Digital autonomy starts locally: how KdG built its own computing cluster
Setting up your own computing cluster? That sounds like something for big universities or tech giants. Not for a university college with no HPC experience. Yet that is exactly what they did at Karel de Grote University of Applied Sciences and Arts. Researchers from the Sustainable Industries Research Centre decided to build one themselves, for good reason: more control, lower costs, and digital autonomy. What started as a leap of faith is gradually growing into a solid foundation for AI research, collaboration, and education.
Why KdG as ‘beginners’ still chose to have their own computing cluster
At the Karel de Grote University of Applied Sciences and Arts, we work on many data-intensive projects at the Sustainable Industries Research Centre. Until now, we mainly used regular workstations or cloud solutions. But we increasingly ran into limitations in terms of scalability, collaboration, and control. For instance, working from remote desktops, on a multi-user workstation, was not ideal. However, storing large, growing, datasets in the cloud also comes with a hefty price tag. Therefore, despite our limited experience with high-performance computing (HPC), we decided to take the leap: setting up our own computing cluster. To do so, we appealed to VLAIO’s ‘Research infrastructure for universities of applied sciences and arts’ grant.
Why HPC?
Our researchers and students increasingly need computing capacity, especially for projects around artificial intelligence, Large Language Models (LLMs), and simulation. We also noticed that collaborating on data and models requires shared infrastructure. An infrastructure that ensures everyone works in the same environment, without depending on proprietary hardware.
Local infrastructure also gives us more control. We can better monitor how and where data is stored and processed, which is important for privacy and regulatory reasons. It also gives us more control over costs in the long term: once the investment is made, we can use the resources for a long time with predictable support costs, without variable traffic and tariffs we have no control over.
Something we absolutely did not realise when we started is how important digital autonomy would become for Europe. AI is evolving rapidly, and nobody knows what the finish line looks like. However, it is clear that AI will become a core requirement for organisations such as ours. More and more processes will be driven by AI models, so computing power will become a strategic resource.
Combine that with the fact that many commercial AI providers are able to their services for now thanks to large investments from venture capitalists – resources that are not infinite – and it becomes clear that dependence on these parties is risky. The running costs of AI models in the cloud might be much higher in the future. In light of that, an in-house cluster provides a certain peace of mind.
But, why not just run it in the cloud?
A fair question. The cloud is undoubtedly the default choice for those who want to quickly get started or run temporary projects. The cloud can also be more advantageous for specific applications, especially for low or irregular workloads.
Purely from an economic point of view, you must have a substantial computing power need to justify the switch. In our view, however, a computing cluster is not something you opt for purely to save costs. A crucial insight, for example, is understanding what is really goes on during a computing job and where the bottlenecks lie. Privacy and data management are also a challenge, especially with sensitive datasets such as medical data (think GDPR requirements). Finally, we wanted to make collaboration with students and colleagues easier through a shared platform with central login and access to local data.
For us, the choice was not a rejection of the cloud, but a conscious decision to start locally where it has added value.
As novices, why start on your own anyway?
We are not HPC experts, but that is precisely what made it interesting for us to set up the infrastructure ourselves. There is more and more open-source software and knowledge available. Flanders has a strong HPC community in the form of the VSC (Vlaams Supercomputercentrum).
Thanks to a VLAIO research infrastructure grant, we were able to start a tender for the placement of the computing cluster. This is how we ended up with Clustervision as an HPC provider. With them, we discussed how to translate our computing needs into the specifications of the cluster. In our case, the cluster emphasises the use of GPUs (graphics processors). In the early days, we consciously opted for extensive support in the form of remote system administration. This way, we can count on stable operation, while at the same time taking the time to build up the necessary knowledge internally. Step by step, we grow in our role as system administrator.
That time is much needed, because there is indeed a considerable learning curve: from configuring job scheduling, user management, Single Sign On, Jupyter Notebooks, integration in the KdG network, to firewall settings.
