SC24 | The Future of Direct Liquid Cooling featuring, Dr. Luca Amalfi, Seguente & Tim Shedd, Dell Technologies & Dion Harris, NVIDIA

Clips
News
More from SC24

Dr. Luca Amalfi

CEO

Seguente

Tim Shedd

Engineering Technologist, Office of the CTIO

Dell Technologies

Dion Harris

Sr. Director, HPC, Cloud, and AI Infrastructure GTM

NVIDIA

Cooling evolution: How Direct Liquid technology powers sustainable data centers

It’s back to the future when it comes to cooling data center servers. In the era of mainframes, liquid cooling was an important method to cool silicon. But when the microprocessor took over to power general purpose computing, air cooled systems went mainstream. Today’s data centers have been outfitted mostly with air cooled infrastructure but with AI, liquid cooling is back and is one of the most important topics in AI server designs. In fact, direct liquid cooling is the future of energy-efficient data centers, according to Seguente Inc., a global energy and thermal management company, which participated in a panel on theCUBE at SC24

play_circle_outline Exploring Seguente, NVIDIA, and Dell: Innovations in Liquid-Cooled Server Deployment and Two-Phase Cooling Technology

play_circle_outline Evolution and Future of Data Center Cooling: Transition to Liquid-Cooled for Efficiency and Scalability

play_circle_outline Challenges and solutions for transitioning to liquid-cooled data centers

play_circle_outline NVIDIA's focus on accelerated computing model and infrastructure innovation

play_circle_outline Exploration of different cooling options including single-phase and two-phase cooling

play_circle_outline Dell's focus on single-phase direct to chip cooling and journey to 3000 watts per GPU

play_circle_outline Maximizing Data Center Efficiency: Two-Phase Cooling Benefits and Supply Chain Diversity

play_circle_outline The Future of Data Center Cooling: Tech Innovation and Collaboration for Sustainable Power Solutions

Info
Transcript

The Future of Direct Liquid Cooling featuring, Dr. Luca Amalfi, Seguente & Tim Shedd, Dell Technologies & Dion Harris, NVIDIA

Dr. Luca Amalfi

CEO Seguente

Tim Shedd

Engineering Technologist, Office of the CTIO Dell Technologies

Dion Harris

Sr. Director, HPC, Cloud, and AI Infrastructure GTM NVIDIA

The panel at SC24 in the Atlanta Convention Center includes Dr. Luca Amalfi, CEO of Seguente, Dion Harris, director at NVIDIA, and Tim Shedd, an engineering technologist at Dell Technologies. Seguente focuses on direct liquid cooling for AI and HPC, Dell specializes in thermal strategy for data center products, and NVIDIA champions accelerated computing. The discussion revolves around the shift from air-cooled to liquid-cooled data centers, with direct liquid cooling enhancing heat transfer efficiency. Challenges in transitioning to liquid cooling include lea... Read more

explore Keep Exploring

What does the name "Seguente" mean and what is the focus and specialty of the company? add

What is Luca's perspective on the progression of data center cooling from air-based cooling to rear door solutions? add

What is the transition from indirect liquid cooling to direct liquid cooling in relation to energy consumption on HVAC systems? add

What are some of the main challenges faced by customers when transitioning from air-cooled to liquid-cooled data centers? add

What are some potential advancements or value propositions in utilizing a dielectric fluid with a low boiling point in cooling systems, such as the Coldware system? add

What technology has Dell selected for their business case that they believe will provide the best value for their customers at this time? add

What is one of the reasons why two-phase cooling is seen as more challenging to deploy compared to single-phase cooling? add

