SC25 | Sudip Dosanjh, National Energy Research Scientific Computing Center (NERSC) & Tony Rea, Dell Technologies

Clips
More from SC25

Sudip Dosanjh

Scientific Division Director

National Energy Research Scientific Computing Center (NERSC)

Tony Rea

Global AI Infastructure Lead, Global Tech Officer

Dell Technologies

pause_circle_outline Empowering Women in Tech: Highlights from AWS Museum Event at Supercomputing 2025 with Speakers Dave Vellante and Jackie McGuire

play_circle_outline Sudip Dosanjh's role at NERSC and its mission related to high-performance computing.

play_circle_outline Enhancing Scientific Workflows: The Doudna System's Role in Managing Data and Real-Time Computing Collaborations with Dell and NVIDIA

play_circle_outline Overcoming Challenges in Liquid Cooling: Enhancing Energy and Water Efficiency for Sustainable Supercomputer Infrastructure

play_circle_outline Notable achievements, including 14,000 publications and two Nobel Prizes from NERSC users.

Info
Transcript

Sudip Dosanjh, National Energy Research Scientific Computing Center (NERSC) & Tony Rea, Dell Technologies

Sudip Dosanjh

Scientific Division Director National Energy Research Scientific Computing Center (NERSC)

Tony Rea

Global AI Infastructure Lead, Global Tech Officer Dell Technologies

In this interview from SC25, NERSC Scientific Division Director Sudip Dosanjh and Dell Technologies Global AI Infrastructure Lead Tony Rea join theCUBE’s Dave Vellante and Jackie McGuire to discuss the collision of AI and high-performance computing. The conversation centers on the upcoming deployment of "Doudna," NERSC’s next-generation supercomputer designed to handle the paradigm shift from traditional simulation to complex, data-driven workflows. Dosanjh explains how the system addresses the deluge of data from global sensors and telescopes, enabling real-... Read more

explore Keep Exploring

What is the significance of the venue where the event took place? add

What is the mission of NERSC and what types of research do they support? add

What changes have been observed in the data processing and computing needs at NERSC? add

What type of cooling system is being implemented in the IR7000 rack, and how does it contribute to sustainability? add

What impact has the use of GPUs at NERSC had on scientific research and achievements? add

bolt Powered by CUBE AI

Sudip Dosanjh, National Energy Research Scientific Computing Center (NERSC) & Tony Rea, Dell Technologies

search

Dave Vellante

>> Hi everybody. Welcome back to St. Louis. My name is Dave Vellante and I'm here with Jackie McGuire. And we are inside of Supercomputing 2025. This is day four of the conference, day two of theCUBE's wall-to-wall coverage. Jackie, the museum last night was 2C?

Jackie McGuire

>> 2C. Yeah.

Dave Vellante

>> Pretty amazing inspirational-

Jackie McGuire

>> Shout AWS. That was a really cool venue.

Dave Vellante

>> Inspirational art sort of celebration of women and some really interesting artifacts there, which were quite remarkable. But anyway, we're here to talk AI supercomputing. We've been waiting for this moment for years, maybe even decades. And those two worlds have collided. Sudip Dosanjh is here. He is the scientific division director for the National Energy Research Scientific Computing Center, otherwise known as NERSC. And Tony Rea is back. He's the Global AI Infrastructure Lead for Global Tech at Dell Technologies. Gentlemen, good to see you. Welcome.

Tony Rea

>> Good to see you too. Good morning.

Sudip Dosanjh

>> Thanks for having us here.

Dave Vellante

>> Yeah, good to see you back again on theCUBE. Appreciate your support. Show is quite amazing, I was saying. I remember in the late '80s thinking, "Wow, AI is here. High-performance computing." I was talking last night to somebody about Thinking Machines and Kendall Square Research-

Tony Rea

>> Oh, Danny Hillis' company.

Dave Vellante

>> I have a Thinking Machines T-shirt still.

Tony Rea

>> Right. So do you know who wore a Thinking Machines T-shirt? Richard Feynman. He was actually an employee for a while at Thinking Machines. Danny knew him through Marvin Minsky at MIT.

