NVIDIA GTC Studio with Insights from Vertiv

0:10 1 Hi, everyone.

0:10 2 Welcome to the NVIDIA GTC Studio.

0:13 3 My name is Tiffany Jansen,

0:15 and today I 4 am joined with a very special guest, Scott 5 Armel from Vertiv.

0:19 6 Scott is the Chief Product and Technology Officer.

0:22 7 Scott, how are you today?

0:23 8 I'm doing well.

0:24 9 Thank you very much for having me.

0:25 10 Thank you for being here.

0:26 11 Scott, I have to admit something.

0:28 12 When I saw that we were having a 13 sit-down conversation,

0:31 I had been really looking 14 forward to this because what you are doing 15

0:36 at Vertiv is solving a really big challenge

0:39 16 with AI and infrastructure that I think,

0:42 I 17 mean, I know you know this for many, 18 many years now,

0:45 but I think the general 19 public is starting to become aware of.

0:48 20 I think that's fair.

0:49 21 From a Vertiv perspective,

0:51 I think we've gone 22 from important to the data center in the 23 industry,

0:56 but relatively unknown to folks who are 24 actually paying

0:59 attention to who we are and 25 what we do now.

1:01 26 Vertiv, we're the power and cooling and overall

1:04 27 solutions infrastructure folks from a data center level.

1:07 28 We're helping to actually build and fabricate

1:10 and 29 design the data centers of the future.

1:12 30 I would say a lot of the awareness 31

1:14 and the industry recognition comes back to some 32 of our performance,

1:16 our growth, and all of 33 those things,

1:18 but the partnership with NVIDIA and 34 kind of jointly

1:21 developing a lot of the 35 things we're talking about,

1:22 I think has really 36 put us on a map to a lot 37

1:25 of folks that weren't previously paying attention

1:27 to 38 this space or to us in general.

1:29 39 Absolutely.

1:30 40 Before we started filming, I know we were 41 speaking about just that and you

1:33 mentioned now 42 my neighbors actually know what I do.

1:36 43 How do you explain to your neighbors, for 44 example, what does Vertiv solve?

1:41 45 What challenges does it solve?

1:43 46 Yeah.

1:43 47 So Vertiv, like I said,

1:44 we're really kind 48 of the critical infrastructure folks.

1:47 49 So when we think of powering data centers,

1:49 50 we think of cooling data centers,

1:51 and we 51 think of how we actually deploy and bring 52 up the physical

1:55 infrastructure of the data center 53 building

1:57 and all of the interconnective pieces themselves.

1:59 54 That is what Vertiv does.

2:00 55 We design it.

2:02 56 We think through kind of the future state

2:03 57 of where products and solutions need to go 58

2:06 to enable the AI solutions and the chip 59 architectures

2:09 that Jensen and the NVIDIA team are 60 talking about.

2:12 61 We help as kind of the foundational

2:14 infrastructure 62 level bring that to reality.

2:17 63 Absolutely.

2:17 64 When I was speaking to my mom this 65 morning and sharing about Vertiv,

2:21 she's like, well, 66 what do they do?

2:23 67 And here's how I explained it.

2:24 68 Tell me if this is accurate at a 69 very high level.

2:27 70 I said, you know, models we use that 71 run, you know,

2:31 different chatbots that we interact 72 with, they need thousands of GPUs to run.

2:36 73 But for those GPUs to run,

2:37 they need 74 two things effectively, which is power and cooling.

2:41 75 And those two things can bring challenges.

2:44 76 And basically that is what you are solving.

2:46 77 100%.

2:47 78 To query that chatbot, to run that AI 79 inference,

2:52 to do all of those things that 80 a normal

2:54 person would interact with AI or 81 candidly any data center,

2:58 those servers, those chips, 82 those GPUs have to sit somewhere.

3:02 83 They sit in a data center.

3:03 84 That data center has to deliver power to 85 that chip so that it can run.

3:08 86 And effectively that chip creates heat.

3:10 87 And a lot of the solutions that Vertiv 88 deploys,

3:13 whether it's liquid cooling or air cooling 89 or other types

3:15 of approaches is about removing 90 that heat from the chip,

3:18 getting it out 91 of the racks,

3:20 getting it out of the 92 data hall in the building

3:21 and figuring out 93 how to get it to the atmosphere

3:23 or 94 ideally to figure out ways to reuse it 95

3:26 to make all of these data centers we're 96 driving more efficient,

3:29 more effective and maybe less 97 impactful to the world as well.

3:33 98 Absolutely.

3:34 99 You had some big announcements this week here 100 at GTC.

3:37 101 Could you share some of them with me?

3:38 102 Yeah.

3:39 103 The biggest announcement I think was

3:41 the creation 104 of the Omniverse Rubin DSX, our one core 105 solution for that.

