Name: Designing and Operating the AI Factory: Continuous Performance Optimization at Scale
Uploaded: 2026-05-19T17:30:20.566Z
Duration: 1 h 9 min 48 s
Description: Designing and Operating the AI Factory: Continuous Performance Optimization at Scale

Transcript for "Designing and Operating the AI Factory: Continuous Performance Optimization at Scale": Welcome, everyone, to today's webinar. We will begin in a few minutes. Welcome everyone to today's webinar. We will begin in a few minutes. On behalf of Cadence, I would like to welcome you to today's webinar, designing and operating the AI factory, continuous performance optimization at scale. My name is Cale Hawkins. I'm the field marketing manager here at Cadence. Before I introduce today's presenters, I'd like to go over a few housekeeping items. All attendee lines are muted to prevent background noise. There will be a q and a segment at the conclusion of today's webinar. If you have any questions during the presentation, you may type it in the ask a question section at the bottom right hand side of your screen. We'll attempt to answer all questions at the end of the webinar. However, if we run out of time, questions may be answered via email within the next few days. Please ask questions early so that we can get them queued up and ready to go for the q and a session. Today's webinar is being recorded, and a link to the recording will be emailed to you in the coming days. Final note, we will be controlling the movement of the slides. Some slides include graphics and may take some time to advance depending on the speed of your connection. At this time, I'd like to turn the event over to the team, Sherman Ikemoto and Mark Fenton. Sherman, the floor is yours. Thank you, Kale. Good afternoon. Good morning, everyone. I'm Sherman Ikimoto. I'll be, tag teaming with Mark Fenton today to talk about this, webinar. I wanna make sure, just can everybody hear me okay? Can you give me a a signal? Yeah. I can. hear you fine, Okay. Great. Great. Alright. Just a quick slide for disclaimers. We'll leave this up for a second, and then we can advance to the next slide. So, again, I'm Sherman Ichimoto, sales, development group director for Cadence for the Reality Digital Twin product line. And with me today is Mark Fenton. He's our product engineering group director. I'm located in San Jose, California. Mark is in, London, and, and we're happy to bring this, webcast to you today. So this is the agenda, four main sections. We'll talk a little bit about cadence, then we'll go into the AI factory design and operating challenges, and and then talk about how those challenges could be tackled, using the Cadence Reality digital twin platform and also look at some AI factory design and optimization case studies towards the end. So just to get us started, Cadence is a market leader in AI and digital twins. We pioneer the application of computational software to accelerate innovation in the engineering design of silicons and also systems. Our design solutions, based on Cadence's intelligent system design strategy, are essential to the world's leading semiconductor and systems companies to build their next generation products from chips to full electromechanical systems. And we serve a wide variety of markets, including hyperscale computing, mobile computing, and communications, automotive, aerospace, industrial, life sciences, and robotics. Cadence is a global company with over 14,000 employees in 40 different development centers. We invest over 32% of our revenue in r and d to fuel innovation. Our 2025 revenue reached 5,300,000,000, and this is a continuing trend of consistent growth. We're recognized as the top global company by the world's best publications for various achievements in technology, business, and management, and we offer a comprehensive AI driven design solution for electronic systems, sciences, and applications. So our customers include semiconductor companies that design and manufacture every type of device and systems companies that design and manufacture products that contain these semiconductors. And at Cadence, we view our solutions as a three layer stack that underpins all of our products. We integrate accelerated computing hardware, AI and physics based simulation and optimization software into a unified offering. So each layer can be used independently. For example, you can run our software on your own compute platform. But when they're taken together, they deliver the highest levels of performance and productivity. So with that, let's get into the AI factory design and operation. So let me start, with this quote from Jensen Huang at Davos. He noted that we are now facing the largest infrastructure build out in human history, and the numbers back this up. This is, the trend for global or projection for global data center capacity, over the next decade. And we're projecting to go from 82 gigawatts today to 219 gigawatts by 2030. This is a staggering number. And the light blue bar, which is the AI workload portion, that's what's driving almost all of this growth. But this these numbers don't capture just how fast and how in the scale of this build out. If you compare the AI factory infrastructure build out to two other transformational build outs that we've experienced, aviation and automotive. Aviation took a hundred and twelve years. So to to reach, about, you know, 4 to $8,000,000,000,000 in total, infrastructure CapEx cost. And in automotive, it took a hundred and twenty five years to reach, 12 to 22,000,000,000,000 in cumulative CapEx. But AI factories are projected to deploy up to $50,000,000,000,000 in just twenty years. So the scale is bigger and the time frame is much, much shorter. Again, staggering numbers, and, it just underscores what Jensen Wong, said at Davos, earlier this year. So let me let me start also by, describing, at a high level, the AI factory business model. This is a business model for a one gigawatt AI factory. These numbers are representative. These numbers will change based on, specific cases, but it's it's it's good to illustrate the, the the business model. And, again, this is an idealized case. So, we've heard, numbers for total cost of ownership for a one gigawatt AI factory anywhere from 30 to maybe $82,000,000,000. That's total cost of ownership, CapEx and OpEx for, in OpEx for about eight years. And you can see the breakdown in the box. And what this gets you is a one just about a one gigawatt data, AI factory that is producing tokens continuously for about eight years and one upgrade cycle within that eight year time frame. And roughly speaking, based on quoted, revenue production numbers, you can