Summary
- The University of Ljubljana has launched FRIDA, a modular rooftop HPC data centre at its Faculty of Computer and Information Science.
- The system combines 104 GPUs across seven generations, including Nvidia Blackwell B200 and B300 units, with hybrid air and direct-to-chip liquid cooling.
- FRIDA adds national AI and supercomputing capacity in Central and Eastern Europe for research, public-sector, and industry use.
The University of Ljubljana has launched FRIDA, a modular high-performance computing data centre on the roof of its Faculty of Computer and Information Science, giving Slovenia a new AI and supercomputing platform built around dense GPU infrastructure and hybrid cooling.
The system was brought online on June 16, with the university publishing details on June 17. FRIDA is intended for artificial intelligence, machine learning, large language models, scientific computing, and big data processing. The university said the platform will be available to researchers and students at the faculty as well as companies developing advanced AI solutions.
The university’s project update says FRIDA combines 104 GPUs across seven generations, including Nvidia Blackwell B200 and B300 GPUs. It describes the system as the first academic supercomputing infrastructure in the region with AI-oriented Nvidia Blackwell B200 and B300 graphics processing units installed. FRIDA can reach up to 708 petaflops at lower precision, with exascale potential for sparse matrix calculations at lower precision.
Cooling sits at the centre of the design. FRIDA is built as a modular containerised data centre and uses hybrid cooling, combining air cooling with advanced liquid cooling directly on chips. The university says that combination is needed for modern AI servers with very high thermal loads.
The facility gives Slovenia a more visible position in Europe’s distributed AI infrastructure map. It does not compete with the largest hyperscale campuses or national exascale supercomputers on raw scale. Its role is more specific: it gives a university-led ecosystem local access to high-end AI hardware and a physical platform that can support research, prototyping, public-sector work, and industry collaboration without every workload leaving the country.
Dense compute in a constrained footprint
FRIDA’s rooftop location shapes the engineering problem. Data centre development is usually discussed in terms of large campuses, edge sites, or retrofits inside existing industrial buildings. A modular rooftop HPC installation brings a different set of constraints: structural load, access, cooling, power routing, fire safety, maintenance, noise, and resilience all have to be managed within an academic site.
The choice reflects the use case. This is not a speculative colocation hall waiting for external tenants. It is a university-controlled compute platform intended to sit close to researchers and collaborators. Proximity can shorten the path between hardware, research teams, software environments, and industrial users testing new models or applications.
The facility also shows how liquid cooling is moving beyond the largest data halls. Direct-to-chip systems are no longer confined to hyperscale AI deployments. As GPUs push beyond the limits of conventional air cooling, universities, research institutions, and smaller HPC facilities are facing the same thermal questions as commercial operators, often with less space and less budget flexibility.
Hybrid cooling is a practical route through that constraint. It allows parts of the system to be cooled conventionally while using liquid where chip-level heat density demands it. That reduces the need to redesign every part of the facility around liquid cooling, but it still requires careful commissioning, maintenance skills, leak detection, coolant management, and service procedures.
A national compute layer
The university says FRIDA offers twice the learning capacity and energy efficiency of Vega, Slovenia’s earlier supercomputer launched in 2021 as part of the EuroHPC ecosystem. That comparison should be read as workload-specific rather than a simple replacement story. Vega remains a major national and European supercomputing asset, while FRIDA is more tightly focused on AI and machine-learning development.
FRIDA also lands in a European policy environment where access to AI compute is increasingly treated as industrial infrastructure. The EU is expanding AI factories and planning AI gigafactories, but those programmes do not remove the need for national and regional platforms. Smaller countries need local capability if universities and companies are to develop models, train specialists, and prototype applications without being fully dependent on remote commercial clouds.
The facility can help bridge the gap between research output and commercial deployment. Access rules, pricing, support, software environment, data governance, and security will decide whether companies can use FRIDA effectively or whether it remains mainly an academic resource.
The installation also gives Central and Eastern Europe a working example of modular high-density compute. Much of the region’s data centre conversation still centres on land, power, and hyperscale proposals. FRIDA shows another route: targeted compute capacity, installed in a constrained academic environment, with cooling designed around AI hardware from the outset.
The next questions are operational. Workload mix, utilisation, industry access, cooling performance, and support capacity will decide how far FRIDA extends beyond the university. Even so, the project adds useful evidence that Europe’s AI infrastructure build-out will be spread across universities, public research bodies, compact modular sites, and commercial campuses rather than one facility type alone.
