Accelerating Next Generation Chip Design With Agentic AI and GPUs

Welcome to GTC. Welcome to Nvidia. My

name's Rambo Jacobe.

You know, when most people meet me and they find out I work at Nvidia, there is one common question that they all ask.

They say, you know, we hear about Nvidia all the time, most valuable company in the world, the lynchpin for AI.

But the question they ask me is, what does Nvidia do? What does Nvidia do? And

that is a really complex question to answer because we are a very complex company. We're complex in the products

company. We're complex in the products that we develop. We're complex in the ways that we work with our customers.

We're complex in the way that we work with our partners. So, what I want to do for the next 40 minutes or so is I want to try to explain who we are, what we

do, how we got here, where we're going, and how we're going to get there. And

the great news is that I'm not going to show you any slides. Okay, this isn't a slide. This is wallpaper. Okay, this is

slide. This is wallpaper. Okay, this is wallpaper. But what I'm going to do is

wallpaper. But what I'm going to do is I'm going to try to show this to you in terms of demos of some of the tools that we develop and actually showing off some of the hardware that enables this

agentic AI revolution that we're experiencing right now. Now, two things that I hope that we can do for the session is one, I don't want this to be me talking to you. I want this to be me

talking with you. So, I'm going to try to make this as um as connected as I can. And please humor me if I throw out

can. And please humor me if I throw out questions at you all. I hope that you all try to answer it. And the second thing is right after the session, I'm going to host a connected with experts session back in the main hall where you

can come, you can ask any questions that you want from the content that I've showed or just any questions about Nvidia. You can see and you can hold

Nvidia. You can see and you can hold some of the hardware that I have with me here right there. And and it's something I hope you don't miss because there's not many people in the world get to that that get to hold a B100 chip. So please

take advantage of that. And uh without further ado, let's let's go ahead and get started. So what we're going to do

get started. So what we're going to do is first I want to start off by talking about some of the hardware that Nvidia develops. Right after that, we're going

develops. Right after that, we're going to do a hard switch and we're going to talk about all the software that Nvidia develops for agentic AI. And then we're going to do another hard shift again and we'll talk about all the software that

we developed for physical AI and we'll try to see if we can have a couple minutes towards the end of the conversation for live questions. So

let's go and get started. I'm going to do the grand reveal.

So the first couple things that I want to show you is this right here. This is Blackwell. This

is a product that we have called B100.

This is the GPU that we sell to all of our customers who develop AI models, who launch those models, that run inferencing on those models. So

customers like AWS and GCP and Oracle and customers who put their own on premises data center solutions together.

This is the product that does all those mathematical computations for training those AI models. This is the product that does all the mathematical calculations for being able to run

inference of those models. Now let's

talk about what a GPU is fundamentally.

A GPU is essentially a massive parallel processor. It can do lots of things in

processor. It can do lots of things in parallel. So to give you a sense for

parallel. So to give you a sense for that today, if you look at one of the highest performing x86 CPUs that go inside of a data center, CPUs

traditionally have about 300 cores inside of them. So they can do about 300 things in parallel. Whereas this machine here, this has 15,000

cores inside of it. It can do 15,000 things in parallel. So while CPUs are wonderful for being able to do serial calculations, these machines are wonderful for being able to do a lot of

parallel calculations very quickly.

That's why they're so good for AI. Now,

this is probably one of the most complex machines ever manufactured by humankind.

This is over 200 billion transistors in this machine. This is worth this took

this machine. This is worth this took almost $10 billion for us to develop to develop the architecture for it. So, so

really an incredible machine. Now,

typically when people see this, this is where their misunderstanding about Nvidia begins because they look at this and they say, "Invidia, I understand them. They are a GPU company." Okay? But

them. They are a GPU company." Okay? But

we're not a GPU company. Can Can anyone out here guess who we are and what we are? It starts with a P. Any guesses?

are? It starts with a P. Any guesses?

Yell it out.

platform platform. We're an AI platform company. Let's talk about what that

company. Let's talk about what that means to be an AI platform company. And

a great product to be able to show that with is this machine here. This is our GB 300

super chip. Okay, this super chip has

super chip. Okay, this super chip has two of those Blackwell GPUs that I was just holding up, but it also has our own

Grace CPU. This is a ARMbased CPU that

Grace CPU. This is a ARMbased CPU that we've developed. It's not based on x86.

we've developed. It's not based on x86.