generate about a $127,000,000,000 of revenue from that one gigawatt AI factory. And after factoring in, OPEX costs, the EBITDA output is about $87,000,000,000 So that's a it's it's a very good, obviously, a very good, business case, 69% EBITDA margin. And the this is really the driver behind the the scaling up of AI factories, the scaling up of the market. Now, here's a key insight though. At this scale, every percentage point of token throughput is worth about $1,300,000,000 in revenue over the six years. And every percentage of PUE improvement is worth about $60,000,000 in OPEX savings. So optimization is is not a nice to have. It's a must have, and it determines whether the AI factory will clear its cost of capital. Now remember this this is an idealized calculation. The numbers assume everything is operating perfectly end to end. But in the real world, that's not what happens. And so let me talk about that next. So in every AI factory, in every data center, there's a hidden cost that doesn't show up on on the spreadsheet. We we call it the stack tax. So an AI factory and all data centers actually are are built in parts. The chip is the baseline. And from there, you have the PCB that is designed for the chip, the compute tray designed for the PCB, the rack designed for the tray, and so on to the pod and all the way up to the facility. And each of these layers are designed by different teams using different tools, sometimes designed by even different companies. And every team faces the same challenge. They don't know exactly what layer above what the layer above is gonna ask of them. Protect against their their this unknown by adding a margin of safety. And those margins don't just add, they they compound through the stack. So as a result, the facility at the end of the stack costs much more to build than it should and delivers less throughput than it could. And this overdesign is built into every level. So the costs add up, and no single team can see the full picture today until the the tax has to be paid. So how big is this tax? Let's quantify it. So in this calculation, we're gonna assume that every layer above the chip or, actually, below the chip in the stack overbuilds by 35%. And this is not a wild number. It it's roughly what we've seen consistently over the years, when you account for the safety margins that each team builds into their layer. So those margins, they compound across the five layers, and you get at the end of the stack at the bottom of the stack, you get a four and a half times overdesign factor at the facility level. Staggering. And this amount of compounding, this amount of overbuilding, leaves you with only 22% of the built capacity that can actually deliver, useful throughput, and the rest is stranded. So on an eight year financial basis, that turns our $87,000,000,000 profit engine into a $12,000,000,000 loss. So the factory doesn't just underperform, it actually loses money. Now watch what happens as we close the integration gaps between the layers. So now we're designing each part against the actual integration requirements, that they must meet. They're not generic or worst case requirements. And the comp the compounding factor starts to drop. So as we continue to right size each layer for the next, we cross into profit and and even beyond. So every layer we integrate has two effects. There's less overbuilt hardware, that means lower CapEx, and also a smaller compounding factor, which means more of our installed capacity will actually deliver. So after fully integrating up to the pod, so we're stopping at the pod here, we're now at an EBITDA of roughly about $40,000,000,000 for this one gigawatt of AI factory and a compounding factor down to 1.4 six four. So we we've recovered nearly $50,000,000,000 of value by closing the integration gap between all of the layers in the IT stack. But we are not done. There's one more layer, and a very significant one. It's the layer between the IT stack and the, the facility itself. And that's where the $23,000,000,000 opportunity exists that we're gonna talk about today. So, historically yeah. So so here's the here's the need now. The need is we need to optimize across the entire stack. So on the left, you see, a problem that the industry has largely solved. For example, NVIDIA, is, has adopted extreme codesign, capabilities within their company to tightly integrate the IT stack from chip to pod. So all of the margins are understood, and they're engineered out. And tokens per watt is optimized throughout that part of the stack. But when you reach over to the facility, that part still today is unsolved. So the facilities are still designed for worst case design assumptions. The margin stack is unseen, unquantified, and the tokens per watt are left stranded. And, again, this is where the $23,000,000,000 opportunity exists. So what are the challenges in closing this gap? Well, the first is, an obvious one. IT reinvents itself every year, and the facility is designed to operate for twenty years or even longer. So a new GPU generation comes out every twelve months now. We can see the trend on the, on the chart. And down at the facility facility level, once it's designed, poured into concrete, wired, you know, power distributed, coup cooling is is deployed. That facility is supposed to last for decades. So as the IT site keeps moving, the facility can't easily, and, this creates a tension between the those two layers of the stack, the IT and the facility. So this is a life cycle gap that is inherent in the industry. Can't avoid it. We just have to be able to deal with it in the best possible way. A second challenge is that every data center has has multiple mandates. There's one system, but it has multiple mandates that conflict with each other. For example, a a CFO of an organization running an AI factory will ask, what is my capital generating in returns? So they wanna measure token throughput. CTO is asking, is this system running at all times when the AI workloads demand it? So they're concerned with availability and uptime. And then the people running the the data center or in the data center, they need to adapt quickly to IT requirement changes. They need time to token to shrink to down to zero or as close as possible to get, IT systems, up and running quickly. So these are three legitimate questions, three different metrics, and pulling in opposite directions. So and and then here's the problem. So in today's tools, these three metrics, they trade off against each other and, and conflict. Every move you make will trade one mandate for another. So what's needed, what's missing is to be able to see all three at