What are the benefits and advantages of using two-phase cooling technology in data centers? add

bolt Powered by CUBE AI

The Future of Direct Liquid Cooling featuring, Dr. Luca Amalfi, Seguente & Tim Shedd, Dell Technologies & Dion Harris, NVIDIA

search

Dave Vellante

>> Hi everybody. Welcome back to Atlanta. We're here at SC24 in the Atlanta Convention Center. My name is Dave Vellante. Savannah Peterson is also here with John Furrier and we're super excited. Back when I started the industry, mainframes were everything. They were water-cooled mainframes. Every time you did a briefing at IBM, they'd bring out this big cold plate and they'd give you 20 minutes on that. Well, when the microprocessor took over, air-cooled became the thing. Well, liquid-cooled is back to support and power the AI wave. We're super excited to have a direct liquid cooling panel. Dr. Luca Amalfi is to my left. He's the CEO and founder of Seguente. And to my right is Dion Harris, who's the director of Accelerated Data Center, go-to-market at NVIDIA. You've heard of them. And Tim Shedd, he's an engineering technologist at Dell Technologies. Gentlemen, welcome to the panel. Welcome back to theCUBE. Great to see you guys so much .

Dave Vellante

>> Thank you for the invitation.

Dave Vellante

>> So Luca, let's start with you. We know Dell, we know NVIDIA, their household name, Seguente. First of all, what does that mean? What does that stand for? Tell us about the company.

>> So actually Seguente is an Italian award, which means next, whatever comes next. And since Seguente is a technology player, many of the co-founders that they have Italian roots, we basically stick with an Italian word. So again, pleasure to be here today.

Dave Vellante

>> Okay, great. And so your specialty is with Seguente, obviously liquid cooling, direct liquid cooling.

>> So yeah, let me elaborate a little bit more here. What Seguente does, I believe we have the mission to support the scaling of artificial intelligence technology and HPC, high performance computer in a way that we would like basically to deploy the full end-to-end solution. When I say full end-to-end solution, is IT, the hardware, the networking piece, and the storage. Basically we rack and stack into racks of course, and then basically we deploy according also with the cooling systems. But the flagship that we have in Seguente is really our unique passive direct to chip two-phase cooling technology, which basically operates without any water and any pumps. So we minimize the energy consumption.

Dave Vellante

>> Okay.

>> So this is the value added that we provide to our users.

Dave Vellante

>> And we'll come back to that. I want to dig into the technology and how you guys do that. But I want to go to Tim. I was just in your lab a couple of weeks ago down in, I guess it was in Round Rock-

Tim Shedd

>> Correct.

Dave Vellante

>> Right, which was awesome and loud. We were listening to GPUs, we were exercising them, we were feeling the heat coming off. And you guys have done some amazing work there, but explain your role at Dell and a little bit about the lab that we saw?

Tim Shedd

>> Sure. First, thanks for the invitation. It's great to be on this panel. I work in the CTO's office at Dell and I lead our thermal strategy for all of our data center products. So whenever a challenge comes up like needing to deliver racks, hundreds, thousands of racks of 64 NVIDIA GPUs, we need to find a way to cool them and more than that to make the entire rack operate very efficiently. So really cooling is just a tool to enable energy-efficient computing. And so what you saw and what to describe, what we did with the H100 servers that you looked at, we're deploying those 64 GPUs with a rear door heat exchanger. So we liquid-cooled the rack while keeping water out of the chassis. And it turns out that met a whole bunch of TCO energy efficiency and the ability to deploy very quickly to the customer met all those requirements all at once. What you also saw was the beginning of some hints that are next generation of once the GPUs are getting to higher and higher powers, we are bringing liquid into the server, right to the chip in collaboration with our partners at NVIDIA so that we can continue this trend of offering dense energy-efficient, rapid times to deployment.

Dave Vellante

>> And we hear at OCP this year, the call for using a warmer liquid. Of course Dion, everybody at NVIDIA, almost everybody has to be technical. So you're working with customers, helping them figure out how to design to do accelerated data center. Explain a little bit about your role?

Dion Harris

>> So basically as I'm part of the accelerated data center team, and so what I say there is that NVIDIA is no longer selling a GPU. We're really selling customers a data center. We're obviously working with our partners in ecosystem to enable that. But essentially by sort of championing this new computing model called accelerated computing, it's also requiring us to think differently about how we think about infrastructure, how we think about building out these data centers, how we think about the workloads that run in these data centers. And so through doing that, like I said, we're looking at lots of different innovative ways to give more performance out of that computing stack, out of that networking stack. And oftentimes it involves driving more compute density, which therefore requires different innovations, whether it be through liquid cooling, through the bus bars, through the entire infrastructure that supports the specific computing infrastructure.