3:47 106 Really what that means from a Vertiv perspective 107 is an end-to-end,

3:51 pre-designed, pre 108 -fabricated,

3:53 converged systems engineering approach to a full 109 data center.

3:57 110 So we think of all of the elements 111 of powertrain and thermal chain,

4:01 the liquid cooling 112 we talked about, the heat rejection of a 113 data center.

4:05 114 We have partnered with NVIDIA and others to 115

4:09 pre-define what all of those interfaces look 116 like,

4:12 what that site solution ultimately looks like 117

4:14 and to do it in a way that 118 is maybe changing how we would have done 119 data

4:19 center design previously and that we're doing 120 it digitally.

4:22 121 So within NVIDIA's Omniverse,

4:23 we're trying to drive 122 towards more physics-based and simulated types of 123

4:28 designs to speed up a lot of the 124 converged infrastructure approach

4:32 so that long before a 125 shovel ever hits the ground

4:35 or a construction 126 company is trying to put up a data 127 center,

4:37 we understand the performance of our gear.

4:40 128 We understand how the chips and the clusters

4:42 129 and all of the servers working together will 130 perform in that environment

4:45 and we can launch 131 that to the industry

4:48 or to all of 132 the customers so that we can help speed 133 deployments

4:51 and we can help enable us to 134 get time to token a little quicker.

4:56 135 Absolutely.

4:57 136 And that's one of the biggest things

4:58 that 137 comes to mind for me anyways when I

5:00 138 think about digital twins is really how they

5:04 139 enable companies to move so much faster now.

5:06 140 Do you have any examples, maybe not even 141 specific numbers,

5:10 but examples of how you've seen 142 using digital twins speed up the process?

5:15 143 I think it's critical that digital twins are 144 really

5:19 an enabler for us to get out 145 of this bespoke on-site

5:23 custom construction world 146 that we've previously lived in to something

5:27 that 147 is truly more pre-designed and thoughtful around 148 the building

5:32 blocks that have predefined interfaces and 149 connection points and how

5:35 ultimately our gear needs 150 to be deployed on-site in a way

5:38 151 that doesn't require the complexity of figuring out 152 after

5:42 it's already been delivered or after it's 153 already been commissioned.

5:45 154 That's really been the name of the game 155 for us is

5:48 figuring out an enablement path 156 for us to do it digitally with physics

5:53 157 -based capability within the products themselves so that 158 we know

5:57 those simulations and those emulations of 159

5:59 how the infrastructure will behave are true.

6:01 160 Exactly.

6:02 161 Yeah, I like that.

6:03 162 The word thoughtfully.

6:05 163 I think that's such a powerful word

6:08 to 164 describe what digital twins are doing.

6:09 165 It's simple, but it's very powerful.

6:11 166 Absolutely.

6:12 167 Scott, when it comes to converged physical infrastructure,

6:16 168 how do you explain this topic to folks

6:21 169 listening who maybe aren't as familiar with it?

6:23 170 It's a lot of big words.

6:24 171 Potentially a lot of big words and a 172 lot of buzzwords as well.

6:27 173 We hope it's not just corporate speak,

6:30 but 174 really converged infrastructure is trying to get out 175

6:33 of this idea of individual point products

6:36 that 176 weren't designed together that were

6:38 potentially designed by 177 a bunch of independent companies all trying to 178

6:42 come together on-site and it becomes either 179 a consulting engineer's problem,

6:46 a construction company's problem.

6:48 180 Converged infrastructure for us is really about,

6:51 I'll 181 use the word thoughtfully again,

6:52 the predefined block 182 sizes, the interface points,

6:56 the deployment model for 183 that infrastructure.

6:58 184 We do it in a digital simulated environment.

7:02 185 We do it in a CAD environment so 186

7:04 that we can actually see how all of 187 the piece

7:06 parts will exist in Omniverse or 188 in a digital

7:09 environment long before it's ever 189 put out into the marketplace.

7:13 190 Those key interface points help us to actually 191

7:17 design better total solutions as opposed to better 192 point products.

7:20 193 Exactly.

7:21 194 Total key solutions it really comes down to.

7:23 195 When you think about that, the total key 196 solution,

7:26 we spoke about this a little bit 197 throughout the conversation,

7:29 but where would you place 198 Vertiv throughout the entire AI infrastructure?

7:35 199 In the five layer cake analogy that has

7:39 200 become so fun to dissect this week, we're 201 the infrastructure layer.

7:43 202 The most important layer.

7:45 203 It's foundational.

7:46 204 Now you said it, I didn't, but I 205 will run with that.

7:48 206 The most important layer because it's the foundation

7:50 207 of everything that eventually gets built on top.

7:53 208 When that foundation is correct and when

7:56 that 209 foundation is defined and deployed

7:59 in a way 210 that is thoughtful and purpose

8:02 built and with 211 these end solutions in mind,

8:04 it allows us 212 to take that infrastructure layer,

8:07 pair it with 213 how energy comes into data centers,

8:09 pair it 214 with a thoughtful and more efficient way

8:12 in 215 which the DSX deployments need to be run.

