Nvidia’s Chips Power the Supercomputer That Could Change AI
17:24 Share

This episode is brought to you by Shopify. Forget the frustration of picking commerce platforms when you switch your business to Shopify, the global commerce platform that supercharges your selling wherever you sell. With Shopify, you'll harness the same intuitive features, trusted apps, and powerful analytics used by the world's leading brands. Sign up today for your $1 per month trial period at shopify.com slash tech, all lowercase. That's shopify.com slash tech.

Late last year, at a ceremony featuring both the King of Denmark, Frederick X, and the King of Artificial Intelligence, NVIDIA CEO Jensen Wang, a very powerful supercomputer was plugged in for the first time. Hello, humans. My name is Keefian. I am a supercomputer created to serve and assist you. I've been trained on billions of data points, and I learned from you.

Standing between those two kings was the woman in charge of this AI supercomputer, Nadia Carlsten. She's the CEO of the Danish Center for AI Innovation, or DCAI. It was established as a joint venture between chipmaker NVIDIA, the Novo Nordisk Foundation, and the Export and Investment Fund of Denmark. The DCAI gives both public and private sectors in Denmark access to Gefeon for ambitious, large-scale AI projects that could have global impact.

Think new drug discovery, mass weather prediction, and modeling molecules. Carlston, who spent much of her career in tech management and quantum computing, has high hopes for Gefion and its 1,528 GPUs, or graphics processing units, the advanced computer chips NVIDIA is known for. There's so much you can do. It's almost the question should be, what can't you do with it?

From The Wall Street Journal, this is the Science of Success, a look at how today's successes could lead to tomorrow's innovations. I'm Ben Cohen. I write a column for the journal about how people, ideas, and teams work and when they thrive. Today, I'm speaking with Nadia Carlston and Mads Kruisgaard-Thompson, the CEO of the Novo Nordisk Foundation, about Gefeon, how it got built, and why they believe it's plowing a road to the future.

I started by asking Thompson to explain exactly what differentiates an AI supercomputer from any other supercomputer and why the Novo Nordisk Foundation wanted to build one. As a foundation, we did a survey some time ago in the Danish life science ecosystem about what are the biggest stumbling blocks in terms of our ability to do research based on data-intensive areas. And the short answer was,

strong compute power. Denmark didn't have access to such a Gen-AI, Gifion-like computer, and that was a strong desire, both among academics, clinicians, and even among the companies. But I think I'll let Nadja address the question about the difference between the computers.

There are many supercomputers in the world. The top 500 list ranks all of the supercomputers in the world and Gifion in November ranked number 21 on that list. Gifion is optimized for AI, so that makes it special and that's why we call it an AI supercomputer and not just a supercomputer.

Everything is designed from the architecture to the networking to make sure that AI, especially AI training, is happening in the best way possible. So for us, it means everything from the selection of what type of processors to use. So Giphion is based on the GGX SuperPOD system from NVIDIA. We're using the H100 chip from NVIDIA, and that is designed for AI workloads specifically. One of the benefits is that when you have workloads that demand

parallel processing of huge amounts of data. That's what AI computers are super good at or optimized for also. And I have to say, it's not only the acquisition of the Gephion computer, it's actually also a strategic partnership with the NVIDIA company where we benefit from their super experience within R&D. And of course, they learn from the little Danish kingdom where everything is relatively agile and moving fast in terms of how they should develop maybe their next generation of software and so on.

Mads, I want to talk about parallel processing and huge amounts of data and how that applies to the work that you do. How long have you been in the field of drug discovery and research and development? Oh, don't ask that question. I've been there for more than 30 years. I was at the New North Company for 30 years and I was head of research and development for more than 20 of those 30 years.

From the early days where so-called rational drug design was done with the 3D glasses on and looking into a simple computer screen with a very low resolution all the way up to, you can say, almost Gen AI. I think I left Nukunjordiske company just before Gen AI came of age. So these are new times.

Before this AI supercomputer, how were you already using AI in that work? Yeah, so we've used machine learning. So one of the drugs that is going to undergo approval right now, actually for hemophilia, that was the first example of the company using machine learning in the design of the two specific arms of that particular antibody, how they should interact with each other to optimize the binding to the drug target.