Dave Vellante

>> Okay. So let's talk about liquid cooling. We have all these data centers that have been built for air cooling and now we're transitioning. Luca, let's start with you. What's your perspective on where we've come from, where are we today? And then we'll get into where we're going.

Tim Shedd

>> Sure. So I guess if we talk about the landscape, so we started very early stage with the aisle containment is air-based cooling across multiple lang scales, which means inside the servers, the rack and the room. And then the progression was as Tim mentioned, that to be able to basically reduce the energy consumption actually move to a rear door solution. With the rear door solution, what we basically do, the primary heat transfer mechanism within the server is still air, air-enforced convection, but the rear door will allow us to basically capture the heat before goes into the aisle of the data centers so that you can basically offload some of the energy consumption on the HVAC systems.

Dave Vellante

>> now.

>> But then of course NVIDIA and Dell and all the big players, we are pushing the power, the TDP of these chips. So we need to go localize on the CPUs and the GPUs. And that was the transition from I call indirect liquid cooling to direct liquid cooling, basically flowing water into the cold plate and the cold plate are directly attached on the CPUs and GPUs.

Dave Vellante

>> So we've seen some of that walking around the show. So it's now direct liquid cooling, meaning the liquid, whether it's water or something else, right on top of the GPU, correct?

Dion Harris

>> Yes. So you basically have your cold plates, which allows you to eject the cooler fluid in to retract the warm fluid out and then go through some heat exchanger to dissipate the heat. So it's a more efficient way of transferring the heat off of the GPUs. And it also allows you to have a much more densely packed compute rack. And so that again leads to a lot of the efficiency that you get out of there as well.

Dave Vellante

>> And Tim, we're used to when you go walking around a data center, you see a hot aisle, you go to the hot aisle, you put your hand back there, it's hot. Well in your lab, the hot aisle's inside the rack so there's not super hot air coming out. Is that a new innovation, relatively new innovation? Explain that.

Tim Shedd

>> Sure. So no, it's not new and all of this technology as you alluded to already was actually originally used for mainframe era, kind of fell out of fashion, but it's really come back with a force. Dell's been shipping tens of thousands of liquid-cooled servers a year since 2018. So this is actually something that's just a continuation and obviously a rapid growth from there. But what you noticed and what we like to think about and I think is consistent with what you all are doing at Nvidia is talking about instead of hot aisles and cold aisles, we're talking about wet aisles and dry aisles now. We're going to be distributing liquid through in one aisle and feeding liquid to those racks and maybe to redo heat exchangers, maybe to DLC, maybe to both. And then what this does is kind of opens up space, it opens up the operating environment so that we're no longer constrained by all of the challenges of getting air to the racks. We can now focus on the compute and building those that compute that our customers need and deliver it, like I said, as fast as possible.

Dave Vellante

>> So Dion, when you're talking to customers, what are you learning about the main challenges of moving from air-cooled to liquid-cooled? I've been talking about potentially could be the weakest link of this, hundreds of billions of dollars and we've got bad connection, integrity. Boy, we would be a foolish industry if we didn't take care of that. You got to design it in upfront almost like security physical.

Dion Harris

>> Yeah, well I think what's interesting is we've been working with a number of partners to actually meet the customers where they are. In some cases they're looking to do a retrofit, like you said, sort of a hybrid where you have a hot, not just hot and cold but wet and wet and dry. So we do have some customers who are deploying hybrid solutions and retrofitting parts of their data center for liquid and leaving others that are completely air-cooled. So working with various partners in the infrastructure ecosystem, it's allowing us to meet the customers where they are. Obviously if you have a green field, you can approach it slightly differently where you're building from the bottom floor and understanding how and where you want to deploy. Say for example, a fully liquid-cooled GB200 MVL72 for example. But there's some instances where we're talking to customers and they're deploying a Blackwell and they want to have some liquid-cooled. They want to have a air-cooled H100 aisle. And so you're starting to see a lot of these mixing and matching where it's not one-size-fits-all anymore. It really is going to be dependent on the workload and then therefore dependent on the customer and where they are in their journey.

