Required cookies

This website uses cookies necessary for its operation in order to provide the user with content and certain functionalities (e.g. language selection). You have no control over the use of these cookies.

Website visitor statistics

We collect visitor statistics on the use of the site. The data is not personally identifiable and is only stored in the Matomo visitor analytics tool managed by CSC.

By accepting visitor statistics, you allow Matomo to use various technologies, such as analytics cookies and web beacons, to collect statistics about your use of the site.

Change your cookie choices and read more about visitor statistics and cookies

CSC

The resources were granted in the Fourth Finnish LUMI Extreme Scale Call, which shared resources from Finland’s country share of the LUMI resources.

The accepted projects

ComPatAI

AI-Based Computational Pathology at a Scale, PI Pekka Ruusuvuori, University of Turku, co-PI Teemu Tolonen, Fimlab laboratories

The project builds on successful piloting of developing computational pathology tools for digital images of histopathological samples. So far, the research group and its collaborators have developed methods for prostate, breast and colorectal cancer diagnostics. In addition, the team has ethical, data access, and tissue processing permissions in place for a very large dataset consisting of altogether 600 000 whole slide images from routine clinical workflow from Fimlab. In this project, this dataset will be using modern AI on LUMI HPC environment.

Have a look at this video where the project members tell about the pilot project: LUMI – A Force for Cancer

Lumi-GPT-V

LumiGPT-Vision Multimodal Foundation Models, PI Aarne Talman, Silo AI, co-PI Sampo Pyysalo, University of Turku

This project will develop a family of massively multilingual and multimodal foundation models that can process both text and images, and generate text in multiple languages as the output. The project will leverage the vast computational capacity of the LUMI supercomputer and the extensive text resources of the Horizon Europe project HPLT and vast image datasets by LAION to train a family of multimodal and multilingual foundation models. The largest models will be over 60 billion parameters in size and trained on trillions of words and hundreds of millions of pictures from diverse sources, establishing a new state of the art for open multilingual multimodal models. The resulting models will enable a broad range of advances in multimodal AI across dozens of languages. The project will contribute toward transparent and trustworthy European AI by making all of its tools and models as well as their detailed training process openly available.

MOOMIN

Modular and open multilingual NLP, PI Jörg Tiedemann, University of Helsinki

The project aims to develop large modular language and translation models using the MAMMOTH neural framework. Leveraging LUMI’s capabilities for efficient parallelization, the team plans to scale up sequence-to-sequence models covering hundreds of languages. The final model’s components will be open-sourced for reuse, with a focus on efficient inference and exploring energy savings. The team commits to transparency by openly publishing all procedures, scripts, software packages, and datasets for replicability.

NewGEMS

New Generation Digital Earth, PI Heikki Järvinen, University of Helsinki, co-PI Jouni Räsänen, University of Helsinki

The project will deploy two new-generation Earth system models to simulate decadal climate at kilometre-scale resolution, and to provide a proof-of-concept for algorithmic uncertainty quantification of these models. The models are linked to the EU’s Destination Earth initiative, creating a climate information service with digital twins of the Earth’s climate system, as well as the Horizon 2020 project NextGEMS, preparing these digital replicas. They also relate to the EuroHPC Center of Excellence ESiWACE, offering solutions for data/compute-intensive workflows on pre-exascale heterogeneous computer architectures.

SASSE

SAturation of Small-Scale dynamos in convEction zones of stars, PI Maarit Korpi-Lagg, Aalto University, co-PI Jörn Warnecke, Max Planck Institute for Solar System Research

The research group used LUMI’s GPU capacity in earlier VISSI and SISSI projects to study the small-scale dynamo (SSD) in conditions resembling deep stellar convection zones. Surprises included, for example, easier SSD excitation than expected, indicating its feasibility in stellar convection zones. To confirm findings from the earlier projects, more simulations in realistic settings using LUMI’s computing capacity are crucial for determining the importance of SSD in stellar convection zones.

Read more about the earlier projects VISSI and SISSI, and have a look at the video interview on the LUMI website.

The next Finnish LUMI Extreme Scale Call will open in spring 2024.

More information about applying for LUMI’s resources on the LUMI website’s Get started section.