SC25 | Glenn Dekhayser, Equinix & Alan Bumgarner, Solidigm

Clips
More from SC25

Glenn Dekhayser

Global Principal Technologist

Equinix

Alan Bumgarner

Director and AI Technologist

Solidigm

Why Solidigm and Equinix believe the AI race will be won through data strategy

Artificial intelligence models are growing more complex and data-hungry. Across the industry, organizations are rethinking how they manage and refine data to keep pace with accelerated computing and real-time intelligence needs, driving a new focus on AI-ready storage infrastructure.These days, many enterprises are trying to pour as much power as possible into massive graphics processing unit clusters

play_circle_outline Building On-Premises AI Stacks: Addressing Data Control with Co-Location and Neocloud Solutions for Modern Data Management Challenges

play_circle_outline Equinix's role in providing flexible interconnection for various computing needs and environments.

play_circle_outline The need for clean, organized data before implementing AI models and computing processes.

play_circle_outline Importance of establishing sovereign data locations to avoid vendor lock-in and egress costs.

play_circle_outline Engaging a wide perspective in tech education, emphasizing overlap between different technology disciplines.

Info
Transcript

Glenn Dekhayser, Equinix & Alan Bumgarner, Solidigm

Glenn Dekhayser

Global Principal Technologist Equinix

Alan Bumgarner

Director and AI Technologist Solidigm

In this SC25 interview, theCUBE’s Dave Vellante and Savannah Peterson sit down with Glenn Dekhayser from Equinix and Alan Bumgarner from Solidigm to discuss the critical infrastructure required to support high-performance computing and AI. The group moves beyond the hype of GPU shortages to analyze the "warehouse-style computer," emphasizing the symbiotic relationship between massive compute power and the storage required to feed it. Bumgarner explains the logistical challenges of transforming raw, "unpretty" data into the clean datasets necessary for accurat... Read more

explore Keep Exploring

What are organizations in the finance sector currently considering regarding the use of AI and data management? add

What are the best practices enterprises are using to integrate AI across different environments and what role does Equinix play in that process? add

What are the challenges and considerations involved in managing and preparing data for storage and usage, particularly in the context of using AI models to identify duplicates and clean files? add

What are the considerations to keep in mind when starting AI projects in relation to cloud infrastructure and data sovereignty? add

What is the definition of an AI factory and how do different perspectives perceive its value? add

bolt Powered by CUBE AI

Glenn Dekhayser, Equinix & Alan Bumgarner, Solidigm

search

Savannah Peterson

>> Good afternoon HPC fans, and welcome back to lovely St. Louis, Missouri. We're here midway through day, one of our three days of coverage here on theCUBE at SC25. My name's Savannah Peterson, bringing you all the best and brightest and nerdiest and even some outs today with Dave Vellante. That was a cool segment we just did.

Dave Vellante

>> Yeah, I mean, there's just the applications of high- performance computing. It's just fascinating. I mean, just mind-boggling actually.

Savannah Peterson

>> Yeah. Well, and there's so much synergy now between the high-performance computing community, scientists, other people doing cool stuff. You got to have relationships like data and storage and data centers all coming together, which is exactly what our next guests are going to tell us about. Glenn and Alan, thank you so much for being here.

Glenn Dekhayser

>> Thanks for having us. - Hey.

Savannah Peterson

>> Glenn, you must really dig it up here.

Alan Bumgarner

>> This is our second time today.

Alan Bumgarner

>> I know. It's kind of new for me too, so I'm excited.

Savannah Peterson

>> Well, you've done great. Did you secretly just want to sit in Gary's seat?

Alan Bumgarner

>> Yes. Yeah, because I mean, how many times do you get to interview a legend.

Savannah Peterson

>> Right. Oh no, I know. I know. I was absolutely honored to be a part of that situation. So everyone is talking about two things they haven't always talked about in the technology hype cycle. We're actually talking about storage. We're actually talking about data centers. It's gone beyond. We need the AI ready infrastructure to make the magic happen. So I'm curious, Alan, what the relationship is like and what had you bring lovely Glenn to us today? Why is this such an important tension? Not tension collaboration.

