Podcast Archive - StorageReview.com - Podcast #133: Effortless Liquid Cooling with JetCool

Episode Date: September 12, 2024

Closed-loop liquid cooling from JetCool delivers power efficiencies and increased CPU/GPU performance. Liquid cooling… The post Podcast #133: Effortless Liquid Cooling with JetCool appeared first on StorageReview.com.

Hey everyone, welcome to the podcast. Today I've got an interesting conversation, I hope, with Bernie Malouin, the CEO and founder of JetCool. JetCool is something that we've had in the lab recently. They're a really interesting player in the liquid cooling space. We just had an R760 in with their technology. It's a self-contained cooling loop. And I know it's cooler than that. Bernie's going to tell us all about micro jets and things like that, I'm sure. But I ran into these guys really at Super Compute last year, but again at Dell Tech World this year in May. And we got to see some of their demos on the floor. There was a lot of excitement around this form of liquid cooling because it's self-contained it makes it really

easy for enterprises and even some really unique situations to adopt liquid cooling without the complexity of a full liquid loop facility water and all that sort of thing but bernie thanks for coming in and joining us absolutely thank you brian for having me this is a great topic and we're excited to share some of the technology and some of the testing that's been going on here at Jekyll. Well, there's nothing hotter right now, I suppose, than the liquid cooling trend. And AI is dragging us one way or another into liquid cooling. But the R760 we looked at that you guys shipped over had no GPUs at all.

So there's definitely a case to be made for efficient cooling for CPUs as well. Yeah, I think we're seeing that too, Brian. So a lot of hype on GPUs and AI for good reason, getting a lot of use cases. And Jekyll does those types of applications too. But there's still a core need on the more generalized in HPC through CPU computing. And that's something that, of course, we provide in that 760 product that you evaluated. In the use cases, they are still very, very strong.

Because it's not just GPU power roadmaps that are going up. It's CPU roadmaps that are going up too. And so, the play for liquid cooling kind of spans both of those regimes. Well, the thing that strikes me is every time we look at these technologies and the same thing with yours is that it's not just supporting those high power CPUs, the efficiency message is much greater than that.

By being able to cool the CPUs better, you typically will get not big, but a little bit of performance boost. The CPUs will be cooler than air cooled so that the chips themselves and the components there will run cooler. And even though with air cooled, they're still within spec, I think we can all agree that running things cooler should help with longevity for any silicon or any small electronics. And then the other thing is fan speeds can at times be cut down, which is one of the biggest power sucks of any server. So talk a little bit about how your solution is sort of addressing some of these concerns and why it's more than just enabling the consumption of these high-power chips.

Yeah, no, that's absolutely right. It's not just the high-power chips that people are using Jack Cool for. Because, look, I think today with AI and the power consumption from GPUs, there are real constraints in power budgets for a lot of our customers. And we have solutions, of course, for the 100, 120 and higher kilowatt racks, but a lot of our customers are running at 10 or 12 kW racks. That's their power budget. And it gets back to exactly what you said, Brian,

that if you have a very power-constrained facility or rack, if you can increase the efficiency, even of the generalized compute of the CPUs, that gives you more capacity to get more compute in. If each server uses 15% less electricity, that means that rough math every seven servers, you can get an H1 in that power budget and increase your overall compute output. So that's been very, very popular for us. And we do that through the efficiency of building that liquid cooling into the server. It's still ultimately air cooled. It's a liquid-assisted air cooling, if you will. But those fan speeds go down a lot. And we're dialing those back 40, 50, 60%.

And that can pull out 100, 150 watts for every server. Otherwise, you'd just be spending turning fans. For nothing. The fans have a job, but it's really kind of excessive at this point when you think about when we're looking at like an R760 under heavy compute load, air cooled, those fans will spin at 100%. And like you're talking about with HPC or financial analysis, or there's so many other workloads that are heavily computational on the CPU, that will spin 100%, 100% of the time. And that's a tremendous power budget for just the fans. And I think that's one of the things that I think we all know intuitively in the IT industry.