once, to see the impact of one decision on the other two or or one objective on the other two so that every trade off becomes visible before you commit it. And this is missing today. So this is, a a a big gap that has to be filled. And then the final one is that, we have a, again, a fragmented ecosystem with no bridge to to to to build them or or to bridge them. My computer is stuck or lagging at the moment. Yeah. Here we go. Okay. We have a a fragmented ecosystem with no bridge between them. We have IT manufacturers developing, systems independently of facility operators who are developing their, facilities independently of power and infrastructure manufacturers. And there's no effective way to consolidate these, these products into a cohesive data center, into a cohesive date data center system. And in addition, today's tools, lack something very critical. They measure what's happening today, but they cannot predict the future. And that's what we're seeing on the screen here. So what operators will see today are things like inventory, real time monitoring, historical trends. But what they need to know in addition to what they see today are what will happen in the future. What will my workload do to my my token output? Will I will I experience availability problems with this next workload deployment? So in effect, the operators cannot see the future. You cannot optimize what you can't predict. So here are the gaps, life cycle gap, ecosystem gap, physics gap. Can't see what what you can't predict. And the impacts, again, are on throughput, agility, and availability. And given the lack of visibility of these three objectives, what will happen in most cases is that operators will have to sacrifice to a large extent computational output and agility for availability. That's because availability availability cannot be sacrificed. So the other two are are sacrificed. And this, illustrates the level of maturity for data center design and management that is in the industry today. This is a maturity model that was written by the Uptime Institute a few years ago, which describes where the industry is today and where it needs to head. And, if you read this report, you'll be able to see that the industry is at about level three where, DSM tools are are, take the industry from level one to level three. But what the industry needs to do is to progress up to level five where you can start predicting the future so that you can optimize your your performance and then eventually through AI, self optimize. And that's where Cadence Reality digital twin platform will take us. So let me just go through what makes Reality Digital Twin different. It is a physics based and AI accelerated solution. So built into the platform is the ability to predict the the physical behavior of the data center. And from that, you can get, a a visibility of token throughput, availability, all from a simulation, and and then that gives you the ability to optimize. Second, digital twin is continuously synchronized with the real world. So every change that occurs in the physical twin is reflected automatically and instantly in the digital twin. So the the digital twin model never drifts from reality. Third, it's a multiscale integration solution. So the the model itself represents the data center from chip to chiller all in a single model. This allows operators to optimize to the highest level of granularity possible to get the most out of their AI factories in terms of token throughput and also to understand the margin on availability as accurately as as possible. And then finally, the fourth piece is within the platform is the concept of DC elements. So this is a library of of, components. It's actually a reflection of the market, the the entire AI factory and data center market. Any product that you can buy and put into, a data center, you can find a behavioral model represent representation in the DC elements database. It's all there, and these are the building blocks of digital twin models. So just another, little bit more details on the DC elements. There's over 15,000 different models now. These represent the make and model and behavioral characteristics of IT servers, CDUs, air handling units, power distribution units. They're organized by vendor, so you can search for a specific computer model from a specific vendor and find it in the, library and then build your digital twins from those building block elements. My computer is lagging again. Give it a second to catch up. I think I need to to restart this. Apologize for that. Okay. And, the final thing I wanted to describe, for Cadence Reality Digital Twin, it's a platform. So it's a a communication or a shared model that connects the ecosystem to the operator, and it's made up of multiple tools. So what you see on the screen are the the the two tools in the platform that are used for designing the components of the data center. These are IT systems, power units, cooling, and facilities themselves. And the output from these tools are behavioral models that go into the DC elements library. And then operators draw from that library to build their digital twin models, create AI surrogates of those, and then use the models to plan and optimize their AI factory. I apologize. My computer is is lagging. So let me, that was actually the, the last slide. I I I can stop here, and let me stop sharing and turn it over to Mark Fenton. We'll take you through the Cadence Reality digital twin platform in more detail and some, design and optimization examples. Mark? Hi there. Okay. So let me share my screen. Hopefully, you can hear me, and, hopefully, I have less sharing problems than Sherman. Yeah. We can hear you. got. it? Alright. Brilliant. Okay. Well, this is me, and, thank you very much for everyone who's joining us. My name is Mark Fenton. I'm the product engineering group director for the Cadence Reality software. I've been working in the space, for the in fact, my entire career, twenty three years going now. So it's been a real privilege to see how data centers have changed over that time and how we've been able to deliver digital twins for the data center market. And, really, I thought this is a nice opportunity to talk to you about, how what we're building is really helping with AI factory design and optimization. So I've got some some nice case studies that we can go over, but also wanted to show you some of the features that we're developing in the software that is is kind of taking our offering beyond the traditional, simulation technology that maybe you would have recognized in the past and really how things are changing to adapt, to to meet the the current, requirements. So