Now, a common question that people have for us when they see this, when they hear about our development of a CPU is why? Why did you go out and reinvent the

why? Why did you go out and reinvent the wheel? There's plenty of x86 CPUs on the

wheel? There's plenty of x86 CPUs on the market. Why can't you just utilize

market. Why can't you just utilize standard x86?

Well, there are two things that Nvidia focuses on maniacally. The first is performance, but second, and probably one of the most important things that Jensen hit on a lot during his keynote,

is performance per watt. It's power

efficiency. This CPU has higher AI performance than x86 CPUs do, but with half the power. Half the power.

Everything that we focus on is on performance, but performance per watt.

And I'll come back and I'll talk about power efficiency a little bit more in a moment. Now, besides the GPU and the

moment. Now, besides the GPU and the CPU, we are also one of the largest networking companies in the world. Not

many people know that about us. We

develop all the networking solutions for being able to connect hundreds of thousands of GPUs together. Okay, so

we're also a networking company. But

probably the most important aspect of being a platform company, the fourth and most important aspect, any guesses what it is? Software. We are a software

it is? Software. We are a software company. I'll talk about software a

company. I'll talk about software a little bit more in just a few moments, but I want to talk about power efficiency just a little bit more. So,

this machine, this GB300 goes inside one of these Blackwell racks, one of these NVL72 racks. Each of these racks has 36

NVL72 racks. Each of these racks has 36 of these. So, about 72 GPUs per rack.

of these. So, about 72 GPUs per rack.

So, a couple of years ago, about seven years ago, we helped build one of the world's largest supercomputers. was a

supercomput called Summit. We built it for Oakidge National Labs. They were

doing things like weather simulation and molecular simulation with it. Now,

Summit was about 120 pedlops of performance. A flop stands for

performance. A flop stands for floatingpoint operations per second. 120

pedlops. Incredibly performant machine.

It was 27,000 GPUs, so about a 6,000 square foot data center. And it was 13 megawatts of power. Now today, just 7 years later, one of these machines is

capable of 300 pedlops of performance.

So more than double the performance.

It's only about, you know, one square meter in size, but it's only 120 kilowatt.

So let that sink in. Over the course of 7 years, we've more than doubled the performance of one of the largest supercomputers in the world, and we're now running it at 1/100th of the power.

Everything is about performance but it's also about power efficiency. Now to talk about how we're able to get that power efficiency. When we built Summit 7 years

efficiency. When we built Summit 7 years ago, the only thing that we contributed to it was our GPU. Whereas today we provide the platform. Every aspect of

this machine is controlled by Nvidia. We

develop the GPU, we develop the CPU, we develop the networking, we develop all the software, we develop all of the cabling. We control every aspect of this

cabling. We control every aspect of this machine. So there are so many ways that

machine. So there are so many ways that we're able to optimize that machine to be able to get these performance and power efficiency improvements. We are

not a chip company. We are a systems company. We're a platform company. Let's

company. We're a platform company. Let's

talk about software. We'll come back and show a little bit more hardware in a few minutes.

So we at NVIDIA, we are a tiny company.

We're a tiny company. We have a lot of reach. We have a lot of influence. But

reach. We have a lot of influence. But

from a employee headcount perspective, we're only about 40,000 people. We're a

very small company. About half of those 40,000 people that work for us are software engineers. We build a lot of

software engineers. We build a lot of software. We build a lot of foundational

software. We build a lot of foundational models. And let me explain what a

models. And let me explain what a foundational model is. Today, if you're an enterprise and you want to be able to use what's called a reasoning agent to be able to improve the per the performance of your workloads, to be

able to get more efficient at the work that you do, what you can do is you can go out and you can build a reasoning model from scratch. But the problem with that is that it'll cost you billions of

dollars. It'll take many years for you

dollars. It'll take many years for you to do it. But the other option is you can use a foundational model, an open- source foundational model that has already been developed that you can

build on top of. And that's what Nvidia does. We develop a lot of foundational

does. We develop a lot of foundational models that we open source. That was

another big point of Jensen's keynote.

