A technical university in Sweden is getting a new supercomputer

The KTH Royal Institute of Technology (KTH), based in Stockholm, is getting a new supercomputer that promises to add around 13.5 petaflops of performance, or 6× the current supercomputing capacity.

The HPE Cray-based high-performance computing (HPC) system has been named Dardel and is powered by AMD Epyc CPUs and AMD Instinct GPUs.

“We found that the performance of an AMD Instinct MI250X is comparable to that of two Nvidia A100 GPUs, but it consumes about half the power,” said Professor Dirk Pleiter, director of the PDC center for high-performance computing at KTH. “Most of the systems that have been deployed recently are based on AMD Epyc processors because, for typical scientific applications, they provide outstanding performance.”

KTH is deploying a total of 1270 nodes in the CPU partition and 56 nodes in the GPU partition of its new cluster, an HPE Cray EX system. Each CPU node is powered by dual 2.25GHz 64-core AMD Epyc 7742 processors. The majority (524) are equipped with 256GB of memory, but there are also some nodes with 512GB or 1TB and two have 2TB RAM. The GPU nodes will also use 64-core AMD Epyc processors, but in single-socket configurations with 512GB of memory. Each will be equipped with four AMD Instinct MI250X GPUs.

According to Pleiter, the facility needed to use GPU accelerators to meet the requirements as a partner in the Swedish National Infrastructure for Computing (SNIC) consortium, and the research tasks SNIC enables.

“Many researchers are running computational fluid dynamics (CFD) applications, as well as more general engineering applications,” says Pleiter. “In the CFD area, there are extremely scalable applications, such as Nek5000, alongside powerful applications that are notorious for their scalability issues, like OpenFOAM. Another major group using our HPC systems are researchers from the fields of biophysics and genomics. Other large user groups are from materials science and computational chemistry.”

The new supercomputing power will be used for research into several areas, including Alzheimer’s disease and sustainable sea travel.

Sweden's WASP is also upgrading its Berzelius supercomputer

Another Swedish supercomputer is also getting a GPU upgrade. Berzelius is operated by the country’s Wallenberg Artificial Intelligence, Autonomous Systems, and Software Program (WASP) out of the National Supercomputer Centre at Linkoping University.

Unveiled in 2021, the system is currently based on 60 Nvidia DGX A100 systems, WASP is planning to add another 34 A100s in the coming months.

The DGX A100 systems include graphics processing units (GPUs) with 80GB of memory each. Based on the Nvidia DGX SuperPOD, the GPU company said the addition should push the performance capacity to nearly 'half an AI exaflop' (around 470 'AI' petaflops).

The HPC system has developed large language models capable of drafting a speech or answering questions in Swedish, and sports 20 billion parameters. Over time, researchers hope to expand this to 175 billion parameters and to make the language models also fluent in Danish and Norwegian.

“Supercomputers are helping to solve some of the hardest problems of our time — and with the Wallenberg Foundation, we’re providing the conditions to fuel AI innovation and groundbreaking research,” said Nvidia CEO and founder Jensen Huang. “Sweden has a vibrant AI ecosystem, and Berzelius is fundamental in supporting the country’s competitiveness.”

Berzelius is funded by the Knut and Alice Wallenberg Foundation.