So to be honest, there's been use of rational drug design, molecular modeling, 3D based in the old fashioned way with classic computing for many years, but most often it didn't work. The first time we had a real success, and that's a quite big one, was actually using these machine learning algorithms, not with the Gen AI computer, but just with standard high performance computing that probably cut the process from idea for this new hemophilia drug until it was ready as a drug candidate.

by a couple of years. We did it in about two years rather than normally you do it in four years.

And Mads, what is the order of magnitude difference between what was possible before and what you are hoping will be possible now? Well, I would say it is orders of magnitude. I can't put a number on, but I can say you have iterative cycles going from chemistry to testing biology, going back to chemistry, iterating, making new molecules, testing them. Now, this whole thing was very sequential in the old days. Today, we can do that

In parallel, in the labs, we can even automate the synthesis of compounds and using these machine learning algorithms, we can do something that typically would take

maybe two to four weeks in the old days, one iterative cycle. We can do that in a few days, maybe two days even. A chemist can only, in his or her brain, encompass so and so many molecular structures and have ideas about what to do. You can feed the computer with all the data you have in your database, and it will come up with a hundred suggestions for molecules to make.

And you start making them. And already after you've done 10 or 20, the biological data comes back into the computer and he starts realizing, oh, I didn't get it right the first time. This was good. This was not good.

And then it immediately changes its strategy, so to speak. And that can happen within a few days when the old days, before the biology data came back, the chemists already done new molecules that went in the wrong direction. So it is a smart way of working. Nadia, before we look ahead, let's look back a very long time. Before you took this job, did you have any idea who Geffion was?

I did not. I found out who Gephion was when I moved to Denmark and I asked the team, how did you come up with the name Gephion? And they told me the story. All right. Tell me the story. What is the story of Gephion? Who is this? What is this name? Yeah, I think it's a beautiful story. Gephion is a mythical goddess who was told by the king that she could have as much land as she could plow, but only in one day and one night.

So Gifjorn, being a very resourceful woman, decided to turn her four sons into oxen and use them to plow the land during that time. And they were able to plow so much that they created the island that is now the central part of Denmark. So that's the legend. And I think it's just wonderful and in many ways a great name for our computer.

Yeah, Mads, how did Nadia's retelling of Norse mythology sound to you? Fantastic. I couldn't have done it better. Can you describe, Geppion, for me, how big is it? What does it actually look like to people who have never seen an AI supercomputer? It's very large. So it's the size of at least a basketball court. It's inside a commercial data center. So very few people have actually seen it in real life.

It is entirely based on the TGX system, so it actually looks really nice. It has the trademark NVIDIA gold boxes on the inside and you can actually walk through rows of them. It's very large. It's 1528 GPUs all assembled into one large super pod. When we come back, the future that Gefion, the AI supercomputer, could plow for Denmark, AI innovation and the rest of the world.

This episode is brought to you by Indeed. When your computer breaks, you don't wait for it to magically start working again. You fix the problem. So why wait to hire the people your company desperately needs? Use Indeed's sponsored jobs to hire top talent fast. And even better, you only pay for results. There's no need to wait. Speed up your hiring with a $75 sponsored job credit at indeed.com slash podcast. Terms and conditions apply.

Turning to the topic of sovereign AI. According to NVIDIA's blog, that company defines it as a nation's capabilities to produce artificial intelligence using its own infrastructure, data, workforce, and business networks. Mads, what does the idea mean to you?

When we say it's a sovereign computer, it means that basically there's access for everyone in society. Of course, computing costs money, but there's access for everyone who has an ethical purpose for which they want to use the computer. It's a national computer. Corporates, clinicians, hospitals, university researchers, entrepreneurs, everyone can access this computer, obviously. Now,

Nadia, does controlling access to 1,528 GPUs make you popular in Denmark?

It does. What we try to do is also make sure that we are allocating that capacity in the right way. And so one of the things we try to do is make sure that there's a balance of the academic and the commercial users. So we are now starting to take on customer workloads. We're still in this pilot phase. So right now we have a very high level of control about everything that is running.