We open source all of our models and we make it widely available for any of our customers just go ahead start using it and building up their own applications on top of it.

One of the sites that I want everyone to be familiar with is a site that we have called Build.

Build.envidia.com.

Build is where we take all of the software that we develop. We develop

software for reasoning. We develop

software for vision, visual design, retrieval, speech, biology, for all the different verticals that we work in. And

we make it available here on build such that anybody, this is a public facing website, any of our customers, any college students, anyone that wants to learn about AI and wants to learn about how these models work and experiment

with how they can work in their own workflows, they can go, they can experiment with them here. They can run inferencing on those models. They can do it on our dime. It all runs in our own

data center and try to really understand how it works for their own applications.

Now, we have hundreds of models here. I

don't have time to show you all hundred, but I want to show you just a couple of them. And one of the ones that we're

them. And one of the ones that we're really excited about is this model here called Neatron. Neatron is our internal

called Neatron. Neatron is our internal reasoning foundation model. And I want to show you what Neatron can do. So, I'm

going to go ahead and click on Neatron.

I've already opened the tab up. And this

is where you land.

Now, you know, one of the things I really don't like about this page, about this landing page, is that we give these really silly examples of what you can ask this foundational reasoning model to

do. How many Rs are in strawberry?

do. How many Rs are in strawberry?

Right? Who cares? Who cares? We want to see how these machines can be used in real world applications to improve enterprise efficiency. So, let me go

enterprise efficiency. So, let me go ahead and show you one of those examples. One of the types of customers

examples. One of the types of customers that we often get at NVIDIA, we get a lot of manufacturing customers that come to us. And one of the challenges that

to us. And one of the challenges that manufacturing customers have is they have a challenge with predictive maintenance. Predictive maintenance is

maintenance. Predictive maintenance is when you want to be able to take all the machinery that you have in a factory, as an example, and be able to predict when that machine is going to break down so you can go out and can fix it before it

actually breaks down and takes down your entire manufacturing line with it. So,

as an example, they might want to be able to figure out when a robotics's arm is going to break down. Now today this is a solved problem. You're able to do predictive maintenance with machine learning algorithms with traditional

machine learning algorithms. But the challenge with traditional machine learning algorithms is that they are really good at taking in a lot of data and identifying an outlier. They can

take a lot of temperature data as an example for a robotics joint and be able to detect when the temperature is going a little bit higher than it should and say, "Hey, something is wrong. You got

to focus on that machine." But they can't tell you what's wrong. they can't

tell you how to fix it. Well, let me show you how you can do that with a reasoning model. So, we're going to come

reasoning model. So, we're going to come here and we're going to ask this reasoning model a question. We're going

to say, "Hey, you are a predictive maintenance assistant. The below data set contains

assistant. The below data set contains the generated RPMs of a GE Vernova." You

all know Genova? They make wind turbines. They make turbines that turn

turbines. They make turbines that turn gas and electricity. One of our one of our great partners, Jensen, mentioned them as part of his his keynote. So you

are a predictive maintenance and you're going to take these generator RPMs of a GE VOVA 1.5S 1.5 megawatt wind turbine under full load. Please review the data and determine if the system is working

properly, if there are any issues beginning to form and what the potential issue may be. Very very real practical application, right? Very real problem

application, right? Very real problem that some industries face. So, what I'm going to do is I'm going to come over to this tab. And here I have 300 data

this tab. And here I have 300 data points. 300 generator RPM data points

points. 300 generator RPM data points for a Genova wind turbine. I'm going to take these 300 data points and uh we're going to go ahead and put

it into this model. And uh

I'm just going to paste it in and we're going to let it think. I'll turn on the chain of thought for you so you can watch it think. What this is doing right now is it's going through it's looking at all that data. is trying to find relationships and patterns of the data.

It's going out and researching that Genova wind turbine and very quickly in a couple seconds it comes up with a full analysis. Let's take a look at this

analysis. Let's take a look at this analysis and see what it says.

