. You start to see the building oscillating. And when it oscillates has nothing to do with any kind of way or critical mission threat to the building. It just has to do with low level of accelerations that are at the top level. At the top floors. Those are tiny accelerations but they're enough to make the occupancy very difficult at those levels. And since that's the prime real estate, you do many things on the geometry of the building to deal with that. One of a great example is the Taipei 101 tower, where basically they have a uh, pendulum of about 1100 tons at the top of the building, five stories swaying to kind of keep the building without any vibrations. Yeah, it's insane and it's a huge price to pay for such a small phenomena. So my PhD was looking at trying to mitigate those structural motions with using not exactly this type of accurateness, but using air that the building would eject externally using the H Vac in order to attach the airflow around it and prevent that sway.

Speaker B: So uh, you fell into cfd?

Speaker C: I fell into cfd. But actually a lot of my work which led to doing this for vehicles, uh, a lot of wind tunnel experiments. So most of my work was uh, actually wind tunnel experiments at small scale. Started with buildings, moved to small 1 to 14 scale. Kamiyah trucks, toy trucks with smaller actuators or simulated actuators. That's actually how we've done the proof of concept.

Speaker B: So um, I want to get back to your product. But tell me, are there buildings where I can go see a stationary building with a full swinging pendulum inside? Is that, is that the, is that what the actual experience is? You go and you're, you're at the, you're at the penthouse and there's a giant pendulum in there and it is moving, but the building is staying still. Is that you?

Speaker C: You, you basically have one of these examples is uh, the tower called Taipei101.

Speaker B: Okay.

Speaker C: And in that case, uh, the pendulum is so large that uh, all the top floors are evacuated. There's nothing. But you can go inside and see it's a huge ball at the end of the day, period work. But that's where I started with my PhD from there started to look into aerodynamic drag. And, and as we went forward with the concept and with the device, uh, we started. Look, one of the biggest, uh, part of the journey for entrepreneur is obviously the finding the product to market fit. And for us what started with, okay, let's try to see what we can do with a PhD on what a smaller scale on vehicles.

Speaker B: It's still a very large scale compared to where you are right now. Where the, how large were on a vehicle, would it be like full, like paddle sized or is it pocket sized? And you do a lot of these, a lot of these devices A lot

Speaker C: of, A lot of these. A lot of these. And the size was about 4 inches in diameter.

Speaker B: Okay. Okay.

Speaker A: Fairly large.

Speaker B: Okay.

Speaker C: But yeah, would that be on the,

Speaker B: on the exterior or would you have um, like vents to where you're kicking

Speaker C: air out or now installed on the backside of the, what's uh, called the aerodynamic uh, fairings that are on the back of the, of the tractor. So closing the gap between the tractor and the trailer.

Speaker B: Okay, so mostly trucks. Yeah, yeah, yeah, yeah.

Speaker C: So yeah, some of the work was on trucks. We ended up doing some, some work also in the croy in, in with OEMs in on SUVs to reduce their drag. This was before the electric, uh, the electrification, uh, revolution. Now obviously you see how much effort is being put into the aerodynamics of uh, electric vehicles.

Speaker B: Oh yeah, they look totally different than cars did 15 years ago.

Speaker C: Ah, exactly. So there was also some interest there. But again, uh, automotive is definitely uh, on the exterior side is not an easy task. And we were constantly pushing forward to find uh, the best AI application for the actuators.

Speaker B: You're in data centers now. How did that happen?

Speaker C: A few years ago we were still kind of working hard in the transportation space. Let's say automotive and transportation space. Ah, we were approached by a very large uh, let's just say train manufacturer which was very interested to reduce the uh, or to increase the power density by removing the heat. The passive and the active solutions that they had to cool some of the electronic components on their uh, electronic boards and by slimming them down, being able to just put in a rack more, more boards. They knew about us from uh, work that we've done with them on, on a separate project. And that was when the first time, when we were requested actually to use the actuator in order to cool uh, a cpu. The work uh, went very well. And from there we started to look much more in depth and try to understand also the, the new ecosystem, the new landscape of data centers as it started to shape out. Obviously there's huge progress on everything that has to do with large computational models. That that's obvious. Uh, we see the hardware lagging behind. So for us that gap started to be very interesting. Furthermore, the distinction between the processing hardware and the networking hardware started to become also more and more obvious to us as a good place to interject.