But I just don't think, unless you've done the math on how much power those fans eat, a lot of people still don't know. Yeah, it's really shocking to see how much of your IT budget, the server power consumption, goes to turning fans. And you don't get anything as a return from that. In fact, you get worse than nothing from it because you end up operating in your facility. So let's talk about the technology, because you did make the clarification, liquid assist, and I think that's really critical. The loop that you have, and I don't want to diminish it, so we'll get into the technical stuff that makes it cool, but this is not

dissimilar from the concept in a PC gaming rig, where you have a cold plate on the CPU, you've got a closed loop and a radiator in the system, fans blow across the radiator, which helps distribute the heat and out of the system. Yours is not dissimilar from that overall concept, right? That's right, Brian. So, you know, very similar concepts. Think of ours as kind of the industrialized jazzed up version of that, right, for today's modern kind of computing needs and power levels. But at principle, it's actually, you know, very, very similar concept, but we've had to make some innovations throughout to be able to hit the power levels of today's packages in this compact design. So let's dig into the pieces then that make this different. And I think the cold plate

is probably your highest point of innovation. The cold plates, it's remarkable. Three years ago, they obviously existed, but not in the way they do today. In fact, I get hit from companies I've never heard of before. Two of them this week, in fact, that have new cold plate technology. And I just can't believe how many companies are out there with these cold plates. It's, I don't think it happened overnight. Maybe they were always in it and now it's just really popular, but there's a tremendous amount of engineering potential in the plates themselves. I know I teased micro jets at the beginning and you were showing me some of that at Dell Tech World this summer. But the nuance there in terms of efficiency of your ability to cool these systems, I think

is a major differentiator. So tell us a little bit about that. Yeah, so our cold plate technology is different from many of the others that you'll see, perhaps all of the others that you'll see, perhaps all of the others that you'll see, where others focus on general ways to increase area within the cold plate to get a little bit more cooling surface area. JetCool's approach is fundamentally different. And what we've built our company on is our proprietary approach

of taking these arrays of tiny fluid jets. So inside our cold plates, from the outside, they look like any other cold plate. But on the inside, there's these arrays of tiny fluid jets that are targeting hotspots on the silicon. And so we work very closely with the chip makers today to understand where all of those heat loads are on the processors. And then we're architecting these arrays of jets to target those hotspots. And when we do that, we end up with anywhere between 500 and 1,000 little tiny jets cooling a single piece of silicon. And that makes all the difference because then you can think of us as really the surgical way of removing heat instead of trying to take the more generalized approach.

So let's dive into that. Does that mean then with your microjets that the cold plate understands where on that CPU lid there's more heat than not? Probe into that a little bit further because I think most of us think of cold plates as water in, water out, and then some sort of like little labyrinth in between, possibly other disruptors and other technology in there, but how specifically targeted can you get with this? Yeah, so it's, I mean, from, the great thing about it is from the use case, it actually is that, right?

It's very simple. It's water in and water out. But inside, we are designing to hit those individualized hotspots. Now, that's something that we do during the design process of a module for a particular chip family. And then, so then it doesn't take any other power consumption of adaptive things inside of the cold plate. So it remains very efficient and very passive electrically in that way. You're profiling the CPU early on so that you understand that for Xeon, fifth gen, whatever,

that the typical hotter areas are distributed maybe a little differently than the chip before it. That's right. That's right. And we'll test it under different workloads to make sure we get all those spots that are lighting up. Okay. Yeah, that's a, I'm sure a perpetual challenge with every development of, from,

you know, Intel, AMD, NVIDIA, whoever else that the one that you're cooling today is probably different than the one you're going to cool tomorrow. Well, that's why our relationships with the chip makers are so important, right? And we have great relationships with all three of the big chip makers to make sure that, you know that we're ahead of the curves on those. And then we've done some things on our design side to make sure that we streamline that and make it very, very quick to make sure that we're to market when those new chips are released. So those guys are your partners. So don't throw them under the bus, but I've got to imagine if I'm you, there's almost no chance I get from them exactly what I need to design a cold plate for their chips because they're not designing for

that. They're designing for other specs and other outcomes. And they probably, I'm guessing, don't give you a nice little heat map and say, here, Bernie, go cool here, here, and here more than the rest of the chip. You still must have to do a tremendous amount of hands-on characterization to figure that out. Yeah, we've actually had great partners on the chip making side. And so we do get a good amount of detail from them. And then we supplement that, you're right, Brian, with a lot of internal validation here at Jekyll, because the picture is never complete. And then we have to put that together, too, for when we're actually looking at test campaigns for different validations. Because a very popular technique will be to use a TTV for validation.