And over time, as we open the doors even wider and we have thousands of customers, hopefully that's going to be something that we adjust daily or even minute by minutes to ensure that A, we are at full capacity as much as possible, but also that we're maintaining that balance over time of meeting the mission of having very high impact science that we're enabling, but also that we are successful on the commercial side.

Nadia, there are six pilot projects currently working on Geffion. What fields are they in and how are they going so far?

They're going really well. So everybody started having hands-on access basically by December. So they've been playing around now for a few weeks. They are all over the place in terms of what type of use cases and what success looks like for them. And that was very much by design. We were looking for diversity, both in the types of users. So we have some startups using the computer, and then we also have some academics using the computer.

but also in the fields that they're in. We have a very large autumn related use case happening that's led by the Niels Bohr team, and we're very excited about that. But we also have an application that's looking into using data for weather modeling. We also have some business use cases that are running on Gifion.

Success looks different for all of them. So the goal with the pilots was giving them access and enabling them, but it's also for us to really test multiple ways that Gifion can break and then making sure that it doesn't break when we're finally open to everyone. Matt?

Mads, when Jensen Wang came to Denmark, he spoke a lot about digital biology. And I'm curious, what excites you the most about digital biology? What we've been lacking is the ability to take the complexity of the biological world, because in a human organism or even in a mouse,

Or for that matter, in the DNA of plant organisms all over the world, there's such a diversity of molecular constructs, of the interplay between, for instance, microorganisms in our gut. There are so many different ones and they interact in highly complex ways that you can't model today.

All of this is starting to become possible now. So what we are actually doing is putting biology on steroids, so to speak. We're able to do things much faster and moreover, we're able to do things we never could do before. To give one example, we're funding a big project with another Nordic name, Aegis. And it's all about ancient DNA analysis where we will follow how plants have developed according to changes in climate.

hot periods of the world or during the agricultural embryonic era 12,000 years ago in Mesopotamia in China. How have these different crops and plants adapted to climate change via mutations that were beneficial and in case they were not, then they wouldn't survive. That is something that the center will look into geogenetically all over the world at least 100 places all over the world. And the amount of data coming in, some of which is a million years old,

has to then be held up against in a bioinformatics system, up against a number of reference genomes for existing crops and plants and so on. That's something you normally can't do. So when we started this project, Gifion was not up and running, but this is a project, once they've done their samples, it can only be done at this fantastic center that simply aims to mimic evolution

in today's age, and rather than use a million years to do so, we want to do it in maybe one or two or three breeding seasons. That's only possible because we have access to something like the Gifjom. That's amazing. And it sounds like understanding ancient civilization is going to be part of our future. Mads, one last question.

What else comes next? Ultimately, what are you hoping to accomplish with Gefeon here? Well, right now, I hope, as Nadja has explained, we hope to accomplish that we can elevate the whole ecosystem. Whatever you're doing research in the life science or other arenas, could be logistics, could be

every single AI amenable area of the Danish world and also in international collaborations, by the way, to be able to elevate that, the speed and quality with which we do the R&D and other efforts. That's the here and now thing. But I would also say that if you mention quantum, to make an error-tolerant quantum computer designed for life sciences or drug discovery, whatever it might be,

that calls for really close integration with high-capacity AI. And that's why one of the four areas we're working with also NVIDIA on, but also with the Nespor Institute and our content program is to make that happen. So you can actually say the future of computing is not either or, it's both and. And typically, probably also hybrid computing where you will need

both high-performance computing of the classic type with the CPUs. You need the AI component and you will need the quantum component with the QPUs. So I think that's really the future, to create a world where computing is not limiting progress in science. Nadia Mads, there is no one I would rather speak about ancient mythology and the future of AI supercomputers with. Thank you both for your time. Thank you.

And that's the science of success. This episode was produced by Charlotte Gartenberg. Michael LaValle and Jessica Fenton wrote our theme music. Our fact checker is Aparna Nathan. I'm Ben Cohen. Be sure to check out my column on WSJ.com. And if you like the show, tell your friends and leave us a five-star review on your favorite platform. Thanks for listening.