Pioneer Centre for AI

About P1 Compute Resources

The Pioneer Centre for Artificial Intelligence (P1) is a Danish cross-university research centre and consortium dedicated to advancing AI research through collaboration with partnering institutions.

This documentation aims to centralize practical knowledge and provide a concise guide for P1 affiliates on how to access and use the high-performance computing (HPC) facilities available to them.

A central mission of P1 is to foster broad research collaboration, both within the centre and across its affiliated universities. We fund dedicated partitions at partner HPC facilities and help run them, with the compute coordinator holding admin on the systems and working alongside the providers' HPC teams, so that P1 researchers across partner institutions can take on projects of varying scale and complexity.

For more information, visit the P1 Homepage or P1 Computing.

Available Compute Resources

While we prioritize smaller projects as they represent the majority of research needs, we also support larger projects through our tiered compute resources, including the capability to train very large models and support European-level workloads on Tier 2/1 facilities. But, very large scale projects might be better supported within their own funding.

In practical terms this means that while training a large language or vision model from scratch might not be suitable for our Tier 3 facilities, such projects can be accommodated on a Tier 2 national facility (ex. Gefion HPC), which is specifically designed for large-scale model training and can even support some European-level (Tier 1) workloads.

Support

In general the P1 does not have official support staff. Instead we rely on the community to help each other. However, for questions regarding the cluster, we have the following two contact e-mails:

For technical issues, like broken or missing packages, contact support@hpc.dtu.dk
For policy issues, like priority access, queuing system etc. contact compute-governance-p1@aicentre.dk

If your signup has been pending for more than a week, you can contact the member of the P1 HPC governance board of your institution. General announcements are made in the #compute channel on the official P1-slack. Similarly if you have general questions regarding compute, including using the P1 partitions, feel free to post those questions in the #ask-compute channel.

HPC Governance Board

The board has one representative per partner university. If your cluster signup has been pending for more than a week, or you have governance questions specific to your institution, contact your university's representative.

Investing Your Compute Budget

Generally, P1 can help purchasing hardware for you if it is to become part of a P1 partition. In this setup you would have priority access to the hardware that you purchase, but if it is unused other researchers at P1 can use it as well. Most HPC hardware has an expected utilization time of approximately 5 years.

An example of a compute budget expenditure:

A researcher might have 400k DKK for non-GDPR computing. In that case P1 can purchase a 2xA100 node, with the researcher getting priority access. With the last 100k DKK, we recommend a diversification of the researcher's compute portfolio: some of it on personal computers with powerful GPUs, and a significant portion reserved for later. That reserve could go to storage or cloud compute. Having the budget to spring for cloud compute if peak pressure on other computing resources makes training a model in time for a submission deadline infeasible can be necessary.

Planning a purchase is easier before the money is committed. Tell the compute coordinator what you have, what you plan to run, and how long it has to last, and you will get a recommendation on the split between cluster hardware, storage, cloud, and workstations.

Investing Your Compute Budget · Budget, workload, and the portfolio split. Nothing here commits you.

New tab

Hierarchy of Compute Resources

Depending on what you are doing, you will need different levels of compute. The following table gives an overview of the different levels of compute and what they are suitable for.

4 Personal

Personal laptop and/or desktop with either no GPUs or a single one. At this level compute is limited and mostly consists of code development and running experiments with small datasets and models.

3 Institute

Compute clusters containing either several smaller GPUs or a few larger ones. Multiple experiments can be run in parallel and models can be parallelized over a few GPUs, but it is not feasible to train very large models from scratch. Expect limitations to wall time (a couple of days), storage (around 0.5-1 TB) and number of nodes/GPUs used (around 2-16).

2 National

National HPC facilities provide the necessary resources for large-scale model training. Expect fewer restrictions to wall time (1-2 weeks), storage (around 1-10 TB) and number of nodes/GPUs used (around 16-64). Some national facilities can also support European-level (Tier 1) workloads. For example, Gefion, where P1 holds a sandbox partition, is a national facility that can support both Tier 2 and possibly Tier 1 workloads.

1 European

European HPC centers offer significantly larger scale compute resources, typically providing easy access to 8-16 GPUs with the possibility of scaling up to 500-800 GPUs depending on project requirements and the specific HPC center. Unlike lower tiers, walltime is usually not restricted; instead, limitations are set on total GPU hours used. See Karolina and Meluxina for examples of European HPC centers with GPUs.

0 Global

The highest level of compute. Possibility to use above 1000 GPUs and perform calculations at exascale. Usually only available to large research projects and collaborations with industry. See LUMI for an example of a global HPC center with GPUs.

Tier	Organization level	Description
4	Personal	Personal laptop and/or desktop with either no GPUs or a single one. At this level compute is limited and mostly consists of code development and running experiments with small datasets and models.
3	Institute	Compute clusters containing either several smaller GPUs or a few larger ones. Multiple experiments can be run in parallel and models can be parallelized over a few GPUs, but it is not feasible to train very large models from scratch. Expect limitations to wall time (a couple of days), storage (around 0.5-1 TB) and number of nodes/GPUs used (around 2-16).
2	National	National HPC facilities provide the necessary resources for large-scale model training. Expect fewer restrictions to wall time (1-2 weeks), storage (around 1-10 TB) and number of nodes/GPUs used (around 16-64). Some national facilities can also support European-level (Tier 1) workloads. For example, Gefion, where P1 holds a sandbox partition, is a national facility that can support both Tier 2 and possibly Tier 1 workloads.
1	European	European HPC centers offer significantly larger scale compute resources, typically providing easy access to 8-16 GPUs with the possibility of scaling up to 500-800 GPUs depending on project requirements and the specific HPC center. Unlike lower tiers, walltime is usually not restricted; instead, limitations are set on total GPU hours used. See Karolina and Meluxina for examples of European HPC centers with GPUs.
0	Global	The highest level of compute. Possibility to use above 1000 GPUs and perform calculations at exascale. Usually only available to large research projects and collaborations with industry. See LUMI for an example of a global HPC center with GPUs.

The P1 DTU and NGC HPCs sit at Tier 3 in the hierarchy. Gefion itself operates at Tier 2 and can support some Tier 1 workloads, but the P1 sandbox on it is Tier 3 capacity: enough to develop and mature a workload on the hardware it will eventually run on. Running at Tier 2 scale on Gefion is a funded step of its own, and the sandbox exists to make you ready to take it.

Reading these docs with an AI agent

If you work with a coding agent or a chat assistant, you can hand it this site rather than pasting pages in by hand. Three routes, in increasing order of how much they load:

https://hpc.aicentre.dk/llms.txt is an index of every page with a one-line summary, following the llms.txt convention. Point an agent here and it can fetch only the pages it needs.
https://hpc.aicentre.dk/llms-full.txt is the entire site as one markdown file, for pasting straight into a context window.
Any page works as plain markdown by adding .md to its path, so https://hpc.aicentre.dk/clusters/dtu.md is the DTU page without the site chrome. The View as Markdown link in the page outline does the same thing.

These are generated from the same source as the pages you are reading, so they cannot drift out of date. Treat them as a snapshot of guidance rather than live system state: queue limits and hardware change, and the provider documentation linked on each cluster page is always the authority.