Speaker B: And how are you, um, on an energy efficiency standpoint, I would assume that that's a big part of the equation for you in these data centers.

Speaker C: So look, the two things. First of all, we're great on the energy efficiency. We uh, the numbers that I usually use are it's a device that is about an eighth of an inch, 3 millimeter thick. It generates under 5 watts, about 150 to 200 miles per hour. Okay. So obviously it's not a thruster. The decay rate is fairly fast. But for very precise spot on heat removal it, it's awesome.

Speaker B: And are these just like constantly going

Speaker C: or is it um, constantly going? Basically it sucks and blows fairly similar to your mouth and lungs. And by that it also provides a uh, much more, it's called an anterior jet, but it's a much more turbulent, a full of vorticity jet where you can think about smoke wings that rotate around themselves but also rotate around an axis. So it definitely, it provides a very turbulent jet. It is very strong in terms of um, heat removal, mixing. So that definitely a larger aspect of how efficient it is. So the ratio between how much it can remove and how much energy you actually put in is very high. But in addition to that, not, not less important is it's actually the fer. The fact that you can look what you see now more and more in servers is the density of equipment just increasing, increasing, increasing. So in the same, the same server racks, if the early 2000s you would uh, you know, consume 3 kilowatts, you're now looking at 120 kilowatts. So the size didn't change much. The real estate is the same real estate, yet you're trying to do much more. And now you're starting to see very dense places where it's very hard to reach with fans or kind of from afar. The surfaces that you used to have for passive removal, such as heat sinks on uh, a port are getting smaller and smaller. Uh, you're starting to see optical transducers like QSFPs, uh, and so on, much m smaller. So the amount of surface that you have to dissipate passively is, is smaller and smaller. So being able to get very close to let's say a hot pot or an area that becomes really, really warm and there's no way, no other way to get to that. That becomes super important if you want to remove bottlenecks. We hear a combination of yes, it's about energy efficiency, but it's also about the accessibility, the ability to be very close to the hotspot.

Speaker B: And so you're um, customers, pilots, ongoing conversations. Who are you working with, um, for these deployments? Is it with the actual like you're being deployed in the rack mounted servers or the external to the equipment? Is it the IT manager? Is it the actual manufacturer of the chips?

Speaker C: That's an excellent question and we're paddling uh, through the ecosystem. But the immediate answer would be we are installed directly on electronic boards for devices that have to do with connectivity. Those are uh, network interface cards, ah, data processing unit switches and so on. So that kind of equipment. And in those cases we really need to be integrated ah, onto the board itself. I think the major players in uh, in connectivity, network and in communications, um, fairly, fairly large.

Speaker B: And earlier I came up, prior to the call or um, once we got on, you mentioned that you believe this is an IoT product. Like how, what category do you feel like you live in? This is um, such an interesting niche here.

Speaker C: It's definitely a cooling device that I can tell you. Uh, but um. But yes, at the end of the day it's definitely IoT device. Obviously you have a piece of hardware and you definitely have a lot of options to operate it in a data driven form. And that goes through definitely anything that has to do with the direct operations. You can provide different waveforms, you can provide duty cycles, you can play with so many ways to operate it, uh, into function, uh, to make it function on the concrete or discrete component level. Then you definitely have any kind of access to predictive maintenance, to telematics, to anything where the device can be either spoken to or you can read data uh, of it. But then I think the most important part is the strength of this device and to me also the strength of the IoT in general is by enabling to use this in a uh, distributed control manner where you have multiple parts that are part of a larger uh, system and can be. I see this also the future of generally speaking of cooling in data centers, but at the end of the day a uh, large AI driven multiscaler system of systems, whatever you want to call it, but where you have the H vac, the nic, the dpu, the gpu, all of them and the different cooling systems that are running them operating together because only that way you can actually achieve energy efficiency to me at the macro level.

Speaker B: So what's the um, current control plane for this? So like how they're not just running autonomously, there's some local management interface or.

Speaker C: Yeah, but that's something that we really leave to the customer. From our point it can be pretty much like a speaker. You can do whatever you want with it on the amplitude, on the frequency, on the waveform and so on. You can decide to accelerate, you can decide to stop it. You can decide to. Now uh, you know, if you have a triangle of trade offs between for example uh, acoustics, power consumption and uh, jet velocity, you can play between all these. And to say right now I prefer to have much uh, more of a echo mode versus uh, an emergency mode where I blow all out versus uh, you know, whatever you know, roll down lookup table that you want.