Dave Vellante

>> So you guys expect that hybrid, that air-cooled, liquid-cooled hybrid will be the norm going forward, is that right or do you see it as a stepping stone?

Dion Harris

>> Well, I think depending on where the customers are, oftentimes most customers are kind of challenged to spin up a net new data center. And so oftentimes as they're looking to bring on new workloads and adopt a lot of the latest and greatest technologies, oftentimes it involves retrofitting existing data centers. And so through that replacement mechanism, you're starting to see a lot more of these hybrid designs where customers are actually running both liquid-cooled devices along with air-cooled as well.

Dave Vellante

>> Now I would imagine that customers are concerned about leakage. They got some pretty expensive assets on the floor. So how are you assuaging that concern?

Tim Shedd

>> So number one rule we will all agree on don't leak, right? So the first thing first is focus on the quality. This industry, although it's been around a while, is just having to mature really rapidly. And so it takes folks like leaders like Nvidia and Dell to get in there and go to the suppliers and say, "Okay, your quality needs to be here, not here. And so we're introducing greatly enhanced quality checks. We're checking all along the way from integration to deployment to ensure that as well as innovating on leak detection technology." So if a leak does occur, we can react appropriately as the customer wants and you'll see new technologies coming out for rack isolation, for server isolation so that we can address these and address these fears first by don't leak, getting our quality to better than five nines. But then if a leak does occur, minimizing any potential blast radius or damage that could occur from that leak.

Dave Vellante

>> Okay. So we got air-cooled, we got water-cooled, we got warm water-cooled, we have immersion, we have phase change, we have phase change immersion. What am I missing Luca? Lay it out for us the different options that we have. And I want to get into phase change and get your guys' thoughts on where the industry's headed. We're seeing with Blackwell there's some challenges in cooling. Blackwell and the industry's going to solve that. It's just amazing this industry, we run into roadblocks, we blow through them. What's your perspective on the options that are out there and how they're evolving? Lay them out for us.

Dave Vellante

>> Sure. So I guess Tim was mentioning about the water-based cold plate. So we can basically run full water or we can run also a mixture between the water and glycol. This is part of a single phase. What is the definition of single phase that during the heat transfer the fluid stays in the liquid phase and then we are basically using the sensible heat, which is the temperature rise of the water across the cold plate to take the heat from the CPUs and GPUs. Then of course there is another mechanism that I believe is very much viable is to say the industry has already made investment in water-based cold plate. So they have cold plate, cold plate loops, manifold and CD use. Can we have something that basically keep the same building blocks but actually can have value proposition in addition to water? And I think the next level that we are actually working very closely at Seguente is to basically, instead of using just water is to use a dielectric fluid which has a very low boiling point and then you store the energy from the CPUs and GPUs into the conversion of liquid and vapor. So when you can take that energy into the fluid and then you have a liquid and vapor, you can do so many things because now you have two option. You can just keep the same active direct to chip two-phase cooling, use a pump to basically push the fluid into the system. Or actually you can use a passive implementation because now when you have the vapor and the liquid, the vapor is lighter than the liquid. What actually we do and can basically use the buoyancy force to drive the fluid up to a top of the rack in rack CD use where you will condense the two-phase mixture back to liquid. And then you basically in a self-sustained system, you can feed the loop again. So you can think about that in the system that we developed that is called Coldware. You basically you use the phase change and the evaporation and the condensation to move the heat from the hot chips to the cold side, which is the condensation.

Dave Vellante

>> Okay. So we've seen phase change in this industry before way back to the days of optical disc. We did a phase change that flipped from a one to a zero. This is a totally different use case. Dion, I'd love to get your thoughts on this because here's Seguente doing this advanced R&D from your standpoint, hey go for it. Anything that will help with the problem. But what's your take on phases? I've read some things that say that Nvidia's a fan of it, you guys are kind of fan of anything that's going to help solve the problem, but what's your take on this?