So what this did is it went out and it researched this machine and it says you know what under full load this win turbine typically operates at around 1500 RPMs. I didn't tell it that. It

figured it out. And what it's noticing is that initially this machine is working really quite well. It's

operating between 1497 to 152 rounds per minute that generator based off of those data points that we gave it. So it says initially this is starting to work really well but over time one of the things that it sees is that there's an

issue beginning to emerge. It's starting

to see that there is all this variation coming up in that winterbond data. Those

RPMs are going as low as 1480. They're

spiking as high as 1519. So it's

starting to detect an instability. Okay.

It's starting to detect outlier data and an instability. And it says, "Well, you

an instability. And it says, "Well, you know what? There's this potential issue.

know what? There's this potential issue.

There's this instability. Let me try to figure out why. Let me try to figure out what's causing it." So, we go a little bit further down and it says, "You know what? Here are all the issues that you

what? Here are all the issues that you could be having. One, you might be having an issue with your gearbox. Two,

your control systems might might be malfunctioning. They're not controlling

malfunctioning. They're not controlling the pitch of those fans very well. Your

bearings might be breaking down. You

might have a sensor breaking down. We

can push this model a little bit further. I can say to it, hey, you know

further. I can say to it, hey, you know what? It's actually really cold and

what? It's actually really cold and rainy outside. And based off of that,

rainy outside. And based off of that, it'll reason. It'll come back and say,

it'll reason. It'll come back and say, you know what, you might have ice accumulating on your wind turbine fans.

It's causing that wind stability. Send

someone out to fix that ice. This

machine very quickly took data points.

It understood that there was outliers in it, and it came back and told you what the issues are and potentially how you can fix it.

Now we at NVIDIA we open source this foundational model and our customers Genova potentially could take this model and they would post train it. They would

do a little bit of additional work to post train it on all of the data that they have about their win turbines and by doing that they can create this agent that perfectly understands their product, perfectly understands the

failure mechanisms and can help them figure out why a machine is breaking down before it breaks down. Okay, we

provide that foundational model. Let me

show you another example. Let me show you another example. I'm going to take the same exact model. I just opened it up in a new tab and I'm going to ask it a different question.

One of the issues that we have in the US is that the cost of health care is going high and we need to try to figure out how to bring that cost down. And one of

the best ways to be able to take out inefficiencies is identifying fraud in the insurance market. be able to identify fraud when people submit

claims. And I want to show you how this same reasoning model that I just used for predictive maintenance can identify fraud. We're going to come here. We're

fraud. We're going to come here. We're

going to ask this machine a question.

We're going to say, "Hey, please review the following dental insurance claims and identify the ones that are potentially fraudulent and require extra scrutiny." So, I'm going to go up here

scrutiny." So, I'm going to go up here and I'm going to copy this set of dental insurance claims and I'm going to give it to that model.

and let's let it think.

Again, I'm going to turn on the chain of thought for you.

What this machine is doing right now is it's going through and it's researching each of those claims. It's researching what the typical cost should be for a claim. It's trying to understand the

claim. It's trying to understand the diagnosis that someone had when they went into that dental office and comparing it to what the procedure that they received is. And very quickly, it comes back for us and it says, "Hey, you

know what? Here are the claims that you

know what? Here are the claims that you should go look at." So, some examples of claims, this claim here from Smilecare Dental, the issue is that a partial denture is justified with a $12,000 full

mouth reconstruction despite simple partial denture needs. So, it understood the diagnosis. It understood the

the diagnosis. It understood the treatment that was being suggested. And

it said, "Wait a minute, that doesn't sound right. Go look at this one a

sound right. Go look at this one a little bit further." Let's look at another one of these claims. Let's look at claim 19 from Bright Smiles Clinic. the issue

with severe periodont periodontitis and it's treated with a $600 scaling and root planning. And here are its red

root planning. And here are its red flags. It understands that this

flags. It understands that this procedure typically requires surgical intervention and not this scaling and root planning that they're being charged $600 for. So again the same foundational

$600 for. So again the same foundational model what we do is we provide it to insurance companies and they're able to post train it with years of insurance

data and develop a single agentic model that is able to go through take care of ingesting all of this claim data and trying to reason about whether or not that claim data makes sense making

people a lot more efficient when they're trying to identify fraud. Now, a lot of you are looking at this and you're thinking to yourself, "Wow, this is

amazing." But some of you are looking at

amazing." But some of you are looking at this and you're thinking, "So what? So

what? We're able to do this with Perplexity. We're able to do this with

Perplexity. We're able to do this with OpenAI. We're able to do this with

OpenAI. We're able to do this with Gemini. The solutions already there. And

Gemini. The solutions already there. And

by the way, those are all great companies. We work with them very

companies. We work with them very closely. They're good partners of ours.

closely. They're good partners of ours.