Alan Bumgarner

>> If you think about what everybody's trying to do today, and everybody's trying to populate, give as much power as they can to this great big graphics cluster so they can run this stochastic gradient descent calculations and determine the future of the world with AI. But when you break down the mechanics of how all of this stuff really operates inside of a data center, it really turns your data center into a warehouse style computer. When you have a very large graphics cluster and it has to... You can't feed that air, so you have to feed it data. And you have this big network that's connected to a lot of storage where all your data is pretty and it's been cleaned and it's nice and it's sitting-

Savannah Peterson

>> Pretty data. Who doesn't like that?

>> Yeah, your pretty data. - Not

Dave Vellante

>> everybody has it. Most people don't.

Alan Bumgarner

>> So then you have to run this command that pulls it all in.

Alan Bumgarner

>> And once you get all this data in and you can run all these calculations, you converge model layers time at a time in a very tail latency sensitive manner. The more power you can give to this thing, doing those calculations, the faster you can get it done, the more accurate you can make your model, the better the fine-tuning. All of the things that make these models not hallucinate or do things that you want them to do becomes more important. And so you have to have a lot of data, and you have to have enough power to give to this thing over here. And so the less power you have to do over here and the more efficient it becomes, the naturally, the better the calculations come, the more accurate your models become, and all of those things happen. So it's a very close relationship between what your storage array can do to keep your data clean and in a proper global namespace to pull it over to make this model more accurate, to do the things that you were trying to achieve.

Dave Vellante

>> So Glenn, six, nine months ago, a lot of discussions with organizations, particularly in finance, they were saying, we are building our own on-prem AI stack because we don't want to just move all the data to the cloud. We got plenty of data up there. We're using the cloud extensively for AI. But it's going to be too expensive to move all this data, and we want to have sort of control over our own destiny. So we're building our own stack. The problem is there aren't a lot of solutions out there, so we got to build our own stack. Today what I'm hearing from many of those, not all of them... JPMC can hire 1,000 AI engineers and build this stuff out. They've got the chops to do that. But many enterprises saying, well, we looked at it, we have to retrofit our data centers. They're really designed for today. We're sort of rethinking that. So we're thinking about co-location facilities or neocloud. So I'm trying to help me understand where you fit now in this new era between, okay, you've got hyperscalers, you've got the neoclouds, you've got Equinix, who's got great facilities expertise. How should we be thinking about Equinix today?

Glenn Dekhayser

>> When it comes to AI, the way that enterprises, organizations of any size are going to accomplish this, the word is and not or. You're not going to go in the cloud or on-prem or to a neocloud or to the edge. And every use case you do will look different. You might use a different model, which may imply a different provider, which may require different network, different data. So what we're seeing enterprises start to coalesce around is some best practices. So Equinix's role in that best practice is allowing a customer to, in a very agile way, interconnect to all of the things, the edges, the neoclouds, the hyperscalers, their own locations, different regionally and allow for all of the sovereign governance that you implied in your question. But what's very interesting from a data perspective, this event here it's about HPC, we talk about big clusters and lots of power and liquid cooling and all that great stuff. And look, organizations are going to get there. Certainly neoclouds are going to be all liquid cooled, right? They're going to have specialized facilities that all they're going to have is GPUs and they're going to be optimized. Their data centers will be optimized for that. Very, very difficult for a to have their own data center that's like that. It just isn't worth, not just the capital expenditure, but the operation of that data center. It doesn't make sense for them to do themselves. But there's going to be a need for them to have some of this processing power in a place that's sovereign so they can do their own models. They don't want to necessarily do that to a neocloud. Maybe they can't move the data. So like I said, there's an and, and all these different use cases are going to imply different requirements. So what you're going to end up with is companies using GPU resources in a hyperscaler to start your POC, perhaps going to neocloud, running your pilot. Maybe you start running production there. But the one thing that's repetitive across all these use cases is that you need to acquire the data. Where are you're getting that from? All of your edges, all of your IOT devices, all your logs, all your customers. You got to acquire all that data, you got to manage that data in its raw ugly format, the not clean format-

Savannah Peterson

>> Yeah. Not pretty.

