In 2024, it doesn’t matter whether we should adopt AI or not. Nor is it a question of whether we need sovereign AI. This is the main reason we are seeing countries and large multinational companies investing in AI infrastructure in ways never seen before with traditional HPC simulation and modeling.
This week it’s Danish.
Back in March, the Danish government, the Novo Nordisk Foundation, and AI systems giant Nvidia announced that they would collaborate to build a supercomputer for the newly established Danish AI Innovation Center. Today that machine is called Gefion. In operation.
The Novo Nordisk Foundation is the non-profit parent company of Novo Nordisk, the world’s largest producer of insulin for the treatment of diabetes and a pharmaceutical giant involved in many aspects of health care, including weight loss with the use of Ozempic and Wegovy. The foundation’s net assets will be $167 billion in 2023, making it the largest philanthropic foundation on the planet. Its wealth is clearly backed by pharmaceutical companies, which had revenue of $33.71 billion and net profit of $12.15 billion last year.
Gefion is the Norse goddess of the plow and harvest, and it’s safe to say that NVIDIA certainly reaped what it sowed. The overwhelming consensus is that generative AI will impact every aspect of business, public, and private life, so countries want to have their own AI supercomputing centers.
As such, it is expected that hundreds of countries will want machines at least as large as the Gefion machines, as well as tens of thousands of other companies that will eventually want to take over control of the AI, albeit on a smaller scale. expected to appear. It’s for hyperscalers and cloud builders to spend billions on their own AI applications, or to rent capacity to the millions of businesses adding AI to their operations.
No matter how you do this math, there are tens of thousands of enterprises with hundreds of thousands of GPUs, hundreds of organizations with thousands of GPUs, and tens of thousands of organizations with hundreds of GPUs. That’s it. And we plan to add more GPU fleets over time. That’s millions of GPUs, tens of thousands of dollars each. This is how we achieve a recurring revenue run rate of over $100 billion across GPU manufacturers. Some say it will be many times larger than this. Let’s take a look.
By the way, DCAI is a new company that plans to sell the machine’s capacity to AI researchers and companies. This is not so much a donation as an investment in new business opportunities by Novo Nordisk and EIFO. The chief executive officer appointed to run DCAI is Nadia Carsten. He is a researcher in nanofabrication and DNA and RNA sequencing at the University of California, Berkeley, and was a program manager and director at the U.S. Department of Homeland Security and product director for the Quantum Center. Computing with Amazon Web Services. In the featured image above, Karsten is in the center, Nvidia co-founder and CEO is on the left, and Denmark’s King Frederick X is on the right.
As such, Gefion is not a particularly large-scale AI supercomputer, but it offers strong mixed-precision performance and 64-bit and 32-bit floating point capabilities, enough to expand Denmark’s AI and HPC footprint. It is certain that it is fully equipped.
Gefion consists of 191 Nvidia DGX H100 servers with a total of 1,528 Nvidia “Hopper” H100 GPU accelerators. Each DGX 100 is equipped with two Intel “Sapphire Rapids” It has an array of switches. The SuperPOD machine is manufactured by Eviden, the HPC and systems division of French IT conglomerate Atos. Atos, which is in the process of being spun off, has annual revenues of 5 billion euros (approximately $5.4 billion). (I still can’t believe a spinout will happen.)
The peak performance of the Gefion machine is 51.2 petaflops on FP64 vectors on the H100 device and 102.4 petaflops on FP64 on the H100’s tensor cores. Without sparsity processing turned on, the 1,528 GPUs in the system peak at a total of 1.51 exaflops at FP16 precision. That’s twice the lowest FP8 resolution supported by the H100 GPU. If your data is sparse and you can ignore half of the zero-filled bits to compress the data, you’ll get an additional 2x increase in throughput.
In terms of price, the Novo Nordisk Foundation is investing DKK 600 million (approximately 14 cents on the dollar), and the Danish Export Investment Fund (EIFO) is contributing a further DKK 100 million. Add this up and multiply by the krone to dollar exchange rate and you get $98 million. The machine will be hosted at the Digital Realty data center in Copenhagen and will be powered entirely by renewable energy. As far as we know, these numbers do not include operating costs, but likely include system software and support costs. When AI and HPC centers talk about costs, they are imprecise and have to make guesses.
Here’s where the problem occurs on a machine with a GPU of the same vintage as the Gefion machine or with a homemade accelerator.
In some ways, Gefion machines are very aggressively priced. That must be because the DGX H100 looks like 2022. We have a GH200 system with Grace CPUs and Goosed GPUs, and the first wave of “Blackwell” GPUs is also in production now. These days, any AI/HPC center is happy to take any Nvidia GPU they can get, but Novo Nordisk and its DCAI startup would definitely have preferred to get a more powerful Blackwell machine.
So how does the Gefion system compare to other Danish resources? Denmark gets 3 percent of the resources available in CSC Finland’s “Lumi” system. Lumi is rated at 531.5 petaflops peak with FP64 precision, which equates to a 15.95 petaflops machine with 64-bit floating point. Through the EuroHPC consortium, Danish companies and researchers can also apply for access to other leading irons in Europe. The Danish Electronic Infrastructure Consortium (DeiC for short) coordinates access to each country’s supercomputers, providing so-called “interactive”, “high memory”, “throughput” machines (presumably the Danish variety (I think it’s installed at the university), but I can’t find it. Feed it to increase its speed. DeiC appears to include Aalborg University, Aarhus University, Copenhagen Business School, Technical University of Denmark, IT University of Copenhagen, University of Copenhagen, Roskilde University, and University of Southern Denmark.
And while Novo Nordisk probably had its own systems that it had been researching for decades, you won’t find any evidence of them in the recent Top 500 Supercomputers rankings. Danske Bank (finance), Moller Maersk (shipping), DSV (logistics), Novonesis (biosciences), Coloplast (medical devices), and a number of other Danish companies are also believed to have supercomputers. .
Sign up for newsletter
We’ll send you the week’s highlights, analysis and stories straight to your inbox, with nothing in between.
Subscribe now
Related articles