Dion Harris

>> So I think in general we're supportive of the ecosystem. And so historically, I mean you have direct liquid cooled options, you have two-phase heat liquid options as well on our Hopper platform. It's just with Blackwell was the first time that we actually I would say productized a direct liquid cooled solution. But it doesn't mean we're not supportive of what the ecosystem is doing to offer innovative cooling solutions. So today for Blackwell it's direct liquid cooled single-phase like we talked about. But like I said, this is very early in the days of these new systems coming online. And so I think it's one of those things that we're always looking at and new and innovative ways to make sure that we can deliver the most efficient solution. So today, like I said, you can see it happening in Hopper. We have a lot of partners who are delivering dual-phase and I think it's a great thing.

Dave Vellante

>> It was a very ecosystem-friendly answer. So well done. Well done.

Dion Harris

>> It is true.

Dave Vellante

>> Okay, but now Tim, you guys are talking massive scale here. Dell a hundred dollar billion company, you're driving huge, huge volumes. We talked and your roadmap Dion is very public and it's not like you guys hide your roadmap, we know what's coming. And we talked in your lab about the march to 3000, 3000 watts per GPU, right? So today you're obviously taking a very practical approach to this. What's your take on the future? How far can your current sort of warm water methodology take you? What's your take tapping into your PhD background on two-phase?

Tim Shedd

>> Sure. So first off, Luca and I know each other pretty well, I love bubbles-

Dave Vellante

>> You love bubbles. We all love bubbles. How can you not love bubbles?

Tim Shedd

>> I was a professor for quite some time and bubbles were my thing. If you look at my publication record, that's what we did. But the reality is, like you said at Dell we don't sell a hundred of something, we sell 10,000 of something or more, right? And what we've done is we've very carefully evaluated the technologies in the field for Dell's business case and for what will provide the best value at this time for our customers. And what we've locked in on is single-phase direct to chip cooling. It's something that we understand, we've got a lot of experience in it and we can scale it. And that's the key thing. We can scale this from tens of thousands to hundreds of thousands to millions of cold plates a year as needed. And we're planning on that. And so that's the real key. The other piece of it is you were saying the journey to 3000, right? So we've got a really strong core team within Dell and we've looked at this and we've said, "You know what? It's not easy, but with some innovation we see that we can keep warm water cooling facility water above 30 degrees Celsius or above 90 F, we can keep that going all the way out that journey to 3000 watts." And that provides energy savings for our customers best TCO and using the power that's available to do the work on the processors to make the money that they're paying for rather than paying for the utilities to run air conditioning. So we do see single-phase taking us all the way out there with some innovation that we're working on today.

Dion Harris

>> And I think one thing I'll just add, I mean one of the main reasons for the transparency and like you mentioned, we've been more public with road maps than we have in the past, is exactly that reason. We want the ecosystem to see what's coming to plan for it and sort of innovate and adopt where needed.

Dave Vellante

>> You guys are transparent to your customers that are trying to build competitive products, which is the right thing to do. I mean, hey, the industry, it's a big business. It's not like any of these hyper-scalers are going out of business, but Luca, I want to come back to you because Tim mentioned TCO and you and I have talked about your philosophy on TCO. So a number of things there. Can you scale two-phase? What about the supply chain? I've read all about condensation and pumps and everything else. How do you solve those problems?