But here is what the so what is. The so

what is if you take one of those models let's call it chat GPT and you post train it with all of your proprietary data about your business you don't own

that model chat GPT is going to own that post-rained model none of that belongs to you it is not your IP those are closed models whereas when you take

Neatron which is an open-source model when you post- train it that IP belongs to you you can use it when you want, where you want, how you want. It's your

IP. That is the fundamental difference between these open- source models and these closed models. It belongs to you.

Any improvements that you make are yours and no one else's. Let me show you another example. Hey, um, curiosity, how

another example. Hey, um, curiosity, how many of you here are uh in the financial services space? Anybody from the banking

services space? Anybody from the banking space? Okay, we get a lot of banks that

space? Okay, we get a lot of banks that come to us at Nvidia, some of the world's largest banks.

And uh one of the biggest challenges that they have is managing all of their cameras. So some of the banks that come

cameras. So some of the banks that come in, they have thousands of branches.

Let's call it 5,000 branches for one of the larger one of the larger banks. And

in each of those banks, they typically have between 50 to 60 cameras that manages security, okay? That watches

everything that happens. And often times when the CEO comes in, I love asking them this question. I look at them and I say, "Hey, you know what? You're very

successful. You have all these branches.

You have between 500 to 600,000 cameras that are recording content on all your branches. And I ask the CEO, I say,

branches. And I ask the CEO, I say, "Hey, who watches all those cameras for you? And what do you all think their

you? And what do you all think their response is?"

response is?" No one. No one watches those c that

No one. No one watches those c that content. They're being very retroactive

content. They're being very retroactive about it. What they do is they store all

about it. What they do is they store all that camera data and when something goes terribly wrong, when they get robbed or some other incident happens, they go back and they look at all that camera data. Well, what we try to teach our

data. Well, what we try to teach our customers is how they can go from being retroactive to being proactive. And we

do that with a piece of software called our video search and summarization engine. I'm going to have you all watch

engine. I'm going to have you all watch a brief video here. I'm going to have you all watch this brief video. Now, I'm

going to make a quick disclaimer before I show you this video. We have amazing employees at Nvidia. This video that I'm going to show you is of one of our own warehouses over in Santa Clara. And all

the employees that you're going to see are our own employees. We have amazing employees, but they are absolutely terrible actors. None of them are gonna

terrible actors. None of them are gonna win any Emmys here, okay? So, please

forgive them. So, in this video, a couple things happen. In this video, a couple things happen.

One, we get this engineer who walks in. He's

carrying all these boxes and oops, accidentally drops a box. That's like a half a million dollars worth of GPUs right there. Okay, I'm going to fast

right there. Okay, I'm going to fast forward. So, he just leaves that box and

forward. So, he just leaves that box and he walks away. I'm going to keep fast forwarding through this. We go a little bit further and we have another employee who walks in, picks up the box. He's

shocked. He's shocked. And he goes ahead and puts the box away in the wrong location, mind I add. So, we'll never find that inventory again. I'm going to keep going. And now, here's something

keep going. And now, here's something interesting that happens. This employee

comes back and what he does is he takes this open space, this open warehouse space, and he starts putting up caution tape across it. What he's doing in essence is he's creating a restricted

environment.

Now, this guy got to fast forward through a lot because he is really slow at what he does. I'm pretty sure he gets paid hourly, so he's still putting up caution tape.

He finishes up hanging caution tape and he walks away and we get another employee who comes in.