>> Not pretty. - Pre-makeup, pre-mani.

Glenn Dekhayser

>> - Not the pretty format is unpretty data.

Alan Bumgarner

>> Soon to be pretty - Right.

Glenn Dekhayser

>> And Nvidia's got all sorts of great software,

Dave Vellante

>> and a lot of companies have really good software

Savannah Peterson

>> to help curate this and get it into these formats.

Glenn Dekhayser

>> And there's processes and known best practices for getting it there. But that data pipeline where you put that data pipeline, that's the repetitive part regardless of the use case. So that becomes a very strategic thing that the companies can do first as they're getting into this journey. And then start to use all these resources and you start to see this real federated or what we're calling distributed AI with all the interconnection that we're enabling. And so this is where we believe enterprises are going, and that's how we've really optimized our network to accomplish that.

>> That very much aligns with the work

Dave Vellante

>> that we're doing at theCUBE Research. In fact, Jackie today posited that the data center is going to become actually that a data center where you store data. You don't necessarily load it up with GPUs, use that existing facility to put your data in. How is your data because you've got that facility, and then use capabilities that say Equinix has, or like you said, neoclouds and clouds, and that's where you're going to do the intense sort of GPU work, as you say, everywhere.

Glenn Dekhayser

>> Yeah. So what Equinix is really trying to do is to provide those things that customers need to consume but may not want to do themselves or can't do themselves. So whether it's a liquid cooled environment for GPUs or whether it's something as basic as tape libraries as a service, providing a sovereign S3 archive for all this AI data, which really is a timely thing given the new shortage of shingled drives that's coming out. So I got to store all this data, I've got all this raw unpretty data or these models, these checkpoints and training. Where am I going to put this stuff if I need to keep it? Because I can't get more drives. Everybody got rid of their tape libraries because they replace it with these high-capacity drives. Oops, I can't buy any more of those and now they're full. What do I do? Well, I've got to evacuate that data. So it's these kinds of things that we're trying to provide these things to be consumed, these building blocks for solutions and interconnection's, big part of that. Some of these other services are part of that. To help customers solve the problems that they're going to face as they get down this AI environment they weren't architected for originally.

Dave Vellante

>> And you can do that economically because you get scale.

>> Oh, yeah. - Yeah, yeah. Absolutely.

Savannah Peterson

>> Well, and you guys are the OGs of doing that.

Glenn Dekhayser

>> So you're definitely the right partner to play in that game. I'm curious because there's so much uncertainty about right now when it comes to people determining their solutions. Is they AI going to be good for us? There's a lot of questions in our industry, as well as a lot of really great hard work that's happening, particularly at shows like this. What's one myth that you wish you could just completely eradicate about? And I'm going to ask you about data centers and I'm going to ask you about storage because I think we're in a place and even the conversations we've had already today, everything that's old is new again. But also we can do things differently now than we could before and we don't necessarily have some of the same constraints that we used to. Alan I'm going to start with you since it's your second round with me today.

Alan Bumgarner

>> So the question was one thing that I could eliminate?

Savannah Peterson

>> If you could just magically educate everyone on one thing about storage right now, what would that be, for all your potential customers who might be listening?

Alan Bumgarner

>> Oh, I have two. So I have to choose.

>> You can have two. I'm really

Savannah Peterson

>> generous today. I'm in the good mood.

>> 1 and 1A. - Yeah.