The first application: ExplainMed
One of the first projects on our cluster was ExplainMed. In this project, we applied AI to medical data. As it involves sensitive data, it was a big advantage that we could rely on our own infrastructure: the data stays local and we have control over access and logging. Training a first full AI model on our cluster, rather than limiting ourselves to fine-tuning, felt like a real milestone.
We also use the cluster for internal AI training, testing parallelisation, simulations, and student projects that benefit from more computing power than their laptops can provide.
What we have learned (so far)
The biggest complexity is not in the hardware, but in the software stack: user management, job scheduling, monitoring, logging, security, updates, etc. It requires intensive cooperation between IT, researchers, and our HPC provider.
We learned that good documentation and user support are essential. A cluster is only valuable if people can work with it smoothly. That is why we invest in short manuals and comprehensive documentation. Understanding a system architecture is also an added value we want to impart to our students and researchers, which will help them understand the cloud later on.
Our key insight? Start small. Test a few typical workflows or the main use case. For us, this was being able to offer Jupyter Notebooks in the browser and a container runtime. Try training a small LLM. Did you succeed? Try it with multiple GPUs and nodes, and so on. Let users give feedback. Growth is always possible.
Invest in tooling. Plug-ins for your IDE can save a lot of time on your command line. With the right tooling, HPC access need not be difficult and it’s like accessing a regular server.
What do we still want to learn/do?
This is just at the beginning. We want to better tailor our job scheduling to different types of users (e.g. students vs researchers). We also want to further roll out tools such as Matlab and links to NAS storage.
Finally, we want to build a user community, with mutual exchange of scripts, tips, and feedback. HPC is a learning process that works better if you do it together.
There are many emerging initiatives around access to HPC such as EuroHPC and AI Factories. This importance seems to be increasing. Additionally, the transition to HPC computing is still very big in our field. We hope in time to be able to assist others, either by allowing access to our cluster or by introducing companies to the world of HPC.
Final thoughts
Having your own computing cluster is not an end point, but a starting point. For small institutions, HPC sometimes seems out of reach, but our experience shows: it is feasible and valuable. You don’t need to be an expert to get started.
Our takeaway: start simple, be open to learn and work together. HPC is not just for big universities or institutions. Although it requires a substantial investment in people and resources, it is an option for any institution that wants to take steps into data-driven research with curiosity.
This blog was written with the help of AI.
Remco van Schadewijk
Dr Remco van Schadewijk (KdG, Sustainable Industries Research Centre) is an interdisciplinary scientist and researcher with a background in life sciences, data science, and science communication. After a PhD at Leiden University, where he researched metabolic imaging with MRI, he focused on using data and AI for industrial innovation. He is currently coordinator of the research line Data Science for Industry at the Karel de Grote Hogeschool in Antwerp, where he leads projects on digitisation and sustainability. He is also active in science communication, including through educational multimedia and educational innovation.
Wouter Deketelaere
Wouter Deketelaere is a researcher at the Sustainable Industries Research Centre at the Karel de Grote Hogeschool. He has extensive experience in both teaching and research, with a strong focus on Data Science, Artificial Intelligence (AI), Machine Learning (including Reinforcement Learning and Large Language Models).
With a master degree in Artificial Intelligence, Wouter combines in-depth theoretical knowledge with practical applications. As a lecturer in the Applied Computer Science programme, he developed and taught courses on Data Science and AI for many years. He now works on the innovative application of these techniques within various research projects.
Thanks to his expertise and experience in both research and teaching, Wouter can build an important bridge between these domains. He focuses on applying research results and translating them into specific, understandable applications for various target groups, which contributes to the dissemination and implementation of project results.
Related courses
HPC Café
Seminarie - Online, Karlsruhe - KIT, HammerHAI, Training geselecteerd in de Belgische AI Factory Antenna (AIFA)
SAIAR Summer Studio: From latent to physical space
Zomerschool - Kortrijk - Howest Hogeschool, SAIARlab