Funny enough, this is one of our VPs and he says, "You know what? I could care less about your restricted area. I'll

just go wherever I please." This

probably happens in a lot of spaces, whether it be a bank, whether it be a warehouse, whether it be a manufacturing site. So, what I did right now is as I

site. So, what I did right now is as I was presenting this video to you, I took it and I passed it through what we call our video search and summarization engine. What this engine did is it

engine. What this engine did is it watched the entire 3minute video that I fast forwarded through. It summarized

the whole thing, created all this metadata about it that we can now go and have a conversation with. And so I'm going to come down here and uh we're going to ask this AI a couple of

questions. We're going to say, "Hey, you

questions. We're going to say, "Hey, you know what? I received some broken

know what? I received some broken GPUs in my warehouse.

Did anything happen that could have resulted in these broken GPUs. I'll fix my typing. I don't have

GPUs. I'll fix my typing. I don't have to. It's a large language model. It'll

to. It's a large language model. It'll

figure it out, but I'm an engineer. I

just can't let it stand this way. Okay.

So, we're going to ask it the question, and we're just going to let it think for a few moments. It's going to think and it's going to come back and it's going to say, you know what,

based on what I saw, there were a couple of things that could have resulted in those broken GPUs. One instance is when worker zero drops a box on the floor, which is considered an anomaly. And

although the contents of the box are not specified, it is possible that the box contained the GPUs that were damaged.

What this model did is it watched the content and it reasoned through it. it

reasoned about how what it saw in the video could have resulted in these GPUs being broken. Let me show you another

being broken. Let me show you another example. Let's come here and let's ask

example. Let's come here and let's ask ask it another question. We could say, "Hey, you know what? Um, I experienced

some theft in one of my restricted areas. Can you tell me if anyone entered

areas. Can you tell me if anyone entered a restricted area in my warehouse?"

I know I have a typo. I'll just let it I'll just let it go and let's see what it comes up with. Let's have a think.

And actually, yes. You know, according to the video summary, worker seven entered a restricted space in your warehouse. The worker crossed across the caution tape and entered the restricted area while carrying a box,

which is a potential safety violation.

And I love this because it throws them under the bus twice. Additionally,

worker 7 was not wearing any visible personal protective equipment at the time. Okay, what this model did is it

time. Okay, what this model did is it didn't just do facial recognition. It

could do facial recognition too, but it was capable of reasoning through what it saw, reasoning about the contents. This

is how any owner of a physical space is able to take all of those cameras that they have and go from being a reactive company to a proactive company. We also

have a piece of software called Deepstream that connects to this model and in real time it connects to all the cameras that are across an organization.

It ingests all of that data and is able to tell in real time when something is happening that shouldn't be. Okay, so

you can run all these models in real time as well. Okay, let me move on.

Everything that I've showed you so far is considered to be reasoning. But one

of the areas that Nvidia is focusing on really heavily is physical AI. We want

to be able to simulate the real world.

And here's an example of how we do this.

And by the way, VSSs, we have a great demo of it on the showroom floor. I I

encourage you all to go take a look at it in real time and experience it and experiment with it. Give it your own questions. So this is a model that we

questions. So this is a model that we have called that we have called Neoysics. And with this model, what

Neoysics. And with this model, what we're able to do is we are able to simulate real physical objects. Okay. So

the example here is that we want to take a car and we want to be able to virtually run it through a wind tunnel computational fluid dynamics to figure out if that car is aerodynamic to figure

out if that design is aerodynamic. Now

traditionally this is something that you would run through a full CFD engine using a piece of application called Ancis Fluent but that is a very expensive process to go through. What

we're able to do is we're able to create models that are able to simulate these physical objects in real time. This is a digital twin of a vehicle that you can start experimenting with in real time

and get a first order approximation of what the aerodynamicity of that vehicle is. So, you know, I'm going to bring up

is. So, you know, I'm going to bring up a car that's a little bit more interesting here. Let's go ahead and

interesting here. Let's go ahead and swap this car for a different car.

Okay.

So, what we've done here is we've created a perfect digital twin of this sports car. We're simulating what the

sports car. We're simulating what the glass looks like. We're simulating what the aluminum body looks like, the friction of every component here, and we're able to run an analysis of seeing,

you know what, what happens if we, let's say, put up the spoiler on this vehicle.