Tim Shedd

>> So this is a great question. So I want to basically talk about the supply chain a lot because I hear multiple times that one of the reason why two-phase cooling is seen more challenging to deploy is because of the supply chain. And what we see at Seguente is really that if we keep things as that are in single-phase cooling, we can basically deploy two-phase cooling system as we do with the single-phase cooling system. At the end what is water-based cooling, it is a copper cold plate loops. Then they have a stainless steel manifold through a metallic manifold and then in the rows CDUs. So what we did at Seguente, actually we keep the same building blocks. The supply chain is diversified enough, so we have dual source or more than dual source for each individual components. So I don't see challenge in scaling because as they produce cold plate loops for single-phase, they can produce for two-phase and that is already in place QA and QC in place as well. So two-phase is I believe can scale and also there are significant advantage. I mean Tim, I respect him a lot. I think he's a great scientist and a great industry player, but imagine, right water base doesn't leak but it may leak. So 1% of the case, but what we see with the two-phase, that is also the advantage that we remove the worry from the customers about the leakage, right? Because if the refrigerant is going to leak, the refrigerant is a low boiling point fluid, so it's going to be vapor at the ambient condition. Plus also is a dielectric fluid. If it's a dielectric fluid and there is a leak, there is no harm to the electronics. So I see a lot of advantage and also I see something that I do like. I do have a PhD in phase change cooling, so I am a little bit biased. I like the scaling. The scaling is the key. If I can store the energy in the energy conversion from liquid to vapor, I have to do less engineering or less innovation to be able to cool a three kilowatts per chip. Now today in my lab we already have a passing two-phase cold plate, which is cooling three kilowatts is an ASICs and is simply by leveraging the latent heat instead of the sensible heat. So this is my thought and I would like to get back to you as well.

Dave Vellante

>> So my understanding was there were additional engineering requirements to accommodate two-phase, I'm inferring from what you just said that it's maybe not as onerous as I think, but I'll ask you and then see if these guys-

Tim Shedd

>> Yeah, so I don't believe it's onerous. We did a lot of R&D to be able to basically resolve the main challenge. But something that also I do like and I did not mention that when you use a refrigerant-based implementation that you can also have a choice in terms of type of metal you use. So you can go from copper to aluminum and if you go to aluminum of course is lower cost, is lower weight, and also you can recycle aluminum much easier than copper.

Dave Vellante

>> So I mean at the end of the day we know in this industry that the lowest cost, highest scale technology is going to win. But I know Dell as well, once you guys are locked and loaded on an engineering design, you got to have a strong business case to change it. So your thoughts on what Luca just shared?

Tim Shedd

>> We are open, just as Dion said, we're always open to new technology, new technology development. We want to be sure that we are providing the best value at scale for our customers. Right now, as I mentioned, that decision is single phase direct to chip, but we're always evaluating and we're keeping an eye on those things. And again, we have a great team that is capable of looking at it not just from the marketing documents but digging into the physics, making sure that it's going to provide the value that we believe it will at time and at scale. And that's really the key. It's easy to do things. And this is not saying anything about Seguente, I'm just saying in general it's easy to do things once on a test bench. It's hard to do things 10,000 times and repeat that over and over again in the data center.

Dion Harris

>> And I mean, I think the other thing that I think he alluded to, which is minimizing risk and especially as you start to deploy hundreds of thousands of GPUs, which is we are starting to see it becomes about how do you mitigate the risk? And then you talked about don't leak, that's the first thing you want to make sure doesn't happen. But at this scale you have to understand that there are going to be lots of risks introduced just by the sheer scale of the infrastructure. And so therefore you want to try to reduce as much as you can and get it down to known quantities as much as you can.

Dave Vellante

>> So Dion, you were mentioning ecosystem, emphasizing that earlier. How do you think about this problem when you think about something like two-phase, is that something that Nvidia would say, "Hey, we're going to do some core research on this," or would you say, "Hey, let the Seguentes of the world do that, we'll observe, partner, whatever." What's your posture on that?

Dion Harris

>> So Nvidia has definitely taken an active approach. So we were just recently in a project called Coolerchips, which is sponsored by the DOE. And so that involved exploring a lot of different direct cooling technologies and we have internal resources. I mentioned Wade and some of our internal teams who are actively looking at a lot of these technologies to really work alongside our partners. So even though like I said, we're enabling our ecosystem, we're active participants. So we're not literally saying, "You guys figure it out." We're trying to be as prescriptive as we can where we can, but we also know that there's value for them to add in terms of being able to build some of the systems as well.

Dave Vellante

>> And I wanted to ask you guys about immersion, because I hear a lot of different stories. I've talked to some of the immersion guys here. The fud is that the Nvidia's and AMD's and Intel's won't warranty yet, I'll say yet, is what I'm learning the immersion, but there are third parties who will, and you guys will probably get there. Do we have thoughts on immersion? Who wants to pick it up? Tim, let's start with you.