8:17 216 It allows us to think of the entire 217 chain.

8:19 218 When we think of that entire chain, that 219 purpose built design,

8:23 those converged infrastructure points, we 220 can unlock better performance.

8:26 221 We can drive faster time to market.

8:28 222 We can drive lower overall TCO for the 223 companies we work

8:33 with in a way that 224 when you think about scaling to gigawatt levels,

8:36 225 we can't be fumbling over ourselves and multiple 226 parties

8:40 and different trades all trying to work 227 at the same time.

8:43 228 These converged building blocks and this converged approach 229 to design

8:47 ideally makes us look and feel 230 a lot more like a factory.

8:50 231 Absolutely.

8:51 232 And that's why it's the foundational, aka important,

8:55 233 most important layer because it really needs to 234 all come together.

8:59 235 Scott, how are you seeing,

9:01 and I know 236 you just kind of gave a few examples 237 of this, but how are you

9:04 seeing emerging 238 AI workloads and system architectures changing traditional

9:08 assumptions 239 about power density and cooling in data centers?

9:12 240 I think in previous years,

9:14 data centers candidly 241 didn't really care what the workload was.

9:18 242 It was going to run on a CPU 243 -based server.

9:20 244 It was going to have a particular heat 245 draw or energy draw,

9:23 and you didn't really 246 need to know or understand.

9:26 247 The difference with AI is it is kind 248 of a unit of compute type

9:31 of an 249 approach where it's a very optimized

9:34 for that 250 either performance parameter or for that workload,

9:39 and 251 we know what it will look like.

9:41 252 When we think of kind of a cluster

9:43 253 size or a super pod in NVIDIA's nomenclature,

9:46 254 we can then back up from those purpose

9:48 255 -built pods to what does the powertrain and 256

9:51 the infrastructure need to look like to best 257 support

9:54 that from a size and optimization and 258 efficiency perspective,

9:57 and we can scale that same 259 analogy and that same approach out to the 260

10:01 entire site so that that site is optimized 261 to work as a single computer.

10:05 262 Wow.

10:06 263 It's incredible.

10:07 264 It's so exciting the times we're in, and 265 I know before we were filming,

10:11 again, you 266 were mentioning that you've

10:14 been with Veritiv for 267 a while now,

10:16 so getting to see the 268 evolution and to where we are now,

10:20 what 269 do you think about when we look ahead 270 a little bit?

10:22 271 I know this can be a vague question,

10:23 272 but when we look ahead a little bit, 273 I need to know.

10:26 274 I need to pick your brain.

10:27 275 What does it look like when it comes 276 to this stuff?

10:29 277 Yeah.

10:30 278 I think we're seeing compounding acceleration,

10:33 which is 279 extremely exciting for us.

10:36 280 From our perspective, our engineers, our service folks,

10:40 281 we're tied into the direction that GPUs are 282 going.

10:43 283 The pace of technology is accelerating.

10:45 284 The roadmap that NVIDIA is putting out there 285 for GPUs

10:48 and performance improvement that is being 286 unlocked is truly mind-blowing,

10:52 and from our 287 perspective,

10:54 we're in a race to try to 288 not only keep up with that pace,

10:57 but 289 we're trying to look multiple GPU generations ahead 290 to say,

11:00 what are the physics-based problems 291 we're going to need to solve?

11:02 292 How do I get this much power into 293 a rack that's 600 megawatts or beyond?

11:07 294 How do I scale to a data center 295 site or an AI factory

11:10 that is a 296 gigawatt in capacity

11:13 and work backwards somewhat maniacally 297 to say,

11:16 all right, here's what the roadmap 298 needs to look like.

11:17 299 Here's what the development needs to look like.

11:19 300 Here are our technology investments that are going 301 to be needed

11:22 for us to make sure 302 that we're

11:23 not tripping over those physics-based 303 problems,

11:25 and we're actually able to turn on 304 data centers well in advance of when

11:29 the 305 GPUs and the clusters will be ready to 306 be deployed there.

11:32 307 Absolutely.

11:32 308 I mean, it's easy for me to say,

11:34 309 but although these are different challenges that you're

11:37 310 identifying and will solve and are solving already,

11:40 311 they're exciting ones.

11:41 312 Scott, thank you so much for your time 313 today.

11:43 314 It was really great to sit down with 315

11:45 you and learn more about what Vertiv is 316 doing,

11:47 especially in partnership with NVIDIA.

11:49 317 Absolutely.

11:49 318 It's my pleasure.

11:50 319 Thank you.

11:51 320 Thank you.

11:51 321 Thank you so much, Scott.

11:51 322 And for those of you tuning in, make 323 sure

11:53 to go check out more sessions on 324 the GTC catalog.