LUMI, Europe's leading HPC system, successfully completed pilot deployment of a GPU partition for AI workloads.

Finland's LUMI supercomputer has completed the pilot phase of its GPU partition, extending the processing power of the system. LUMI is the fastest supercomputer in Europe and the third fastest globally, according to the Top500 list published in November 2022. The system was inaugurated in June last year, though the GPU partition had not been fully installed at that time. A team of Norwegian researchers has now completed the GPU pilot phase for LUMI-G, testing the system's capabilities with the GPU partition installed and operational.

The testing comprised 30 pilot projects covering three main areas: solar atmospheric modeling, natural language processing (NLP), and imaging of seismic data. For the solar atmospheric modeling project, researchers at the Rosseland Centre for Solar Physics at the University of Oslo work with simulations of the Sun's atmosphere and compare them with observations from solar telescopes. They use in-house applications known as Bitfrost and Dispatch, which are coded to be highly parallel, and the research team is adapting the code to use GPUs. The NLP project at the Department for Informatics at the University of Oslo covers research into automated translation, human-computer interaction via spoken language, and content recommendations and contextual advertising, requiring the enormous computing power needed to train advanced deep learning models. Meanwhile, the seismic data project at the Department of Geoscience and Petroleum at the Norwegian University of Science and Technology (NTNU) aims to get detailed images of Earth's interior using seismic data recorded on land and in the sea, employing a new modeling technique where seismic data is simulated using a comprehensive numerical model repeated for many sources.

LUMI is installed at the IT Center for Science (CSC) datacenter in Kajaani, Finland, and is a pre-exascale system provisioned under the auspices of the EuroHPC Joint Undertaking. It is based on HPE's Cray EX architecture, like the Frontier exascale system at Oak Ridge National Lab in the US. The CPU-only partition comprises 1,536 dual-socket CPU nodes, each featuring AMD "Milan" Epyc processors and between 256GB and 1,024GB of memory. The LUMI-G GPU partition has 2,560 nodes, each featuring a single custom AMD "Trento" Epyc chip and four AMD MI250X GPUs. Andrey Kutuzov, a member of the Language Technology group using LUMI, said that onboarding was "relatively easy for those already familiar with HPC environments," and that in most cases, the code designed for Nvidia GPUs was found to run flawlessly on LUMI's AMD GPU systems.

In November last year, LUMI was upgraded with the addition of a quantum accelerator from the VTT Technical Research Centre of Finland called Helmi, which comprises 5 qubits and is being touted as a quantum accelerator or co-processor to work alongside LUMI's classical CPU nodes. The EuroHPC JU has announced a call for tenders to select a vendor for JUPITER, intended to be the first European exascale system, with a tender closing date of February 13. JUPITER will be installed on the campus of the Forschungszentrum Jülich research institution in Germany, owned by the EuroHPC JU and operated by the Jülich Supercomputing Centre. The EuroHPC JU will fund half the total cost of the new system, with the other half coming from the German Federal Ministry of Education and Research (BMBF) and the Ministry of Culture and Science of the State of North Rhine-Westphalia (MKW NRW).