I thought this is a nice time to to to take you through some of those. The first thing I wanna touch on is just some of the the metrics that we're getting from our customers. So these are all numbers that they have provided for us. These are not our marketing, data points. These are all in collaboration with some of our biggest, customers that are using the the reality software. And they we have a second slide on some of the other verticals, but in terms of, looking at enterprise hyperscalers and AI factories, these are some of the the the benefits that you're gonna be getting out of digital twin. And I I just wanna touch on a couple of them. I won't go through all of them. But the first one I think is interesting. One of our largest enterprise customers, and and and Sherman sort of talks about these three mandates where people have got a different lens that they view the data center through and what they, are trying to improve. And, actually, we were brought in, to look at energy. So this 15% reduction in PUE is what they achieved with using digital twin. And they were, you know, really happy with that in terms of the saving, but, actually, what we, uncovered was was a much larger, saving from a CapEx and capacity standpoint. And so, actually, the the the energy piece in terms of a dollar value was was was actually quite small compared to, what they were able to to, improve in terms of the amount of capacity that they could, put into this this hall safely. And I think that's with you know, we're hearing a lot about today the challenges of building new data centers. The time the time takes to actually acquire power, mostly power, and land and get planning is is years in the making. We also hear about how oversubscribed colocation and how difficult it is to get lease, space. So people are looking at what they have, you know, on the books now, and can you make improvements with with your own data centers before looking elsewhere. So, we're seeing this message from enterprises a lot that they're saying, okay. How do we optimize what we have? Because it's getting harder and harder to to build and and harder and harder to lease. So some some good numbers that we got from from our enterprise customers there. When we take kind of take a look at it at it from a hyperscaler lens, and, Sherman kind of pointed about the the fact that our digital twins are operationally always ready, they're always tuned, and, they are syncing live data. It means that you can make those decisions really, really quickly. So rather than having a offline engineering model, where you have to find all the data to make a decision, you've got many different tools that are all, disparate, many different teams that are not talking. Bringing all that information into a single digital twin really does speed up making those difficult decisions about next deployment. And and this data point that they gave us was just the the difference between, having to do that operation manually, versus just opening the digital twin, which has all that information, all of that good stuff automatically consumed and populated. So they have no work to do apart from saying, well, what do we wanna try is our next deployment? So they gave us that, nice stat of of a speed up of a one two eight x. And then, I'll I'll dig into this AI factory one in a little bit more, but because this is, the focus of today's talk, I'll I'll I'll spend a bit more time on that, particular example a bit later on. And then on the second slide, I just wanted to just kinda touch on you know, we do a lot with collocations. There was a question in the chat, already about how how that's an extra challenge because these lease tools have multi tenants quite often, and they all have their own vision for how hardware, will be configured and laid out. And so quite often, the collocation provider is looking at it from a from an energy standpoint and a an SLA standpoint, but the users themselves are looking at it from a from what does my hardware look like? And, actually, this is where digital twin offers really immense visibility and and the connection between those two groups. So our colocation, users are actually using it as a a way to connect to their customers and offer almost this home away from home where the customers can log in to a web portal, see the the layouts of their hardware, make decisions about their capacity, and then the colocation providers get those plans ahead of time and can can, investigate what that looks like from a performance standpoint from the data centers. So, colocation is a big big users of our tools. And then the other one I just wanted to point out is that we now are, used globally by the 10 of the top 10 m and e firms within data centers. So, you know, if if you're talking to a firm about using our technologies for design, then if they're in the top 10, they'll be using our our, reality digital twin. Here is a a couple of quotes with some QR codes that kinda link you to some other, materials if you wanna read a little bit more about some customer success stories across the board. So I will move on, but, there's lots of good reading, good materials to to have a look at how our customers are using us, globally. So let's talk about this, optimization that we did, with NVIDIA. So we were looking particularly at AI factories, and, I'll animate this video here so you can see, the digital twin model played in the digital twin experience, which is powered by the NVIDIA Omniverse. So we launched this in April, so it's just, on the market, and it's a real high fidelity visualization of the data center with all the simulation. I've got a little bit more information about that later on, but just wanted to show you what it looked like for a for a real site. But but here and and, Sherman pointed to this in in his opening piece about, you know, leaving all that headroom on the table really does leave a lot of, the revenue that you can go after, and so optimization becomes more and more critical. So what what we were looking at here is rather than running all these GPUs at the maximum power, that actually there's a lot of data out there, that shows that if you, take those, GPUs and actually run them at slightly lower power, they still have phenomenal performance in terms of the the workload. So, for this is an example training set that we had, for GPT three that you still get 98% performance at, 79% GPU power. So with that in mind, what we can actually do is if we reduce the power going to the GPUs, we can operate more of them for the same power footprint and actually then generate more tokens. And so when we're then looking at that as an optimization