Perun, LUMI, or Leonardo? How many GPU hours do you actually need?
Let’s find the right computing resource for your project.
National high-performance GPU infrastructure available via SK:AI:AT
For initial development, experimentation, and more demanding AI workloads, we can utilize the national HPC infrastructure operated by the Computing Centre of the Slovak Academy of Sciences.
The PERUN SAV supercomputer was commissioned in 2026 and is operated by the Computing Centre of the Slovak Academy of Sciences (SAV). It is a heterogeneous HPC system with a theoretical peak performance of approximately 17 PFLOP/s. Its architecture combines three types of computing resources: a general-purpose CPU module, a memory-expanded cloud module, and a high-performance accelerated module based on the NVIDIA Grace Hopper architecture.
| GPU nodes | 76 |
| Accelerators | 4× NVIDIA GH200 Grace Hopper / node |
| Overall | 304× NVIDIA GH200 |
| GPU memory | 96 GB HBM3 / GPU |
| GPU memory per node | 384 GB HBM3 |
| CPU | 4× 72-core NVIDIA Grace / node |
| CPU-GPU | NVLink-C2C |
| GPU-GPU | NVLink |
| Interconnect | 4× 400 Gb/s InfiniBand NDR400 / node |
| Local storage | 3,8 TB NVMe / node |
Each accelerated node integrates four NVIDIA GH200 Superchips, which combine an NVIDIA Grace CPU and a Hopper GPU. The Hopper GPU features 4th-generation Tensor Cores and supports computations in various precisions—including TF32, FP16, BF16, FP8, and INT8—which is relevant for modern mixed-precision AI workloads.
A high-speed InfiniBand network and GPU interconnects enable the use of multiple accelerators and nodes, even for distributed workloads. Consequently, Perun can be utilized for tasks such as model training and fine-tuning, LLMs and generative AI, computer vision, multimodal models, distributed training, large-scale experiments, and computationally intensive inference.
Perun uses the Slurm workload manager; GPU workloads are run via dedicated GPU partitions.
How can we help you?
You don’t need to know how many GPU hours or what type of node you need.
Together, we will assess your workload, assist in estimating the necessary resources, and—if needed—prepare a test run or benchmark. As we move forward, we can help with environment setup, executing GPU workloads, scaling, and optimizing resource utilization.
European AI infrastructure within reach
Your project does not have to remain limited to Slovak infrastructure.
The European network of AI Factories, being built by the EuroHPC Joint Undertaking, connects AI-optimized supercomputers with expert support, data, software tools, and other services for AI development.
Currently, the EuroHPC JU lists 19 AI Factories and 13 AI Factory Antennas.
| AI Factory | Country | System | Core technology | Note |
|---|---|---|---|---|
| LUMI AI Factory | Finland | LUMI-G | AMD MI250X, 128 GB HBM2e (x4)* | AMD ROCm, large-scale GPU computing |
| IT4LIA | Italy | Leonardo Booster | NVIDIA A100, 64 GB (x4) | Large-scale CUDA/HPC |
| BSC AI Factory | Spain | MareNostrum 5 ACC | NVIDIA H100, 64 GB (x4) | Hopper, high-speed multi-node AI |
| L-AIF | Luxembourg | MeluXina GPU | NVIDIA A100, 40 GB (x4) | CUDA, AI/HPC |
| SLAIF | Slovenia | Vega GPU | NVIDIA A100, 40 GB (x4) | CUDA, AI/HPC |
| BRAIN++ | Bulgaria | Discoverer GPU | NVIDIA H200, 141 GB HBM3e (x8) | Large-memory GPU, GenAI/LLM |
| JUPITER AI Factory (JAIF) | Germany | JUPITER | NVIDIA GH200, 96 GB HBM3 (x4) | Grace Hopper, exascale AI/HPC |
How can one access EuroHPC AI compute?
Playground – I want to try out AI Factory first:
For startups, SMEs, and users who are just starting out with HPC. It is intended for smaller-scale experiments, testing technologies, and getting familiar with the AI Factory environment.
The application is brief, the call is open on a rolling basis, and EuroHPC indicates that access can be obtained within 2 working days. The allocation may last 1–3 months, and a single system is selected.
Suitable for: initial GPU experiments, model testing, benchmarking, minor fine-tuning, or workflow verification.
Fast Lane – I have a ready workload and I need a bigger allocation.
Intended for companies that already have experience with HPC and require up to 50,000 GPU hours. Access can be approved within 4 working days, and the allocation lasts for 1–3 months.
Suitable, for example, for: training and fine-tuning models, large-scale experiments, benchmarking, or developing PoCs with higher computational requirements.
Large Scale – I need more than 50,000 GPU hours.
For computationally intensive and innovative AI projects capable of effectively utilizing a large allocation. Projects may be granted resources for 3, 6, or 12 months. The call is open on a rolling basis with two cut-off dates per month, and successful proposals can be approved within 10 working days of the cut-off date.
Suitable, for example, for: large-scale training, the development of foundation models, extensive generative AI projects, or longer-term, computationally intensive experiments.
AI Factory Austria AI:AT
SK:AI:AT is linked to the Austrian AI Factory Austria – AI:AT, which serves as our primary entry point into the broader European AI Factory ecosystem.
AI:AT combines AI computing infrastructure with expert services for the development, testing, and scaling of AI solutions. Through the collaboration between SK:AI:AT and AI:AT, we are establishing a framework that will allow Slovak users to access AI:AT’s available infrastructure, platform services, and expert support.
We will publish detailed access options and available services soon.