But TTVs don't always represent the actual heat map of silicon, right? And so that's another detail and nuance that we have to work through here, but we've become very good at working with customers and with partners on that. Okay. So the cold plates are there. We've got pumps then that move the fluid around. So let's talk a little bit about that. In our system, there were two on the edges, little tiny. I mean, they didn't even really look like pumps. They're almost like little beehives if you're trying to visualize what these look like on the sides. Talk a little bit about the pumping mechanisms, what sort of technology is involved there.

Is that yours? Is that an off-the-shelf part? How does that come together? Yeah, that's a technology that we've worked with a partner on for adapting for these kinds of applications because to get into these systems, you need something a little bit different. And so we've worked on that with a partner. It's a micro-DC brushless pump, and there's two of them in our systems. They circulate that fluid throughout the server. And does that fluid go at variable speeds depending on the temperature? Is it monolithic? How does that pump intelligence work? Yeah. So it's interesting. These are things that we think about in the design. We've gone through a bunch of those trades, Brian. Today, the pumps are statically set for speed by us at JetCool, depending on the platform and the application. And, you know, there's actually not a lot to be gained from the adaptive part of it, from what we've looked at. You know,

you could maybe get a few watts more efficiency if you scaled it and dialed it based on workload. But it doesn't seem, it doesn't seem worth it, I guess. So for the increased complexity. So today, we have a very simple and reliable solution that is statically set by us. And then we know, and we know with confidence where each of those systems performs. Well, I guess, I mean, one of the things that from a JetCool perspective that as we're looking at these systems is interesting is that if you don't know that the JetCool loop is in there, you don't know it's in there. And I guess that's the bridge that you're talking about here, that you're trying to make it simple yet effective. And so you're going to take

a lot of that tuning work or any sort of maintenance or management work away from the system admin or IT admin in an organization and just say, just treat it like an air-cooled system because fundamentally it still is, just more efficient, right? That's right. And that's the biggest compliment that we get from our deployed base. That they don't know it's in there?

That they don't know it's in there, right? They say this server- It's a bug hate kind of thing, I guess, right? It works great. You can't tell a friend if you don't know that it's in there and we're making you more efficient. Well, so I mean that's, that's actually why we started putting the face plates on the front of the

servers because people had no idea that it works great and it uses less power. What's going on. So $150 bezel probably goes, uh, goes, goes longer than any other spend you make in terms of making sure that your brand's out there. Okay. So then we have the pumps, got the cold plates and then the radiator. Talk about that because outside looking in, it's a decent sized chunk of metal, but I can't see what's going on in there.

So maybe talk a little bit about radiator design and if there's anything cool there happening from a technology perspective. Yeah. bit about radiator design and if there's anything cool there happening from a technology perspective. Yeah. So we've used some very creative radiator technology, again, with one of our partners. And we've actually modified this technology. Its basis is in Formula One racing. So looking at the intercoolers for some of these race cars, we've taken that same technology and built it into a different form factor, of course, with our partners to be able to get it into one of these computer servers. And that's actually true of several parts of the technology, things like the different hose connections and hose materials and things. We've brought a lot of those from performance automotive

into this application. So we like to say that our servers are good for three years or 100,000 miles. Well, I mean, if you get into the rugged edge, you might have to hold up on the mileage warranty. But yeah, I can't remember the hosing specifically, but we've seen a lot of parts like from Continental, for instance, who has a long history, obviously, in rubber and hoses. But that tech's been out there for decades at this point. Do you get any concerns over clamps and hoses? I can't imagine many people