within the space, where is the best place to lay these racks out, but we now move beyond just physical layout into where is the workload best to be positioned, which which racks can hold more than let the workout the the the workloads, which cooling, units are able to handle the additional load, whether it's a CDU for liquid cooling or the cry units for the air side. So the optimization then allowed us to, put in an extra 17% more tokens per watt. So for the same watt, 17% more tokens through the through the AI factory, which when we scaled up to gigawatts equated to about $5.05 and a half billion dollars, per gigawatt. So really a huge savings when you're looking at optimizing a space for for the workloads. And just a little note on kind of how the optimization works is that you're always looking to define what's successful for optimization studies. So, in this case, we're trying to put as much GPU online with as many workloads as possible, but we also very strictly at this point said we don't want any any tokens at risk from a thermal, failure. So those are the the kind of inputs to the optimization study that we say. Try and fill as much and try and push as many tokens out of this site. But as soon as, we have any thermal risk, the optimization learns about that and understands, okay. I pushed it too far in this part of the of the data center. Let's move the workload to a different part. And by running, I think it was about 60 or 70 scenarios, we were able to very quickly narrow down the the the optimal way to run this data center, and that's where that, optimization came from. So you can see this kind of set of bar charts is actually a whole load of different distributions for the workload, a whole load of different number of GPUs. We can also tailor the, the CDU supply temperatures from the water. I think you may have heard Jensen's talking about 45 degree water, at his keynote at GTC. So all of those variables you can now play with in terms of what do we wanna find from an optimized space. And so, that's, kind of where we we settled on with with NVIDIA. So just really quickly, just a a high level view over the product. So there's two parts to it. There's the design part on the left hand side. And then when we talk about, the operational platform, that's our insight tool where we connect out to existing data, so whether there's telemetry, whether there's DSIM, whether there's other ITSM tools, for example. We can connect those directly into the operational digital twin so that it's always on and always ready. So there's two two parts to the to the platform depending on whether you're an M and E consultant that's doing pure design or whether you're operating a data center and you wanna have a a live operational digital twin. There's something that's that's there for everybody. And as I sort of said at the top of my section, we've been doing this since about 2006, so selling the software into the data center space. So it's been, you know, it's built on many, many years of experience dealing with these sites. Okay. I'm just doing a quick time check. So I can see, probably have another five, ten minutes to go, so so we're looking good. Just a couple of the points of of how the platform, can help you with the sorts of things it can look at. So we've touched on optimizing the design. I think quite often we get looked at as just a pure air side simulation, but we've we've moved, very, heavily into liquid cooling solutions as well. So we can look at the distribution through the chilled water system into CDUs and then the CDU doing the coolant up to the liquid cooled IT. So we can now evaluate that entire end to end system from outside the white space into the inside so that we can call the chip to campus level evaluation when we're looking at this very big scale problem. We can only handle those problems when we start looking at GPU acceleration, and I'll touch on that as well a little bit later in the in the deck, and how we're using AI particularly to drive design optimizations. So this is just a high level view of of of how we handle this system wide problem or system wide design and optimization. So we're now able to model an entire campus, so the outside of the entire campus, and that can be many, many buildings. And those buildings are generally connected together from a power perspective and also a cooling perspective, which we can now handle within, the digital twins. So in this middle portion, you can see this is a liquid network, which we can now simulate the flow of cold water maybe from an external chiller yard going through to, a set of CDUs. And then from the CDUs, the the TCS, the the technical cooling supply going to our, hardware. And all of that can be modeled, all the controls, the pumps, the valves, everything can be put into that system. We can we can do transients, so we can do failure of components quite often. That's an important consideration. Do I need a buffer tank? How big does it have to be? What temperature should I run the water at to give myself ten minutes of failure? You know, those those kind of what if scenario questions that people are interested, which always have a sort of time element. And then that that can, again, have multiple data white space data halls connected to it, and we can even go right down it to, surrogate models of the IT, so looking at, temperatures of the chips. And that can all I mean, it is quite amazing. This was never something that we we dreamed we'd be able to do. But with current computing, we can actually now build an entire system model, which goes literally from chips all the way out to the chillers, and the whole thing can cosimulate in one go to give you a full system wide performance calculation. And that, I think, is is extremely unique to Cadence and really valuable. So when our customers realize what they can actually look at, they answer a lot of those what if questions in design and also in operation. Now I wanna talk a little bit about AI and how we're using it inside. So Shannon talked about design for AI and AI for design. So this is a look at how we're using it internally. Now if you've ever used simulation technology in the past, you'll know that there's this waiting element that you build the model or maybe you ingest some data from DSIM. The model is ready. You wanna try some scenario out. You hit solve, and then you have got a wait, and it's a waiting game. And depending on your size of your models or the hardware you're running, that could be an hour solve. It could be a four hour solve. It could be a twenty four hour solve, and that, you know, adds up. And and so what we've what we've been looking