Okay? Now, as I put up the spoiler, one of the areas that I want you to focus on, I'm sorry I can't bring this up any higher. One of the areas that I want you

higher. One of the areas that I want you to focus on is this area right here behind the tail of the vehicle. Let's

see what happens when we put up that spoiler.

You can see we're actually simulating the eddiurrens that occur in the back of that car in real time. So now you don't have to go through the full CFD process.

You can very quickly run this analysis.

We incorporate this software with customers and partners like Cadence as an example. We have a great sample of

an example. We have a great sample of that. We have a great demo of that on

that. We have a great demo of that on the showroom floor. We're able to show you how physics Nemo is able to simulate the aerodynamicity of an airplane and how you can experiment with the structure of that airplane in real time.

Make it very simple for you to come up with things like aerodynamics.

Let me give you a different example.

One of the areas that we're really bullish on and uh you may have noticed this in in Jensen's keynote is we are really bullish on humanoids. We believe

these machines are going to be in our society and prevalent within the next 5 years. And there's two things that we

years. And there's two things that we build for the space. We at NVIDIA are likely never going to build a robot because this isn't where our expertise

is. Our expertise is in two areas. Our

is. Our expertise is in two areas. Our

expertise is in building the edge inference that goes into these machines.

And this is an example of that edge inference. This is a product that we

inference. This is a product that we have called Jetson Nano. A humanoid

takes two of these machines. This is the far edge inferencing machine that runs all of the models that gives these machines the ability to do things like understanding the environment, things like path planning. So we provide this

edge inference machine. But the other thing that we provide is another open- source software that we have called Groot. group is the model that is

Groot. group is the model that is pre-trained by NVIDIA that allows these machines to path plan to understand their environments.

Now, here's a question that I have for you all. When you build an AI that works

you all. When you build an AI that works in the real world, that works in a physical environment, what does that AI need to be fundamentally aware of?

Some guesses like fundamentally, what does it need to understand?

>> Space. Keep going.

gravity physics is what it needs to understand. In our environment, we have

understand. In our environment, we have gravity, we have center of mass, we have momentum, we have friction. So, when you build an AI that works in our environment, what you need to do is you

need to simulate it and train it in an environment that we're able to simulate physics in. And this is what this

physics in. And this is what this environment is. This is Omniverse. And

environment is. This is Omniverse. And

real quick, I'm going to pan this camera so you can see that this is not a video.

This is a simulation that we're running here.

What we're doing here in Omniverse is on this machine. This is one of our RTX Pro

this machine. This is one of our RTX Pro machines. On this machine, we're

machines. On this machine, we're creating a digital twin of an environment. This is one environment

environment. This is one environment that we are using to train these machines on how to do things like walk up and down stairs. But the amazing thing about what we're able to do here in Omniverus is we are able to simulate

physics. We're able to simulate gravity

physics. We're able to simulate gravity so that these machines, these digital twins of these machines can learn how much force they need to put on their legs, on their joints to be able to move

up and down stairs. We can go ahead and pour water across this floor, make it very slippery, and simulate the frictional coefficient to make sure that these machines don't slip and teach them

how not to slip. Okay, but it all starts with a digital twin running on a simulation computer. Now, last thing I

simulation computer. Now, last thing I want to show you is I want to show you another example of a digital twin.

Digital twins are great. They're used

for robotics, but they're also used for being able to optimize manufacturing environments. You can use them to be

environments. You can use them to be able to do anything like optimize space.

This is an example of how one of our first customers using Omniverse was able to create a digital twin. This is BMW was able to use a digital twin to be able to simulate their entire

manufacturing line to be able to optimize the flow, the throughput, simulate the yield, the number of vehicles that they're able to generate.

And once they perfect that digital twin, once they're able to simulate every aspect of it, and by the way, one of our partners that's on the showroom floor is Seammens. Okay? Seammens is critical in

Seammens. Okay? Seammens is critical in this space is showing how they're able to create these digital twins in real time. Okay? Once you're able to create

time. Okay? Once you're able to create these digital twins, you're able to optimize every aspect of the of this floor, okay, you're able to test and experiment, but really the most important thing that you can do with

these digital twins is it can create an environment where you can teach the robotics. Okay, you use these digital

robotics. Okay, you use these digital twins to be able to simulate an environment that you can teach a robot how to move inside of. All right, my time is up. I've only got about two more

minutes, three more minutes, and I want to just give you all a couple minutes to ask questions if you like, but I'm just going to summarize a nearly five trillion dollar company for you in 10

seconds. Okay? We are not

seconds. Okay? We are not a GPU company. We're not a GPU company.