Tim Shedd

>> Sure. So Dell has been innovating in liquid cooling and immersion for a long time. To our knowledge, we're the first to deliver a commercially, a commercial system that was immersion cooled over almost 15 years ago now. So we have a lot of experience with it. And the reality is that over time we've come to realize that from a thermal perspective, there's not as much benefit as it first appears, and that's just due to physics and chemistry and thermodynamics. That doesn't mean that it's not a useful or important technology, it just means that we need to match the use case correctly for the immersion technology and what the customer needs. Again, scaling out to hundreds of thousands of GPUs in a data center, probably not immersion, but an edge deployment, a factory data center in a box deployment, that becomes kind of interesting. So definitely not ruling it out, but we have found that it's probably more useful in those kind of special use cases where the customer can really take advantage of all the benefits of immersion.

Dave Vellante

>> And look we're going to see a million GPU clusters, so it's coming. Thoughts on immersion? You want to?

Dave Vellante

>> Oh, yes, I will be happy to talk a little bit about, I just wanted to differentiate that you have two different approach. So a single phase immersion and two phase immersion. So keeping the same definition. I see immersion also as a viable approach, although there is a lot of, and we can see here a lot of study and exploration to see can we couple actually cold plate an immersion to basically extend the performance? Because Tim was mentioning in some of the use case, if the workload is very intense, you may have some challenge in terms of how you distribute the coolant because you don't have cold plate, so you may have some stratification and dry out challenge. So a mix, let's say between cold plate addressing, let's say the primary components, CPUs and GPUs, and then let's say a single phase immersion cooling for the secondary set components can also work, right? Because you're basically addressing the high power but you also capturing the heat from the secondary side components. I see some challenge also, I mean, I have to say scaling hundreds of thousands of pods in the way you service the pods and the way you deploy the pods, there is I believe a lot of work in consideration to be done to deploy immersion cooling at scale, right.

Dave Vellante

>> Okay. Last question. I'll ask each of you, I'll go around clockwise. So and Jensen describes this so well, I think we can all agree that we have entered the age of accelerated computing. Dion, it's in your title. We're not going back. General purpose computing, I won't say it's dead, but every workload is going to be accelerated. Can agree on that as well, and what Jensen described so well is with general purpose computing when new chips came out, Excel ran faster. Nice. It was great. But now we're looking at a completely different dynamic. New systems and new chips come out and I say systems because it is a system and new software comes out and it's accelerating much, much faster with new capabilities that we never even think of before. So as it pertains to the topic of direct liquid cooling, where do you see this going through the end of the decade? Is it pretty much where we are today, it's just got to be built out and scaled or do you expect that by the end of the decade, Tim, we're going to see dramatic new introductions of architecture in the data center that we're going to have to respond to? What are your thoughts on that?

Tim Shedd

>> So yeah, at the data center level, I think it's really important to keep in mind I mean we're living the excitement of accelerators and HPC here, but by 2030, most of our sales at Dell are still going to be air-cooled. Most data centers are going to be air-cooled. We have to maintain focus on innovation from air to liquid and enabling the customer where they're at right. Now to your question, will we see new technologies and new architectures? I absolutely believe that as we start planning on half megawatt racks, it's a different architecture. That's a different way of thinking about building the data center. And we're working at it at Dell, we know our partners at Nvidia are working on that. We know that Seguente is also looking at innovative ways to implement that. And that's what I think we're just at the beginning of seeing that. I know we're working on thinking about how do we at Dell meet our customer's needs with a complete solution that we can deliver rapid time to deployment meeting our customer's needs where they're at, whether it's air-cooled or liquid-cooled working with Nvidia where we need to say, "Hey, we really need an air-cooled skew here. Can you help us out?" They do because they want to sell GPUs that we want to sell and then we want to enable those for our customers who for whatever reason, they're just not going to go liquid-coll. So we'll be there. We're going to see a really interesting set of innovations here in the next five to 10 years.

Dave Vellante

>> Dion, give us your take.

