GPU & LLM Cluster
High-performance compute,
built for Digital Humanities
Scale up what already runs on your laptop. Train and fine-tune models, run interactive GPU notebooks, and call a hosted OpenAI-compatible LLM and embedding API — without leaving the academic network.
What you can do — and how you get in
Three entry points, from a raw shell to a one-line API call. Pick the one that fits how you work.
Shell & SLURM
SSH in to run batch training and multi-GPU / multi-node fine-tuning over InfiniBand, bring your own containers (Apptainer), and use project storage. Full control for power users.
JupyterHub notebooks
Log in through the browser with single sign-on, pick your resources, and your GPU-backed notebook runs as a scheduled job — no terminal required.
LLM & embedding API
Call OpenAI-compatible LLM and embedding endpoints with a project key over HTTPS — for chat, RAG and semantic search. No cluster account needed.
How it differs from generic HPC
Interactive & API access
Not batch-only. Notebooks and hosted APIs alongside SLURM — we try to outperform generic HPC for DH needs.
Hosted model APIs
A ready LLM and embedding API with project keys — something a classic HPC allocation does not give you.
Part of Austria's HPC ecosystem
DHinfra.at is one piece of a wider landscape, and we would rather connect to it than reinvent it. Austria already has excellent shared supercomputing through Austrian Scientific Computing (ASC) with the Vienna Scientific Cluster (VSC) and MUSICA, and most major universities — including ours — hold partnerships that make these systems available to their researchers, often at no cost. Where your work fits there, we are happy to point you to it.
Beyond the national centres, many Austrian institutions register GPU servers, compute clusters and other equipment in the federal research-infrastructure database (FI DB) and mark them as open for collaboration. Searching it is often the quickest way to find a cooperation partner near you; terms are agreed with the institution that runs the facility.
The skills carry over. If you are comfortable with a SLURM batch job or a JupyterHub session on one cluster, you are most of the way to using ours — and to using national and European systems. So for the fundamentals we gladly lean on the training material and events these centres already maintain, and focus our own effort on what is specific to DHinfra.at: interactive, DH-first, API-driven use.
We are in close contact with these partners in any case. EuroCC Austria and ASC have lent their expertise to the planning of our cluster and to parts of its operation (for example authentication), and we try to give something back by acting as a low-threshold, DH-oriented entry point into this ecosystem.
Hardware at a glance
The cluster at the University of Graz. Figures are indicative.
| GPUs | Memory | Typical use |
|---|---|---|
| 12× NVIDIA H200 | 141 GB each | Large-model training, SOTA LLM serving |
| 11× NVIDIA RTX PRO 6000 | 96 GB each | Medium training, notebooks, model serving |
| 4× NVIDIA L40S | 48 GB each | Embeddings, lightweight inference, courses |
Entry in the federal research-infrastructure database →
The secondary cluster site at the University for Continuing Education Krems features hardware and a configuration closely related to Graz. Both systems are accessed in the same way, but different limits may apply.
| GPUs | Memory | Typical use |
|---|---|---|
| 6× NVIDIA H200 | 141 GB each | LLM serving, Tailored application hosting (e.g. eScriptorium) |
Entry in the federal research-infrastructure database →
Storage is tiered: very fast NVMe right next to the GPUs (served over RDMA InfiniBand) to keep training and inference efficient, plus larger, slower capacity for everything else — including data staging and short- to mid-term project storage. Across the sites that is at least 100 TB of fast and 500 TB of slower storage. Everything runs over the Austrian academic network (ACOnet).
Getting access
1. Authenticate
Log in at login.dhinfra.uni-graz.at with the eduID account of your home institution (ACOnet single sign-on). The documentation sits behind the login.
2. Apply for a project
Submit a short project application — research, course, thesis, event, or exploratory use.
3. Get approved
On approval, SSH, the SLURM scheduler, storage and the model APIs are unlocked for your project.