at with AI is what if we flip that problem on its head and for a for a specific site, what if we actually do all the simulations upfront, and train a surrogate to then understand how the data center will perform? And so this is an example, of of what we've been building where we've trained a model, with with about I think it was about 500 simulations of different layouts, different configurations of hardware, different cooling controls. So all of those things are fed into the training. And then once we've trained this AI surrogate, we're able to make almost instant inference, results of what it looks like when you make a deployment. So let me just run through this example. So here in the web interface, what we're doing is we're looking for a a GB 200. I'll try and hit the pause next time. So we're looking for a GB 200. We pull that into the, into the menu as if I wanna deploy this. And what it's done now is it's gone that pause isn't working. It's irritating. I might be able to do it here. Okay. So, what it's done is it's given us every single row in every single data center that has capacity for that particular, GB 200 rack, and it may be multiple racks that you're looking for. Now this is how we traditionally would have worked, that we would have offered you visibility in where your available capacity is. This is how most decent tools work. They just look at the hard numbers and say, yeah. Okay. This rack needs a 120 kilowatts of power. It needs this much cooling. It needs this much space. And so we surface all of those up. But, actually, if you're trying to make a decision about the best place to put something, that is an a lot of options that you have to kind of think about and churn through. So what we've done now is if I, hit play, hopefully, it'll carry on. What we've done now is added this capacity agent. And what's happening here is that the agent is searching through all of those options, and it's inferring the results almost instantly. So it's come up with those three best options. And And every time I select one, you'll see that the rack is actually moving around on the floor print, but it's giving performance already. So you can see it's already got temperatures that are inferred from the simulation. So I don't have to wait four hours or ten hours or twelve hours. It's giving me the performance straight away. Now if I decide, actually, what I wanna do is take my best three and run a full traditional simulation, I can go ahead and send that off. And that then gives me rather than those quick inference results, which is a quick look at what it might look like, What it's done now is is it's run the full simulation, and you get those three reports side by side to be able to dig into a little bit more detail. Okay. Actually, this is what I get from the inference. It's just a temperature. But what I really wanna do is dig into a side by side of the tables of the data or the full simulation results that we generally can see when we do a full, SIM. So the way that is helping our customers is they get that quick especially operational, they don't have time to run ten, twenty different scenarios if they're four hours each, but they can definitely look at these in the inference as the best, first guess, and then they pick maybe one or two. And it so really is changing the way that people are using CFD, operationally. And and, again, the second kind of piece to that is how are we, able to start moving things to be faster and, moving from CPUs across to doing work on the GPU is actually a massive, massive, accelerator for us. So this is just an example of what's coming. We aim to be, making this available by the end of the year that everybody be able to run on GPUs. And in terms of just a traditional 16 core run for a a 250 rack, two megawatt data center, which is about an eight hour solve on a 16 core CPU, running that on a b 200 was about fifteen minutes. So it's a significant speed up. And, obviously, depending on what hardware you're using, there'll be different speed ups, but we're seeing, like, 20 x on even a a a very sim sort of low end GPU. And then, obviously, b 200 is quite high end, and we're seeing sort of forty forty x, speed up on those, runs. And what's what that really allows us to do is when I so when we talked earlier about this this pretraining of models, well, some of you might say, okay. Well, if I have to run 500 variants and each one takes eight hours, that's still quite a lot of compute. But when we couple the two things together, so when we couple the fact that we've got AI surrogates that are running on GPUs, suddenly the training of a of a data center digital twin may only take twenty four hours, and then you have the full surrogate ready to go. So, really, that's unlocking not only, the AI surrogates, but also this optimization piece that I talked about at the beginning where I ran fifty, sixty different scenarios. Again, if you're doing 60 scenarios and they're ten hours a pop, that's quite a significant bit of work. But if they're ten minutes ago, then actually we can start to look at really exciting optimizations where we never really dreamed of doing it before because it was too compute intensive. But now the fact that we can run on GPUs, it really unlocks all of those what if scenarios that we wanted to try, but we never really had the bandwidth to be able to run them. So I think that's pretty much everything I really wanted to touch on. I think the last thing I wanna just show, and I really did kinda touch on it, is the piece that we're doing with the NVIDIA Omniverse. So, we've built that into the insight platform or operational tool where we're able to take the data center and visualize it on the on the web in in high fidelity view. So you're you're you're looking at it and it and it is pretty photorealistic. It's it's like you're walking around the real data center. We can still push all the simulation results on there so you can have a look at the airflows, the liquid cooling, the IT assets. But what's, particularly interesting about, Omniverse is under the hood is this, Open USD file format, which is built really for multi user collaboration, simulation, and with, a lot of the, the libraries that'll be coming out of this, the SIM ready libraries that are part of the ecosystem for DSX. What we're seeing is that the partners, the OEMs that are part of the ecosystem are starting to provide those SIM ready assets that you can consume that have all of the data available to be able to run these simulations. And so you get almost two views of this asset. You get this incredible visualization that you can look at to see how all the pieces