Dion Harris

>> It feels almost impossible to look out that far because of how quickly things are changing. When I think about if I go five years back, would I have predicted we would be where we are today? And I would absolutely say no. But what I will say is I think in broader strokes, I agree, we will see a spectrum. I think you will still see some workloads, some data centers that are still largely air-cooled and using what I call mainstream parts and just given the trend and the efficiency that we're able to drive by increasing the overall compute density, I think that trend will continue as well. Especially as you start to see these scaling loss take into effect around both training and inference, driving more performance, more compute density. And so therefore, I think that trend will continue where we'll be in five to six years, I don't know. But I do think you will still see a continuum of both air to liquid to very densely populated systems.

Dave Vellante

>> And I think about too, the pace of technology. We always say, "Oh, the pace of technology is accelerating," and it is coming out of the factory thanks to the guys like you, but the pace of adoption is not so easy. We talk about agents and agentic. My view it's going to take a good five to 10 years to actually really build that out. And so applying this technology sometimes is a challenge for customers. They've got a business to run, but Luca, I want you to make the case for two-phase and why you think it will go mainstream. I'll give you the last word.

Dave Vellante

>> Sure. So just at a high level, I want to mention, because when you discuss about the new data center architecture, this is let's say a liquid cooling session or panel, but there will be a lot of challenge also on the power side. So this is something that we did not mention, but we have to cool the chips. But first of all, we need to power those. And this is also something that we need to think about. As well as also I want to say that as we increase the power of those racks, there is something to keep in mind, a new innovation they need to come from the sustainability standpoint, how we actually can capture this heat to be able to reuse. Because if we can capture the heat at a high temperature, then we will be very sustainable because we get directly used. So just stepping back to your question about two-phase cooling for us is not an R&D exercise. So we feel that we have completed the R&D phase. We actually, currently we are shipping product and we actually, we have the strategic partnership in place on the back end with the tier one server vendors to be able to provide our SKUs, which we normally we retrofit, we rack and stack and we ship as well as also we have strong partnership on the front end to make sure that we have a come infrastructure provider company which they can help pass to deploy and also do all the SLA and technical services. For me, I like two-phase. I believe there will be a transition also the transition by the fact that with the water-based cooling, if the heat load is going to increase, the only way to basically dissipate that amount of heat is to basically pump more water. But with the two-phase approach, basically you can evaporate more liquid without vapor. So you can keep the mass flow rate lower. If you can keep the mass flow rate lower, then you can start to basically keep lower size of tubes. You can keep the dimension lower. And also something that I do like is that if there are multiple water-based cold racks, right there are many thoughts about, "Okay, can I deploy cold water cooled to phase cold racks, right? The answer is yes, because with our implementation that we have a top-of-the-rack CDUs, which is dielectric on one side and single-phase water on the other side, any incoming flow rate and temperature from the floor-standing CDUs, our technology can take it. So there is no variation in infrastructure at all. So you can have many water-based, many cold-water-based, and then you keep scaling the data centers. But I want to make a final statement here. I enjoy this session a lot. I believe that everyone will play an important role. So let's say from the chip makers all the way to the company, they provide services because we need all of them to make sure that as the power goes up there is technology innovation across the board to make sure that we can make our customers happy. So I want to say that-

Dave Vellante

>> So you and Tim are friendly. There was a little bit of red meat in there. You want to respond? Okay. Good.

Tim Shedd

>> No. As I said before, we are continually valuing technologies, keeping our eyes out. We talk and we know what's going on and we just want to make sure that we're providing the best value to our customers. And if that's Seguente, we'll take a look at it.

Dave Vellante

>> It's an amazing thing about our industry is innovation. You guys each play a very important role. Thanks so much for coming on the panel. Tim, Dion, Luca, thanks. Really a pleasure. And thank you for watching. This wraps up, day one here at SC24. My name is Dave Vellante. Savannah Peterson and John Furrier here. We kick off tomorrow at 10:00 A.M. East Coast time. Go to theCube.net, check out all the videos, siliconangle.com for all the news and analysis and Thecuberesearch.com for the deep analysis. Don't forget to go to Thecubeai.com, ask it a question, it'll give you an answer. We'll see you tomorrow. Thanks for watching.

Tim Shedd

>> Thanks.