Dave Vellante

>> It's kind of tattered, and I'm like, "Do not throw this out," to my wife-

Tony Rea

>> At this point, it's getting old.

Dave Vellante

>> Pretty cool times. But there was this supercomputer renaissance back then, but the AI just didn't follow. I guess it was too expensive, didn't have enough data. But Sudip, here we are in 2025. It's

Sudip Dosanjh

>> Amazing. Yeah, thank you so much for having me. I'm really excited to be here. I'm always thrilled to talk about the great work that people at NERSC are doing and also our users who do amazing science. And so the mission of NERSC is to accelerate science at the open science within the Department of Energy. So we focus on things like high energy, physics, materials, chemistry, fusion, biological and environmental research. So there's a broad range of applications that run at NERSC and they do really amazing work. And because with their mission computing center, we don't select our users, so we can't deploy a system and then look for users that might benefit from it. So we have to take our vast user community, over 11,000 users with us as we deploy every system. And so it requires really a close collaboration with Doudna, which we're planning to deploy at the end of 2026. It really requires a great collaboration with Dell and with NVIDIA because there are a lot of new capabilities that we're going to be deploying within the system. And so it's really, really very exciting.

Dave Vellante

>> Well, let's get into it. What is Doudna? What's the announcement all about? What's its vision?

Sudip Dosanjh

>> Yeah, so we've traditionally had lots of people doing different types of applications at NERSC, but really the focus often has been on modeling and simulation in the past. And one thing that's changed is that there's more data coming into NERSC than leaving NERSC, which is like... It's a paradigm shift for supercomputing center. Supercomputers were thought of, back in the day, that this box that sat by itself and it had data that was local and you processed it somehow. And now people have data from all over the world, doing experiments, observational science, from light sources, telescopes, genomics facilities, and they're being inundated with data. There's a flood of data. Sensors are getting better, sequencing is getting better. And so there's just this tremendous deluge of data that scientists are facing. And what they want to do is transfer that data to NERSC and analyze it using AI, other data analysis techniques, but it's not just about AI. It's really about these very complex workflows that they want to run because they may want to transfer the data, run a simulation at the same time to help steer their experiment. So one thing that's different is they want feedback in real time. You have a telescope, you need to know pretty quickly when to refocus that telescope. If you have an experiment that's running, you need to get information back fast enough for you to be able to make a change to your experiment. And so if you tell people, "Oh, you need to run in a queue and you'll get your answer back in a few days," that's not so good. And so there's this greater emphasis on near real time computing, bringing in data from the outside, being able to stream it in from wherever it's at, and being able to run these complex workflows, being able to orchestrate all that and the storage. So the vision for Doudna is really enabling these very complex scientific workflows and providing a 10x increase in performance. And we expect to do that through the collaboration with Dell and with NVIDIA. It's really very exciting.

Dave Vellante

>> We heard, I guess it was Monday night, Ian Buck held a session of NVIDIA and he talked about this world transitioning from what has always been a batch world to kind of real time, which has been hard because the system's always so fully utilized. So interesting. Okay, so you're basically harmonizing all this simulation data. Is that right, in providing service?

Sudip Dosanjh

>> Yeah, all this experimental observational data maybe from telescopes. You have large databases sitting at NERSC that they may want to compare with. You have an image from tonight of the night sky and you want to compare it to historical data. You may have a fusion reactor that's running and you want to be able to do a simulation of the fusion reactor as it's actually running and be able to steer what the experiment is doing at the same time. So there's just a wide range of different types of workflows that people want to use.

Jackie McGuire

>> Well, that's a pretty symbiotic relationship too, right? Talking about nuclear fusion, like nuclear is one of the things that we really need to get to commercial viability to be able to achieve the super compute at scale, right?

Dave Vellante

>> Right. Right.

Jackie McGuire

>> It's kind a flywheel of-

Dave Vellante

>> Yeah, if you can accelerate that.

Jackie McGuire

>> I literally just listened to a piece this morning about the administration's endeavor to get four companies to... I'm trying to remember what it's called. It's basically like the... Is it viability... Criticality with nuclear .