care too much about that. No, we don't get a lot of thoughts on that part of it. Again, because we're leveraging components and materials from applications that have a long history of reliability. And they've been used in other applications as well. So that part of it's very, very well proven. And then we put that together with all of these components, including the cold plates. The cold plates are something that originally were developed for aerospace applications. So we were doing it for aerospace applications. So we were doing it for aerospace applications

long before computing. And those were very powerful devices too. We like to say in aerospace, we were cooling a thousand watt devices before anybody cared about thousand watt devices. But now with AI, that's become very popular. Yeah, it seems that aerospace has really driven a lot of this Cold plate technology, which is, you know, I think when we think about liquid and data centers,

generally, obviously, mainframes back in what the 60s were doing this, and then it kind of went away and, and the graybeards still around today, maybe are less scared of it than the general, you know, IT guy that works there today. But yeah, the cold plates, a lot of that seems to be a derivative of some of the space technology that's been used over the last many decades as well. Yeah, and I think it's a great heritage for these kinds of technologies because, you know, they, aerospace balances two things. They balance performance and reliability.

And when you look at the data center space, those are two of the top things that our customers care about too. So I think it's actually a very natural progression. So talk about the liquid then a little bit. That's maybe the fourth component or fifth by now. I've lost track of the pieces. But it seems to me the fluid in the loop is perhaps as critical as anything else that you've got going on there. And from your standpoint, I mean that more from a maintenance standpoint.

So when we look at some of the bigger full facility liquid loops, there's a lot of fluid maintenance that has to go on. Or if you're doing submersion, you're using some oils from BP or Valvoline or whoever it is that has the oils, and that's a special substance that has to be managed. But for yours, talk about the liquid in there. What is it? Does the customer have to do anything to maintain that to make sure it's not gumming up the cold plates or that there's something weird growing in there? I mean, those are real challenges that you must be thinking about. Yeah, absolutely. So fluid is actually really critical for two reasons, both that you hit

on, Brian. So the first is kind of the uh the the uh the environmental impact of it right and so uh we use a water-based solution it's a pg-25 uh we don't include the immersion oils or even some of the two-phase you know dielectrics with with our pfas um so we keep it very simple right with the with the water for the environmental and safety standpoint. But then you have all the complications of, well, even with that, with the PG25 blend. One of the things that JetCool has learned over the past six years is that not all PG25 blends are created equally. And so it really does matter. It really does matter the fluid choice that you use. We have two preferred partners on fluid choice that we've gone through extensive validation on. Because you do have to manage the lifetime of that, the maintenance of it, the bacterial growth, the corrosion concerns, et cetera, with PG-25. And it's even more important when you start working with systems that are high power

because the fluid temperature is going to be higher. And that can accelerate both corrosion and biological growth. So it's really, really important because we design and test and build ours for three years maintenance-free. So they are designed to be three years maintenance-free. And if you want to go longer than that, there are simple maintenance cycles that can be adopted to refresh. But for the first three years, there's nothing that the customer has to do. So what would that look like? Again, R760, I've got 10 or 12 in a rack with your plates and loop in there.

And three years, it's not really that long for a server, maybe for its primary role, but maybe now they get shifted to a DR site or something. So what is that refresh look like? Is that a yank and replace? Is it top off the fluids like a Jiffy Lube? I mean, what are we talking about there? It's very much a Jiffy Lube, right? So it's probably just as fast or faster than Jiffy Lube too. So it's a very simple maintenance cycle. And it's just to make sure that everything is still in good shape. They're designed with test data for the three years

and then if you want to extend it, we're just making sure that everything looks the same and if it needs a little topping off of fluid, we can do that and get you back in service. So talk about the fluid then. PG25 and we're holding, I think we wrote it in the paper but i've already forgotten what eight or ten ounces or maybe less how much is in in your loop typically yeah so yeah we are typically in the eight to ten ounce uh range so hey look at that yeah okay so so it's very small and and that's going to be so we we talked about the the buyers that don't know you're in there, but for the ones that do, I'm sure there's, we hear it every time we talk about liquid

of any kind, there's some apprehension that should a hose break or something leak, now what? So walk through that sort of catastrophic scenario and say your system dumps a couple ounces. What does that look like? Yeah, yeah. So everything shipping from Jekyll gets 100% leak tested and we test over pressure.