connect. Then also, what does it look like when we're simulating that? What's the impact on the performance of the data center? So really an exciting time for us to be part of this, NVIDIA e ecosystem and part of the DCX, blueprint. And, I just wanted to touch on that, that, in case you've seen it, then then you know, what the DT experience is all about. Okay. So, I think we're at the the end of the the show. Just giving you kind of a summary of what we've talked about, Shem, I mean, laid out the business case, really clearly. The numbers involved are are really significant, and we've started to see the market not talking about data centers as a as a cost, but looking at the opportunities for revenue. And when those thin margins, those fine margins take it from being a profit center to to being a loss, you really want to be thinking about optimization in the design and then optimization in operations when it comes on-site. And, really, the digital twin helps you understand that true performance, whether you're predicting the the risk or the cost profile, trying to optimize us for capacity, or whether you've got that lens of energy efficiency and sustainability. All of those questions can be answered within the digital twin. So with that, thank you for your attention, and, I'll turn it over to any questions. Okay. Hey. Thank you, Mark and Sherman. That was fantastic. We will go ahead and take some questions, now. Some of you have been asking them as the presentation has been going on. So as a reminder, you can type your questions into your q and a panel by clicking ask a question in the lower right hand side. Some of the questions have already been asked. So, Sherman and Mark, if you'd like to address some of the ones that are already asked, you can take it away. Yeah. Just having a quick scan down. I don't know, Shammin, if you had a chance to look at anything. Yeah. There's, there's one question. The the AI surrogate modeling capability, is it also available in design in the design tool? Okay. So, yeah, the, what what I showed with the, the capacity agent that's going and making a deployment, that's currently only available in the digital twin platform. The the reason for that really is that, the the online capability connected to a GPU inferencing, it seems to match generally our operations users that are, interested in that quick. Tell me the best place. I've got lots of different places that I could put this workload or this piece of IT or this rack. And so, generally, we we felt that that was more suited to operations. We are using AI, surrogates in different ways in the design tools. So one of those one of those places that we that we're using, is is around the actual models of the IT themselves. So, I think in one of the images I showed, let me just jump up here. So so we have a sister tool that's called Celsius EC Solver that does the, the thermal simulation for the for the hardware design. So that goes right down into the details of the chips, the board, the PCBs. And, realistically, if we were gonna build one of those models and then, load that into a data center model and put thousands of them in, the size of the problem, it gets really gets too too large. So this is where we we actually do the same process that I showed with the AI, surrogate for the data center, but we generate an AI surrogate for the server, and then we deploy that in the data hall. And it gives us even when we're simulating, it still gives us an inferenced result for inside the box. So, actually, if you're interested in GPU temperatures or CPU temperatures, even though we haven't got the the actual, granularity of the components in the model, we do have the fact that we understand exactly how it's gonna behave because we've got this inferred, AI surrogate. So we are using AI in design just slightly differently, but it but it may end up that it that it kind of moves across the design too. Great. Thanks, Mark. Oh, there's another question. The Xeon comparison times, it may not be a fair compare to compare them to a b 200. That's the question. Isn't Xeon from 2012? Is there a. yeah. Yeah. It's ahead. a fact that I think we're giving away what our engineering teams have got access to. But, yeah, it's it's it's probably, it's probably, not not quite a fair comparison. It's just in terms of what we what we've what we've been using traditionally, for doing these runs. But we we we have got some more modern CPUs that we've that we've been running on, and I I can just tell you that the the the comparison, even if it's ten ten or 20 x, this is a game changer. It really it it really is. It's, it's it's gonna change. I I think there's a there was a question around, whether we think people will do less simulation because you can get an answer in ten minutes. And and these are huge problems. You know, these are these are 100,000,000 grid cell problems. And people were say suggesting, oh, well, maybe they'll just do one run. But, actually, I think it will be completely different in in that you'll be able to get an answer in ten minutes, and then you'll be able to ask a different question, and get a different answer. And then you'll be able to optimize between those different questions that you really wanted to answer. But we we always find there's pressure to deliver a project with if it's an M and E firm, they wanna they wanna get this project out. They're on tight deadline, and they can only run maybe six different wind wind setups for an external model. Well, if you put it on GPU, you could run 60 different wind setups and optimize the chili yard at the same time. So I think, like, okay. You're fair to say that those that CPU is probably in an old grade, but it is the kind of CPUs that we see our customers using. And and I think, it it it will be transformative in the way that people use simulation regardless of of hardware. But, yeah, it's a fair point. Alright. Yeah. Thanks, Mark. Oh, Miriam, she had a question, about what do we mean by token deployments in the in the presentation? We may have, mixed up the word deployments with token throughput. We typically talk about deploying GPUs, and, the tokens are what you get from GPUs. They're they're the outputs. But, Mark, unless you had something specific you had a specific meaning by I I don't know if if I said it or if you said it, but that might have been a a misstatement of the terminology. Token deployments? Yeah. Well, I think in terms of the capabilities, the the software can can look at any workload, and the workload is is effectively unitless in terms of you you specify what you care about. So in some cases, people are tracking tokens per second. In some