Tony Rea

>> I just want to make a point and that one of the things this machine's going to be able to do is digest enormous amounts of data. And part of the problem is that the instrumentation, a lot of the instrumentation and all these domains being used, the granularity and the resolution is going up and the amount of information is going up by an enormous amount every year. And that makes the computational power needed to digest that stuff and actually do it real time. A machine like this is really necessary to do that.

Dave Vellante

>> So Sudip mentioned that you have to have deep partnerships with companies like Dell and NVIDIA. What is the system that you're architecting? What are you providing?

Tony Rea

>> So this is a highly dense system. It's our IR7000 rack. It is going to be a Vera Rubin based machine. It is going to be all water-cooled. The interesting thing about this is we worked very closely with NERSC. This is going to be warm water cooling. So there is going to be a sustainability. We're looking at the sustainability angle of the cluster and that providing warm water is going to make it so it's more efficient.

Dave Vellante

>> Like 40 degrees C?

Tony Rea

>> Yes.

Dave Vellante

>> That's what you're talking about here?

Tony Rea

>> Yeah.

Sudip Dosanjh

>> And the amazing thing, we're located in the hills above UC Berkeley. And so it's always cool. If it gets to be 74 degrees, we start to complain about how hot it is. We really have no chillers in the building. Because of the warm water, we can actually do it all with air cooling-

Jackie McGuire

>> Evaporative?

Sudip Dosanjh

>> And with evaporative cooling. And the challenge I'm putting before the community is that it's not just about energy efficiency. So we have amazing energy efficiency in the building thanks to this great design from Dell and the fact that we're doing all air cooling... We're just using the Bay Area climate to cool the system. Our PUE is usually 1.06 or something. So we have amazing energy efficiency. But it's also about water. So it's not just about... If you want to be really sustainable, especially in California, you have to worry about water as well. And so we're installing some dry chillers, which is very new for us, but that'll greatly reduce the water usage. So instead of using evaporative cooling, we'll be using these dry chillers. It'll increase our PUE slightly, but the trade-off will be a tremendous savings in water.

Jackie McGuire

>> Yeah, because you can kind of bring the dew point.

Sudip Dosanjh

>> Yeah.

Tony Rea

>> One of the things about these dense computer systems are extremely elaborate, require tremendous amount of engineering, but also what requires a lot of engineering is actually the plumbing for the liquid cooling.

Jackie McGuire

>> All pumps and the compressors-

Tony Rea

>> That technology is still evolving and Dell is doing a lot to make it as standardized as possible on that, so customers don't have a lot of problems getting their systems installed in their infrastructure.

Dave Vellante

>> So you're talking about the secondary fluid network and how that connects to the direct-

Tony Rea

>> Primary and secondary networks, and even the kinds of fluids that are in them and how you control that and how you route that and using standardized couplings.

Jackie McGuire

>> Well, I was just going to say, even things like hose couplings, where power jobs are.

Tony Rea

>> Let me tell you, if you walk around this show, you're going to see hundreds of types of couplings. And so at Dell, we have to digest all that and determine what would be the best to deploy with a customer. And we worked a lot with NERSC to do that.

Dave Vellante

>> We stopped by the Omni booth yesterday I think, and just looked at the different hose types, the connections. Where are we with regard to the connection integrity?

Tony Rea

>> We have cooling labs at Dell and we have-

Dave Vellante

>> I've seen the Lab in Round Rock.

Tony Rea

>> Yes. But we multiple labs and we test a lot of this. I can't tell you how many different couplings and hoses and stuff we're testing and we computer simulate this, and then we also physically test it in our labs, and then we validate it for reliability and then we make it good for the customer.

Jackie McGuire

>> Well, Dell's part of the Open Compute Project too, right?

Tony Rea

>> Yes.

Jackie McGuire

>> And that's part of the Open Compute Project's-

Tony Rea

>> The OCP, yeah....

Jackie McGuire

>> mission is to standardize how data-

Tony Rea

>> It's our rack. Our IR7000 rack is an OCP rack.

Dave Vellante

>> This is such an important topic that not a lot of people talk about. It's kind of like the unsung heroes of the data center or the AI factory. You think about the tragedy of the Challenger that exploded because of an O-ring that had to withstand... It wasn't built to withstand conditions at Cape Canaveral where the temperature dropped below 40 degrees, something like that.