So we have a lot of confidence and good data on that. But let's say that something happens, right? Because we don't know in the field if things happen. And part of, I think, the benefit of our type of systems with the 8 to 10 ounces of fluid is if something like that did happen, you have some fluid that gets out. The risk mitigation is that the so-called blast radius of that is really contained to one server. There's really not enough fluid. And we've done testing here at Jekyll and in other places to say what happens if all of that fluid drops out of that system. And the damage is contained

to one box. There's no damage that occurs to anything below that in Iraq or other places in Iraq. I suspect though that even in a failure scenario, that it's not like the radiator cracks and dumps out 10 ounces immediately. Typically, and you tell me, typically what we've seen is that there's a very small puncture or a bad connection, bad seal somewhere, and you might be losing little bits over time. And where that would manifest itself would be in CPU temp. You'd see something creeping up there versus a hard stop. There was liquid and now there's not. But what's the most common, I guess? Yeah, no, that's exactly what happens in practice, Brian. So you'll get some kind of very minor leak. You might see that in a slow temperature rise of that CPU.

And then you can respond to that. There's no widespread dumping. I don't know that we've ever seen a situation where the full fluid actually dumps out. Even in our test where we take things and cut hoses, you don't get all of the fluid that dumps out. Right. So you really keep it in other parts of the system, right? Yeah. Yeah, that's right. So one thing about leaks that I think the industry needs to catch up on, and I know OCP is doing some of this work in their working groups, but the big infrastructure suppliers, Dell, HP, Lenovo, Cisco, whoever, could do better is integrating some leak detection or even more intelligence about liquid-cooled or liquid-assisted systems into their management platforms, whether it's

iDRAC for Dell or iLO for HPE or whatever. It seems like there's a little bit there, but the industry could do better at integrating. What's your sense of where we are in terms of monitoring and managing liquid-assist or full liquid loops from a server perspective? Yeah, no, I think that that's a really good area for growth for the industry, right? And, you know, there are leak detection mechanisms built in, especially to some of the larger facility level

implementations. But look, I think one of the real opportunities coming up is we're deploying liquid cooling in many of these facilities to power advanced compute and AI workloads, right? I'll tell you what we're not doing very effectively right now is we're not leveraging the same AI to be predictive. To monitor the infrastructure? For the infrastructure part of it, right?

So as we think about that, there's many things that we can do, given all these signals that we're collecting, things that you mentioned, things like these leak detectors, to be very, very responsive to those or even predictive for those. And I think that's something that's ongoing right now. And we'll likely see a lot of innovation there in the next two years. There's a lot of opportunity for sure. And like I said, I know OCP is working on it. It'll be interesting to see in October kind of at that conference what that looks like. There's a whole day. And I don't know because I didn't go last year. I went two years ago and focused on storage. But there's a whole day this year on liquid. There's two, actually. There's a track on immersion, and there's a track on other loops and such. And that one will be interesting to see

if the creativity that's occurring at Hyperscale, at Meta, at Google, at whoever else is part of the big Hyperscale clouds in OCP, because they're doing this and adopting it now. And I'm hoping that that knowledge transfer really hits home and that the enterprise ISG guys can take those learnings, those notification systems, the little ropes that we see going around CPU trays or cold plates and report that back up and say, and to your point, integrate some intelligence because many of them have AI ops capabilities within those systems. Let's look at more of these things

that maybe you're not used to looking at from an infrastructure standpoint. Yeah, yeah, exactly right. You build in some intelligence there with some of those other signals, because when cooling gets disrupted in any way, there are signals, right? And even before it gets to the point where you have some larger event, there are smaller signals before that that can be folded in potentially