cases, people are just looking at performance, you know, percentage. But but in but in terms of the capabilities, yeah, we we can, simulate different workloads on the on the hardware and translate that to power and then see, you know, the impact of running those workloads in different in different distributions within the data center. I think that's basically what I was hoping to say. Maybe I got it mixed up. That's. the Okay. that's what it can do. Yeah. Right. Right. Alright. And then, Azita asked, can you share your experience and challenges associated with building digital twins? Yeah. It's a great question. I think, running you know, doing this over so long, we've we've really tried to streamline how do we build these digital twins quickly because, quite often, the data is already available in other tools or in other areas, and we don't want engineers to have to redo everything twice. So there's a lot of different ways that we can import data. If you're sort of starting generally, data centers start their life in design in something like a BIM BIM software, from there, we, you know, we have we have mechanisms to pull BIM, three d geometry directly into the software, and it maps to our elements. So you don't have to redo the build. It it will just consume the BIM data and and, generate the the digital twin from that. We have obviously the libraries that, Sherman pointed to, so you're not, having to build those components yourselves. It you know, they're they're drag drop or, import if it's coming from a different tool. And then once we start talking about operationally so I've talked about, sort of, BIM data. If there's if you've got CAD, you know, traditional two d or three d CAD, we can import that too in the same way. And then once we talk about operational data, we, have many connectors that we've built over the years that allow us to ingest that data. So it's automatically populating and then, importantly, it's automatically updating the model. And that speed up that I talked about with the one two eight x is the fact that the customers don't no longer have to go and manually update the models. It just will sync the information. So whether it's a a a sophisticated DSM tool or a ServiceNow or if it's a spreadsheet that somebody loves to use because spreadsheets are still good, we can consume that data automatically and just populate the model. So the the initial, work is to set the model up, but we can streamline that with ingesting CAD. And then once it's built, the up the update of it really is is down to the integrations that we've built over time, and and and and that kind of opens up you just using it to get answers rather than using it to to build the model. Okay. Thanks, Mark. And then, Miriam asked, also asked if, the, DC elements include other products besides NVIDIA servers. Oh, yes. I I do get a sense that we've become a bit too NVIDIA centric in in our recent, presentations because a lot of the, you know, a lot of the AI factories is being driven by the NVIDIA ecosystem. But, we are vendor agnostic, absolutely, a 100%. So we have, from the the cooling systems, from CDUs, from the CRO units, from generators, chillers, all the way to racks, IT, We we we work with all of the main vendors. And in fact, most of them also, a a big proportion of them have our technology. They're using it to design the hardware, and then they give us access to those libraries. So we have, you know, libraries built on data that isn't just, public publicly available. The elements have a star rating system, which runs from one to five five with five being something that the vendors have actually rubber stamped and said this represents exactly how, the the hardware is is running, and they have test data to to prove it. So, with every element that you use in the in the model, you you know that it's backed up by, real data and real information that's provided by from the vendors, or or is public if in some cases, it might be like a public PDF. So so, yeah, totally, agnostic, and, you can use it to design. We talked a little bit about AI factories here because that's the kind of topic, but whether you're doing an old, you know, a twenty five year old enterprise data center that used to be an office or a telecoms, you know, any sort of flavor of data center, we have the we have the elements for it. Yeah. Okay. Thanks, Mark. And then, Danilo asked, how are we simulating machine learning load profile for transient analysis? Oh, that's a good question. So, the so the capabilities that we have with Transient include how power will move, and vary on the IT. So what you can do is basically adjust the load for, let's say, a machine learning job lands. We can use that curve that I put in in terms of the performance versus, power so that we can trigger that at certain point this, this workload will land. The the model will simulate the the time varying nature of the power coming into the IT, and then all of the control logic that's built into the various components will respond to that. So, generally, we've seen, the the air cooling control have been very slow in the past because, you know, you can wait you can afford to wait ten minutes if the racks are 10 kilowatts. But now that we're we're we're working in a world where there's 120 kilowatts of heat in each rack, the liquid cooling system needs to respond to that almost instantly. So we can see how the CDUs will respond to that workload landing and and indeed also when the workload, finishes. So, we are able to do that transient nature. You just sort of build the the performance curves in terms of what the the load from a machine learning job would look like. So it could be you know, the workloads can be really defined by what you're interested in. Some people are doing inferencing. Some people are actually doing a what is it what is it like if this data center does a big simulation? And, you know, whatever the workload is, we can kinda model it. And then because we have that transient time nature, we can we can see the impact over time too. Hopefully, Alright. that answered the yep. Yeah. I think that covers all the questions, and we. are eight eight, nine minutes over. That's alright. Cool. Yeah. Let me turn it back over to you. You bet. Thanks, Mark and Sherman. That was really fantastic. Appreciate you answering all those questions. So, ladies and gentlemen, this concludes today's webinar. But please take a minute to complete this short survey I'm about to launch at the end of this call. It's very, very brief. It'll take you just a few seconds. But on behalf of Cadence, I'd like to thank you for joining us today. We'll definitely see you next time.