Savannah Peterson

>> I started with Gary this morning,

Alan Bumgarner

>> but I'll tell you, here's one

Alan Bumgarner

>> thing that I would love to eliminate. I was in New York with Gary probably at the beginning of the year, and he just happened to be the audience on a panel that I was speaking at. And I was on a compute panel, which was very rare for a storage guy to be on a compute panel. And when they got to me, I asked the audience, so how many people in the audience think their data's clean? And one person raised their hand and that was Gary?

Savannah Peterson

>> Or they don't have any data. - Not right.

Alan Bumgarner

>> He runs the Los Alamos National Lab super computer.

Glenn Dekhayser

>> No, but he's not lying. - Very clean data for reasons.

Alan Bumgarner

>> But the thing that we have

Glenn Dekhayser

>> that I see all the time in storage is I have... And now there's AI models that will go do it. You can go and look at all of the data inside of your environment and find duplicates, and you can find bad files and you can do all of these things with it. And now you've got these pretty object stores where I have as much metadata per object as I want to. I can manipulate all of these things. And I think there's this great big growing argument on how do you keep your data, how do you keep it clean? How do you make it as compact as you can? And what are the things that you want to go do to make it ready to be used? And I think that you'll see a lot of people in environments right now before they even get to the point where Glenn can ingest all of their stuff and let them use these sovereign AI data centers that Equinix builds so eloquently, is you have to go fix all of those problems because you don't want to upload duplicates, you don't want to spend all the money to do all of these things. For me, if it was just an easy way to do that on ingest and vectorize it immediately, I think that would be what I'd do.

Savannah Peterson

>> I love that. Glenn, what's something you wish more people knew?

Glenn Dekhayser

>> That I think the myths that we believe anyway is that-

Dave Vellante

>> Just us. - ...

Glenn Dekhayser

>> you should go all cloud with your AI endeavors from the beginning and then move it out later. Because I'm not saying don't use the cloud. Cloud has an absolute imperative purpose and should be used in those first use cases. But if you don't think upfront about establishing a sovereign place for that data to be, you're going to run the risk of getting locked into a platform that's very difficult to get out of later from a data perspective, especially if you start using interfaces and APIs that are opinionated. So that's the problem is that getting the data out, everybody focuses on the egress costs and sure that's painful and eventually that will be the main driver of you probably doing what we're telling you to do upfront. But doing it upfront, building the beginnings of your AI factory, it doesn't need any GPUs. That's the myth is that you can start an AI factory with just storage and just compute for doing your data curation and pipeline management because that's the beginning of it. You can send that data to wherever the GPUs you want to use are and have at it, right? That's the flexibility. But you start there because you're starting sovereign and you're now in a position of flexibility and leverage.

Dave Vellante

>> Wait, explain that further. So you're saying you can start without the accelerated computing piece of it.

>> Correct. - But with a vision that you're going

Dave Vellante

>> to eventually add

Glenn Dekhayser

>> that in. Is that right, or not necessarily?

>> Well, no.

Glenn Dekhayser

>> You're going to use it, but you don't need it in that sovereign space right away. When you're starting your first use cases, you're experimenting, you're trying what models and whatever, but your data is here, you can project that data. Most of the storage companies that are out there today have the logic built in to project data to mirror it, replicate it in very consistent ways to different cloud platforms. You're looking at the Dells and NetApps, the pures of the world, they all do this. This is table stakes. And even if they don't, you want to use a platform that doesn't have that. There are third-party tools that can move that data in a consistent way up to the cloud. So move it to the cloud, use those GPUs. Move it to the neocloud, use those GPUs. Here's the good news. You've got your sovereign copy. There's no egress. You just delete it out of the cloud. You're just delete this. So you never have to worry about egress ever if you have that plan and start the right way.

Dave Vellante

>> So you talked about AI factories a little bit today. What is an AI factory to you? I mean to us it's like you're producing intelligence. And how is it different from the traditional data center and how do you guys, maybe it's not migrate, but balance the traditional workloads and the accelerated workloads?