into some of these more adaptive techniques. Yeah. Okay. So walk me through a little bit on the procurement process. When we looked at the 760, you guys sent it, showed up, everything was good to go, and we just threw it in the rack and got going. But for a customer that wants to buy a JetCool enabled system, whether it's Dell or any of your other partners, what does that look like? Because I think there's, you know, if you go look at the website, you know, they don't have a liquid option typically in the dropdown. And I know that's not how people are buying in volume, but what does that buying process look like? Do your customers drive that? Do your partners drive it? How does

that work? Yeah. So it's, you, so it's something that people come to us at Jekyll or through our partners, and we connect them through the appropriate channel. Jekyll is available worldwide today through Dell, and that's largely through the channel partners who offer the comprehensive service support and warranty coverage. We're happy to connect it to the right channel partner there. You're right, it's not the drop-down box on Dell.com or some of our other partners today, but it is an upgrade, so it's an upgrade selection that can be made through the channel partners, your Dell rep, et cetera. All right, so your trick is to get more people to know about you to help drive that pull

through the channel and get these systems out there. What is the cost implication? I don't want you to pin yourself into a dollar figure corner here. organization and know I'm going to add 50 or 100 servers or whatever the number is, what sort of delta are we talking about to go from an air cooled 760 or 660 or whatever to a jet cool enabled system? Yeah. So generally speaking, it's around a 10% upgrade for the system price. And that'll get you that liquid cooling completely self-contained within the server and all of the performance and efficiency benefits that go along with it. I mean, you must have a little calculator for people to see that because 10% on the base hardware, you got to get that back. What's your ROI calculator look like? Or do you have a time window for when you break even on that?

Yeah, so in many cases, it's within six months. And that's largely driven by, at least for the self-contained products that we offer, the ability to recapture that additional compute in your power-constrained budgets. So that pays back very quickly with a 10% upgrade fee, again, on the base hardware, not all the other things that you'll be adding to it. Right. Okay. So that makes sense from a math standpoint. It's almost, I mean, at that point, really, if you're pushing these servers hard, I mean, I don't want to sell for you, but it's kind of like, well, why not? Because if you can get the payback that quickly, that's a pretty great value proposition. I'm not sure you guys think so. We're certainly bullish on liquid cooling in general, all of liquid cooling capabilities

that are out there. But certainly that's been fueling our globalization with Dell too, right? It's the very quick adoption of this kind of technology because it's not just suited for on-prem enterprise, it's suited for larger deployments, it's suitable for co-location, right? Co-location where you might not be able to control the infrastructure, but this is a great way to get liquid cooling and efficiency capability into a co-location facility. We have great partners in the co-locations today. Well, it makes a lot of sense. Colos, I know we talked to before about you've got some customers that are physically limited by a certain amount of rack space. So I've got 20U or 48U or 100U,

and that's all I get. So in those scenarios, and they're not going to bring facility water in, I need to maximize my efficiency. Do you have that? That's an interesting one. Are there other spots where you've seen the JetCool solution slide in that's surprising or otherwise interesting? Yeah, you know, we've seen a couple of on-prem deployments that are actually very interesting for that, for largely the same reasons, right? They might have a very small space, rack constraints, power constraints, and it's a great way to kind of upgrade without having to make a lot of investment, really any investment in infrastructure. Co-locations are another one

where the power budgets are typically very constrained, especially if you're in a popular area. So that's another great use case for us as well. So, but we've seen other ones too that are a little bit more hyperscale, where they still think of this technology from the energy savings while they figure out how to make the investments in the CapEx for the full-scale liquid cooling, right? Where ultimately we view this as a bridge. Things will be going to the facility level liquid cooling at some point,

and that's why Jekyll has a separate product line. But this provides a nice bridge as you do the capex planning for those bigger infrastructure, which could take three or four years. Well, yeah, I mean, you talk about the practical application of power consumption in the rack, and it's actually something that we're fighting with even in our lab here in Cincinnati, we have an air-cooled eight-way GPU system coming in, and the power draw on that thing is an absolute beast. And when we designed our data center, whatever, 10 years ago, eight years ago, we had a vision for where power was going, and we were vastly incorrect of these systems, because at that point in time,