We're a platform company. Yes, we build all the hardware that enables AI, but we also develop all the software, all the foundational software that allows customers to be able to create their

agentic AI solutions on top of it, but as well as all the software that enables our customers to be able to build their robotics. That is NVIDIA. That is what

robotics. That is NVIDIA. That is what we mean by being a platform.

Okay, we got four minutes. Are there any questions that anyone has here that you'd like to ask live? and the rest of the questions that we can take inside of the inside of the connecting with expert session. Yes, sir.

session. Yes, sir.

Well, what what happens is a customer first starts designing their digital twin inside of a product like Seammens.

Okay, they lay out their entire manufacturing site inside of a Seammen's product and once they've perfected and this is how you do any digital twin, you start with a product like Seammens or

Dell Cadence. Once you perfect it there,

Dell Cadence. Once you perfect it there, you actually bring it into Omniverse where we give it the 3D realism where we give it all this dynamicity where we allow people to interact with it. So the

digital twin is how the manufacturing side is laid out just based on the the the fundamental design of that space. Uh

yes sir back there.

Yeah.

So great question. The question is do we model all the assets that are here? So,

Omniverse is a platform in the sense that we simply provide the tools and we work with a lot of partners who create assets that they're able to import inside of Omniverse. So, things like

those digital humans that we saw a little bit earlier walking around. We

don't create that. We work with a third party. We create a connector to that

party. We create a connector to that third party application and they create the digital human and they're able to import them into Omniverse. We take USD files. What we what we take CAD files

files. What we what we take CAD files when you typically start laying out a manufacturing environment you do it in CAD. We take those CAD files that are

CAD. We take those CAD files that are provided by other manufacturers. We can

call it one of the robotics arm manufacturers and we're able to import it into Omnibus. We're able to bring that asset in and give it that graphical appearance. Okay. But again, it's not

appearance. Okay. But again, it's not about having the end solution. It's

about creating the platform that others can develop their solution on.

>> Yes sir. you showed a number of use cases for the models that you guys provide. Which use case would you say is

provide. Which use case would you say is the most computationally complex or intense um of all the things that you you show?

>> Um video processing is complex, right?

All the inferencing that we do is complex. Uh just to give you an example

complex. Uh just to give you an example of the amount of compute that's needed for video processing, a server that has eight of our H100 GPUs, H100 being the

previous generation of GPUs, one of those servers can process about 200 camera streams simultaneously. Now, what

you wouldn't do is you wouldn't have those camera streams coming into H100 constantly because often times those camera streams are looking at environments where nothing happens. What

you would do is you would put a smaller model in front of that that's able to detect motion and once motion is detected then you would kick in the larger more powerful model. So based off of that you can say that one of those

machines with eight GPUs can theoretically process many many times more than 200 cameras. But you know that that is one of the one of the uh areas that that that is computationally

expensive. All right one more minute.

expensive. All right one more minute.

Any one last question anyone else? Uh

yes sir.

and uh you know um you know we've seen a lot of success in a lot of the closed models. Just curious on why do you think

models. Just curious on why do you think that's happening? Why aren't people

that's happening? Why aren't people people using open models more especially if they can sort of own the data, own the model?

>> Yeah. So closed models are just really easy to use, right? It's already there.

It's already been developed. With these

with these open models, there is a little bit of additional engineering work that has to go inside of it. So a

company does need to have a little bit of their own expertise, a little bit of their own programmers, but that investment pays off in the sense that you're able to create this model whose IP you fully own.

All right, everyone. Thank you all so much for being here. I'm going to be at the connecting with expert session down the down the hall. If there's anyone that have have any questions, if you'd like to see this hardware firsthand, please feel free to come over. Thank you

all. Enjoy your GTC.