Jackie McGuire

>> They built it in Texas.

Dave Vellante

>> You can't have-

Tony Rea

>> There was Richard Feynman also on that one. Yeah.

Dave Vellante

>> You can't have that in a... I don't know what's it cost to build these giant data centers or-

Tony Rea

>> Yeah, it's also-

Dave Vellante

>> $30 billion.

Tony Rea

>> But to your point, just the O-ring and ceiling technology that goes into this is just incredible. The sophistication of this stuff is just... It's almost like as sophisticated as the GPUs at this point.

Jackie McGuire

>> I happen to be married to a commercial chiller expert, and we started a company together years ago, and yeah, just looking... And the small number of people who understand the complexity of these technologies and can link how they physically work to how they're built. And that's one of the bridges that I think we need to continue building is between design and application. Because there's what works in a lab when you're designing a chiller, but then there's once you get it on the ground and you have to maintenance it, you have to replace parts-

Tony Rea

>> Exactly like that....

Jackie McGuire

>> you have to... Yeah, and to your point, everything from the welds-

Tony Rea

>> It's very complex....

Jackie McGuire

>> to the holes to the rivets to... Yeah.

Sudip Dosanjh

>> Yeah. I think that a lot of people don't appreciate how complex it is, what we do, bringing the entire enterprise together. There's the system, there's this huge infrastructure for power and air and water. And then there's a lot of work that goes on in working with the users as well as far as getting them ready for these systems so that they can effectively use them. So providing user support. We have lots of tickets. We work with, obviously, the companies like Dell and NVIDIA to design the systems and to deploy them. But then we also have to get our users ready for them. And that involves things like training and education because we have a very large user base. And that's where the partnership with NVIDIA has really been critical, has been getting our users ready for each of these systems. When we were deploying Perlmutter, people questioned, given NERSC's mission, "Are you going to deploy a big GPU system? Are the users really ready for that?" And I said, "Yes, of course." But it really required a great partnership with NVIDIA to get users ready for the system, get them on the GPUs. Now we have thousands of users using the GPUs. We've had 14,000 publications since 2020 that have referenced NERSC that have been due to Perlmutter and the GPUs. We've had two Nobel Prizes. We expect more, but our scientists... I've lost count of how many,

Jackie McGuire

>> No big deal.

Sudip Dosanjh

>> Nobel Prize users have won deal. But the other thing that's happening is just the pace of science is accelerating, the pace of what we're doing is accelerating, the pace of what Dell and NVIDIA is doing. So it's just a very rapidly moving world, and it's a huge enterprise to bring this all together.

Tony Rea

>> A machine at this scale enables new research. In other words, the researchers now can solve problems they couldn't solve before. So now it's-

Dave Vellante

>> Let's talk about that. What is the scientific impact, this sort of open science model that you guys have? You mentioned 14,000 papers, a couple of Nobel Prizes. What kinds of really interesting things are you guys doing? I mean, climate, material science. I mean, share with the audience-

Tony Rea

>> Mentioned the telescope-

Dave Vellante

>> Some of the cool stuff.

Tony Rea

>> That's pretty cool.

Sudip Dosanjh

>> Yeah, there's just a huge range of different applications. There's this DESI, this dark energy spectroscopic instrument. If you look that up, they've had images that are just mind-boggling. It's looking for dark energy. Actually, if you go look at the New York Times, it was on the front page of the New York Times, some of the images that they generated, all of those images were generated at NERSC. Streaming the data in and analyzing it, and then oftentimes, refocusing the imaging technology based on the analysis that they've done. So this type of real-time feedback. Because you don't get another chance, you'll never see that night sky again. This is the one chance you have.

Jackie McGuire

>> And it's some of the infrared images too, where the human eye can't see it, so you previously had humans trying to do a whole bunch of math based on infrared readings, and now you can use AI for that, right?