I mean, GPUs existed. Machine learning was a thing, what we used to call business analytics or something before then. But these GPU systems with, with the, the power of, of, uh, an eight way system, we didn't contemplate. And maybe we screwed up or maybe we're just like everyone else or most people that just didn't see that coming on the roadmap. So even if I have the physical rack space, which I do, I can mount this thing. It's heavy, but I can make the intern and Kevin do it, get it into the rack. Power is now really important because my PDUs aren't rated to take all that power, just

forgetting about the efficiency for a minute. So now I've got to balance multiple rails. I've got to balance PDUs. I've got to balance my battery backups. I mean, there's so much that goes into it. And now I've taken my power envelope for six racks in our little data center and put it into essentially one device with just a little bit of headroom. I actually wanna drive these GPUs with something to bring the proper compute in there. Having an extra 10 or 15 or whatever percent efficiency out of my compute servers would go a long way

to help me with that density. I still probably will not go 42U full rack with these GPU servers in there, but at least it gives me the opportunity to to get some more density in there where it would be impractical otherwise yeah yeah i mean for sure right it's a real growing challenge and i don't think anybody a couple years ago uh saw uh saw where it was going brian you know we do some work uh for an arpaPA-E program that we have. And when we were setting the goals for that, this is just a couple of years ago now, and those program goals were being set out,

the targets were 120 KW rack, right? It's at a really hard challenge, 120 KW rack. And now, I mean, geez, you look at the NVL-72s that are coming out, they are 120kW. It's no longer a dream anymore. And, you know, talks about- It happened fast. Yeah, and it happened really fast. So, you know, it's hard to keep up. That's why it's, you know, always looking at the capacity for the future. And, you know, the other part of it is just on the facility side, right? Even with the liquid cooling or for some of the air-cooled GPU boxes that you mentioned, your facility air flows, you start to tap up against what the constraints there are for some of those.

Absolutely. And so there's all kinds of things that you can do there. We've actually worked with some co-locations because of the, our system by turning the fans back, of course, consumes less airflow. So we reduced the server airflow about 30% for the CFM capacity, which has other benefits, you know, kind of upstream at the facility for air handlers, et cetera. I just like that it's quieter, to be honest. That's one of the

best benefits, I think, is walking in there and not having to throw earplugs in. Or I mean, at my age, I've already lost half my hearing anyway, so what's it matter? But for the young guys, it might be more beneficial to work in a quieter environment. Okay. So for anyone that's listening that wants to check out one of these servers, what's a POC look like for you? How can they sort of leg into this and try it out? Do they need to go buy a server with your kit? Are you dropping some off with high-value customers or prospects? What does that look like?

Yeah, absolutely. The beauty of it is that it's very easy to try because you can you can try one server because they are plug and play into existing air cooled racks uh you know folks can come uh come to us here at uh jekyll we'll connect you with the partners and and you know we can make those kinds of things happen uh for uh for for for our high value customers uh and and folks that are really interested in getting getting started with liquid cooling. So absolutely reach out to Jekyll or come and see us. We're on the show circuit, come check us out. I think next big show for us is OCP this year in October. We'll have a full line of hardware there. We will see you at OCP. I presume you'll be in Atlanta for Super Compute as well. That seems like a good spot for you. Absolutely. We'll be there as well with full product lines.

All right. Well, if you're at those shows, check these guys out. Also, JetCool.com, right? That's it. Okay. So we'll link to that in the description. We'll also link to our paper. Bernie and his team sent us an R760. We played with it for six, I don't know, four, five, six weeks.

And it just works. And I think, especially when you look at the ROI on the little bit of investment to get a 760 or whatever Dell product it is that you want with these plates, this loop, it's really a pretty fantastic offering, very low risk, and is the bridge to take you from air to eventually liquid if you want to go there or not. I mean, you may not need to do that. But Bernie, it's really cool tech. We appreciate you sharing it with us. And thanks for doing the podcast today. Absolutely. Thank you, Brian.