Speaker B: And I, I cut you off just a second ago about um, integrating with, with larger systems. So do you, how are you thinking about that?

Speaker C: I, I think many people think about cooling kind of as a, as a relic. Not relic but more in relation to, to processing hardware. So GPUs hundreds of watts, uh, power consumption, large, large calculations and uh, over there you really want to cool in order to save energy to reduce oil consumption. Networking has a totally different currency and has a totally different logic in networking hardware. Let's take a nic. A nic does 50 watts at most to reduce the power consumption. There is mini cube on the data center level. But the NIC is really about increasing throughput, making large amounts of data move around more efficient, efficiently move in time and with very low latency. So that's that uh, that's the currency of networking. And in that case if you're really successful in, successful in cooling uh a nic, you just took it from 800 gigabyte per second to 1.6 tera. It's not about the 15 watts that you cut off. Now on the other hand at the end of the day the data center has a set point, there's a certain temperature that everything is designed to work. Now GPUs are designed for 85, 95, 100 and so forth. So um, a lot of cooling is invested there passive usually in the future probably mostly by liquid cooling, very important. But at the end of the day one of the things that uh, affect really the set point is also the part that are more sensitive to heat or at least that are designed for lower temperature. So for example a nick is designed to work at 70ft. If you can cool uh, uh, a nick enough to allow it to or to enable it to operate at data centers where you actually elevate the temperature. So you elevate the set point temperature of the entire data center. That's where you can start seeing the big gain because now you have multiple devices that can work at higher temperatures. Uh hence you need less cooling. So for me it actually no lift, no, no nick left behind. Uh, if we can make them work at higher temperatures then that would be actually the macro Level contribution of these solutions and it's not the discrete component level cooling.

Speaker B: Yeah. And the whole game here is how to reduce cost on these um, compute clusters past life. I didn't say this earlier but I cut my teeth in grad school on HPC and cfd. I was doing. I was a mesh guy. I didn't quite grock the physics. I was, I did not have the mechanical engineering undergrad. So I was like uh, more of a math, math computer person that got into the CFD world. So I did mostly meshing and it was a lot of fun, great times making pretty pictures. But we would go to um, hyper, what is it? Super compute every year. And it was interesting that the, they'd have the top 500 list which everybody wanted to be on and they'd you know, all the algorithms and you know, math, math people excited to run those, uh, run those problems and uh, metrics, etc. But then you'd have the top green 500 and that was always the, the up and running the next candidate for the next year. And it's really crazy um, just now that all these um, top players in the AI space are just investing so much in energy. Energy is like the, the hot topic right now. And Microsoft's uh, like partner, like what are they like buying nuclear plants right

Speaker C: now and yeah, you have the combo of a nuclear plant in the lake. Yeah, yeah.

Speaker B: And uh, and cooling. People don't even think about it. They're like oh, it's just GPUs are just like. But no, it's that the cooling costs are insane. And so when you walk into a data center for the first time, for those of you that maybe haven't, it's, it is so, so loud. And it's just the H Vac running and just the cost and overhead to keep these systems actually running at the optimal temperature. So this is a very cool localized approach here. How loud are these things? Can you hear them? Is it like you walk in and it's like crickets, like chirping or what's the experience?

Speaker C: So first of all it's obviously it's a sonic device. We operate at uh, the range of 50Hz to about 300Hz, usually speaking 40 to 60 DBA. So you can definitely hear them. But again obviously like you said, when you go into a data center it, yeah it's crazy loud. But uh, it's really not uh, so much something that you hear. I have to say. Generally speaking, acoustic performance in data center is uh, important like anywhere else but uh, fairly A fairly uh, secondary to all the, all the things that you actually are trying to do to the data. But again I think uh, that's important. Also an important distinction between. You see many new technologies, new and old technologies that are using uh, piezo and most of them are coming from ultrasonics. So definitely where you have a preference if you may to have a uh, quiet deployment. And that mainly because they are targeting uh, laptops, we made a decision not to operate in that space. And that decision is also driven by the strength of our technology. And the strength that we have is we can generate a very strong, we have a very strong heat removal capacity. But it definitely comes with round.