Sudip Dosanjh

>> Yeah. Yeah. So David Baker, part of his Nobel Prize was using AI to design materials and proteins. And so it's just amazing work that he's done. He's been one of the biggest users at NERSC over the last four or five years. And so I think what used to be the case, like Saul Perlmutter won a Nobel Prize, he was one of our... He is one of our users, actually, for co-discovering that the rate at which the universe was expanding was accelerating, which was a big thing at the time. At the time, that was totally unexpected.

Dave Vellante

>> And it was opposite of what some people were thinking.

Sudip Dosanjh

>> Yeah, he was the first person to do big data analysis at NERSC. So there's just this wide range of kind of scientific impact. But he won a Nobel Prize for things he'd had done quite a bit earlier. Now people are winning Nobel Prizes for what they did in the last couple years. And so that's really where I see the impact we're going to have. It would be unheard of that, "Oh, we had a user..." One of our big users on Perlmutter and was doing this work in 2021, and it helped... It wasn't the only reason he won a Nobel Prize, but it helped him win a Nobel Prize. And so the pace of scientific discovery is really accelerating. And I think part of that, that's where high-performance computing, AI and data analysis can really even further accelerate what's happening in the scientific community.

Jackie McGuire

>> I keep being hopeful that science will outpace the rate at which I'm deteriorating my body. You hit 40 and then all of a sudden you start-

Dave Vellante

>> Ray Kurzweil says that's already happening. He says, "Each year, we only get, I don't know, 80% older now."

Sudip Dosanjh

>> So Jennifer Doudna, who the system is named after, she's a UC Berkeley scientist who helped develop CRISPR, which have-

Jackie McGuire

>> The DNA. Yeah.

Sudip Dosanjh

>> Potentially wide impact in lots of parts of our lives. That was very recent. And so she does have an account on Doudna... Or she will have an account on... She has a account on Perlmutter, but she'll have an account on Doudna. A lot of it is... Usually the grad students and the post-docs are the ones who are doing-

Jackie McGuire

>> Ones who are still curious.

Sudip Dosanjh

>> She has an account on Perlmutter, so...

Jackie McGuire

>> That's awesome.

Dave Vellante

>> So kind of a macro question. If you had to plot the curve of scientific discovery over the last 40 or 50 years, was there a pattern that you could discern? And has that curve, in your opinion, been altered at all by this new era that we're entering, the compute power that we now have?

Sudip Dosanjh

>> Yeah, I would really interested in an analysis. I haven't seen someone analyze people winning a Nobel Prize and how long ago that work was. But my sense has been just watching... We all watch every year, every time that something is awarded, it seems like it's getting shorter. And so that's my personal feeling, that I can just point to lots of examples where the time delay, it's no longer decades, it's years or a year. And you never used to see that. And so I'd be really interested in a more-

Tony Rea

>> So the barrier of entry to do scientific research is much lower now. And also at NERSC, they're also... I'm taking some of your... But they're doing a lot to enable the researchers to be more productive, to understand how to better use these machines. And this is pretty widespread in a lot of the clusters that are being used today.

Jackie McGuire

>> That hopefully makes science more equitable.

Sudip Dosanjh

>> So for us... So it's not just the... We love these Nobel Prizes, we talk about them, but really, it's this broad impact. It's having deep engagements on particular scientific topics, learning from that, but then making that information available through training and education and hackathons and these collaborations with Dell and with NVIDIA, making that information available to 11,000 users so that we can have 14,000 publications in five years. We have to do both of those.

Dave Vellante

>> And the comes down... That's because you can get infrastructure that doesn't melt GPUs and is on time and can take advantage of the... Hopefully. And doesn't leak. And all the good work you guys are doing. Sudip and Tony, thanks very much for coming to theCUBE today, really appreciate it.

Tony Rea

>> Sure.

Jackie McGuire

>> Thank you.

Tony Rea

>> Thank you.

Sudip Dosanjh

>> This was fantastic. Thank you.

Dave Vellante

>> You're welcome. And thank you for watching. This is Dave Vellante, for Jackie McGuire, the whole CUBE crew. We'll be right back right after this short break from Supercomputing SC25, from St. Louis. Keep it right there. TheCUBE, the leading source for technology news and analysis with exclusive access to the hottest technology events from around the globe. In depth...