>> It's a great question because I was at GTC DC a couple

Glenn Dekhayser

>> of weeks back, and Jensen defined AI factory as a factory it creates, what? Tokens.

>> Creates tokens, yeah. - And I think

Glenn Dekhayser

>> that's one perspective if you're in neocloud,

Dave Vellante

>> because yes, you, you're literally measuring it by tokens, how much power you're putting in and how much tokens you're getting out. That's one kind of AI factory. I think enterprises see a little differently. They don't care about the tokens, they care about the business outcomes. That's all they care about. So that's why an AI factory, that's why when I say can start with just the storage, that's what I mean. You might add GPUs as you want to do training and inference either here or perhaps we've got 277 data centers in 77 countries now, 77 metros and 37 countries now. So running inference, some inference is not latency sensitive. It can be run in a central place. A lot of it is latency sensitive. It's not all about chatbots.

>> No. - So any kind of video or real time stuff.

Glenn Dekhayser

>> So having that distributed inference is extremely important.

Savannah Peterson

>> Being able to deliver GPU power at these edges, you don't need huge super pods for that, you're looking at smaller AI factories. And the floor here has so many... It is really cool to watch so many distributed AI companies that don't know the distributed AI companies. They're doing distributed inference, distributed training, distributed RAG.

Savannah Peterson

>> You are absolutely right.

>> Even companies like Redis that I saw last Friday

Glenn Dekhayser

>> that can do caching and then distributed caching. The opportunities for distributing information and distributing this different stages of inference around to make things very efficient. There's so many opportunities to make things a lot faster and a lot more efficient across the board. But also to create those business outcomes. It's not just about, oh, I'm going to go get a super pod and see what model I can create. Some people need to do that, pharmaceuticals, some people who need that bespoke private model. But most enterprises, they want the outcome. They'll take a foundation model or an open source model, they'll post-train it, and then they'll build a very sophisticated RAG because you can't put that temporal private data into a model. Can't do that. It doesn't come out. So you need to be able to get things out and manage that data. This is classic IT. It's what it really is. This is stuff we've been doing for a long time with ERP, with data warehousing, it's just a new model. And so you need really performant storage. This is where I see really cleaning up in this world. Being able to deliver that kind of high throughput data to local inference at the edge. It's going to be, I think a great play as this all unfolds.

>> That's where the data is. - Data is everywhere.

Alan Bumgarner

>> Data is everywhere. - Data definitely is everywhere.

Savannah Peterson

>> Glenn, I totally agree with you.

Glenn Dekhayser

>> I'm glad you brought up that Jensen point

Dave Vellante

>> because like you, I slightly disagreed and that was not my personal definition. And I think that-

Dave Vellante

>> It's narrow. It is narrow.

Savannah Peterson

>> It's narrow. - I mean it's huge, but it's narrow.

Dave Vellante

>> It's based on your perspective.

Glenn Dekhayser

>> Yeah. I say this with so much love,

Savannah Peterson

>> it's a classic tech nerd response. People aren't spending money for tokens. They're spending money to be more successful. They're spending money to go deliver the thing or build the thing or solve, cure the cancer. They're not spending money because... Most people aren't thinking about it that way. The tech community thinks about it that way, but I mean, he's in front of the entire public sector there at GTC, it's just kind of a very different conversation. All right, last question for you guys to close this out. And I'm going to have to spice it up, Alan, because I asked you my usual last question.

Glenn Dekhayser

>> You're in trouble, man. - All right.

Savannah Peterson

>> We're going to take it a fun direction.

Alan Bumgarner

>> What would you tell a new person to our community... Community has been a bit of a theme today, which is really nice. And someone who's a student at one of these universities or someone who's just starting to get involved in our space, doesn't matter. They're age agnostic. What would your advice be to them about coming to play in the HPC AI space right now?

>> Oh, man. I wanted that question.

Glenn Dekhayser

>> Both get it. - I just had this conversation.

Alan Bumgarner

>> Oh, you did

Savannah Peterson

>> just have this conversation. Well, what'd you say?