Speaker B: Um, has there been any real surprises for you in taking this to market? I know you initially thought automotive. You found a fit. The data centers. Have there been any other like big realizations or um, surprises along the way?

Speaker C: We uh, have surprises all the time. First of all, you started before with a question about uh, the ecosystem. For me it's more the supply chain, uh, just to understand where exactly am I. Uh, in automotive we have the OEMs, we have the tier 1s, we you know, we, we had a backup clarity here. It's a little bit more uh, in the making. But we, we definitely see for example how uh, let's call it um, the accelerated computational technology provider being interested very much in the throughput and in the, the what you can do to data or to the data transfer rate within their devices. But when you go and talk with his integrator, he's actually much more worried about the fabrication process. And what you see is that everybody have thermal issues. And thermal issues are not only at the design stage, but it's also on the manufacturing stage. And what you thought you were doing for one is very different than what you're doing for the other. So for us I think what's interesting is actually the exploration to understand uh, when you reduce heat significantly, who gains? And there are different parties and they have different gains. So that's very interesting for us. Beside that, obviously moving from or between industries, definitely between automotive, uh, in transportation to semiconductor in data centers. It's very interesting to see as a startup how much, how different a customer looks to the startup in both paces. We often, you know, we're not the only one, but we often use the analogy of uh, mosquito dancing with an elephant. You know, you're a startup and then you're starting to do pops with a huge oem. Yeah, in the cultures are so different and the language is so different and the KPIs are different and everything, you know, and it even starts with hey, can we, what, what, what do we sign? Which document do we sign for a poc? No, it, it can't be a supplier. PO like it. I'm not a supplier, I'm a startup. It's a whole different type of engagement. But I think what we really see is, is how much more in certain applications, the customer knows exactly to define what he needs from you. He knows to evaluate what you bring to the table. He knows to actually put a number on that and with that number he understands also to put a value on you, on the product. And for a startup it's really important because the more innovative you are, if nobody can really fize you and accept you just, yeah, you're doing amazing stuff. But, but I don't know to put a dollar on that. And that's very hard for business. So we, we, we just see different approaches to that. And uh, and again I think thermal issues are more understood. The market is much more mature in semiconductors. So they very much know that if you're using you, they're not using something else. And if this is the benefit that you're giving them, uh, the premium that they get for that. And it's much easier to explain your

Speaker B: worth if you make Very cool. Um, you mentioned this was spun out of technology based out of uh, Rochester. Is that what you said?

Speaker C: RPI went clear.

Speaker B: Oh yes. Um, and is the company headquartered in New York or you're in Tel Aviv right now. So where, where is home, where's home base for the company?

Speaker C: In the Navy at Brooklyn. And I'm, I'm, I'm a lot, uh, I'm traveling a lot in between and to customers. But yeah, the company sits in the Navy yard. We've been there since 20, uh, 20. Um, and yeah, it's, it's a, it's an amazing environment and, and place. Uh, it actually dedicated I believe by New York for trying to bring manufacturing startups into that area. Interesting. Um, I think it's also federal land, so they can do like we had autonomous shuttles there for a long period of time. I think that they started to deploy 5G much earlier to allow people to or startup actually do some work there. So file suites are definitely an exciting environment.

Speaker B: What, uh, hard pivot. Uh, you mentioned New York. What's going on with these drones in Jersey? Do you have any idea?

Speaker C: Drive. Your idea, your idea.

Speaker B: This is a crazy times. These are crazy times. I'm ready to be on the other

Speaker C: side of this year, everybody.

Speaker B: Well, this has been fascinating. I really enjoyed the conversation. Um, excited to see your product take to market, and I'd love to have you back on six, eight months here. Progress. And once you get out of some of these pilots and have more. More to share, like, uh, around the early customers and use cases, um, really excited to have you on them. Um, anything else before we.

Speaker C: Before we close?

Speaker B: Any way we can help you?

Speaker C: No, thank you very much, and again, thank you. You've invited me. That's definitely a big help. Thank you.

Speaker B: Yeah, absolutely. Well, we'll stay in touch and, uh, wish you the best luck on your journey. And, um, yeah, thanks for coming on the show.

Speaker C: Thank you, Bill. Thank you very much.