Alan Bumgarner

>> I had a student from Stanford and one from Colorado State

Savannah Peterson

>> and one from somewhere else, I'll remember. But the interesting discussion that we had is we were looking at do you want to be in systems administration or do you want to be in chip design? And the fundamental difference between those two is really not that far apart. Because one of the things that I do in part of chip design when we work with that team internally is I have to go look at all the workloads that are out there and I have to understand all the system administration that's happening. And it's this one big kind of ecosystem that you really have to have kind of a wide view to do some very narrow things. And so the advice that I tried to give this team was when you're thinking about how this works over here for system administration versus what am I going to use from this knowledge to make a chip someday, three years from now, that's going to do something very specific. If you can have a view, a wide view, of the world and think about it that way, and then pick which discipline speaks to your heart the most. But it doesn't matter because even the person over here in system administration needs to understand how that SSD or the chip side of it works because they have to administer it. So it's kind of connected. And my advice was make sure you have options when you're walking through that stage in your life when you're picking what you want to be when you grow up. But don't narrow it down too much.

Savannah Peterson

>> And realize that maybe there's more overlap than you might think.

Alan Bumgarner

>> There's a lot more than you think. Don't let somebody peg you into a specialty. That's not how the real world works.

Savannah Peterson

>> As a proud generalist, this is music to my ears.

>> Here. Here. - Yeah, yeah. Here, here.

Savannah Peterson

>> Glenn, what would your advice be?

Glenn Dekhayser

>> All right. So this isn't Equinix marketing,

Alan Bumgarner

>> when someone asks me who's not familiar with who Equinix is, I like to tell them that Equinix is where stuff, I don't use stuff, gets real. And the reason I say this is I've met software engineers, brilliant kids who are, some of them work for hyperscalers, some of them work in the financial world, who have never seen a physical server, who have never seen a data center, who don't understand power, who don't understand the stuff that they write-

Savannah Peterson

>> Haven't felt the heat.

Glenn Dekhayser

>> They have no idea- - The wind.

Alan Bumgarner

>> Yeah. Quite literally. - Like the ramifications, the

Glenn Dekhayser

>> consequences of the code that they're writing of

Savannah Peterson

>> what the physical manifestation of that is and where it runs. And so very often I do tours for these kinds of folks, I'll bring them to the data center and I'll give them, and if you see their eyes light, they can't believe trillions of trillions of dollars of economic activity going on and it just looks like fans humming and water flowing and whatever and lots of blinky lights. But it's important to understand that as you're getting into this business, because there's so much more than just sitting in front of a laptop and coding. Now, it's not that it's not fun. I love it. I've done my share of it. It's fun. You type the command and the computer does something that you told it to. And how many things in this world do what you tell them to? So that's a powerful feeling. Exactly. But the reality is they had no idea what had to happen for that to be real, for that to happen because they don't see it. And I think if they saw it and we taught them and we brought them more into that world, there'd be a greater appreciation for what they were doing and perhaps a better utilization, better architecture, better engineering out of the whole thing because really we'd like to throw around the words architecture engineers in the IT world a lot. But we really are building a big machine, the software is part of that, but it's a big machine and it needs to be architected like a machine and run like a machine. So I think that's something I would love to have more prevalent in the knowledge base of these kids coming out.

Savannah Peterson

>> Well, thanks to this interview today, there will be more prevalence of that, Glenn. Thank you guys both so much. This was a real fun one.

Glenn Dekhayser

>> Appreciate it. Thanks. - Thanks, guys.

Savannah Peterson

>> Thank you. - Yeah, good to see you again, Glenn. And good

Alan Bumgarner

>> to be here with you, Dave, at one of the nerdier,

Dave Vellante

>> but more fun events that we get to do every year. I hope you're all having as much fun as we are here in St. Louis, Missouri at Supercomputing 2025. My name is Savannah Peterson. You're watching theCUBE, leading source for enterprise tech news.