High-Performance Infrastructure

Kaiser I

Kaiser I is our scalable, high-performance GPU computing cluster designed to advance generative AI models, cognitive robotics, and circular systems engineering.

Documentation Website View Nodes & Hardware

Total Nodes

1 Login, 4 Workers

NVIDIA GPUs

8x H200, 8x RTX 6000

€400k+

Funding

EU & Niedersachsen

Slurm

Scheduler

Resource manager

Project Description

The Kaiser I GPU Cluster

The Kaiser I cluster is established as the central digital platform at the Technical University of Clausthal to unlock the full potential of Generative Artificial Intelligence (GenAI) for systems engineering. The goal is to both advance fundamental AI methods and drive concrete engineering applications - ranging from resource-efficient design and circular production to cognitive robotics and sustainable mobility.

Led by Prof. Dr. Christian Bartelt at the Institute for Software and Systems Engineering (ISSE), Kaiser I provides a scalable, high-performance environment that strengthens research, teaching, and technology transfer while fostering regional and international innovation partnerships.

Prof. Dr. Christian Bartelt and researchers powering up the new Kaiser I GPU Cluster at TU Clausthal.

Kaiser I funded by the European Union & Land Niedersachsen (EFRE/ESF/ELER)

Located at Rechenzentrum der TU Clausthal, Erzstraße 18

Led by Prof. Dr. Christian Bartelt (ISSE)

Open workshops and innovation hub access for industry partners

Hardware & Nodes

Cluster nodes and hardware layout

The cluster is structured into login and worker nodes containing specific GPU architectures.

2 Nodes · H200 GPUs

High-Performance Nodes

Two high-performance worker nodes, each hosting up to 8 NVIDIA H200 NVL GPUs (minimum 8 H200 GPUs total) interconnected via NVLink for maximum data throughput.

Up to 8 H200 NVL per node
Optimized for model fine-tuning up to 70B
Inference of Llama 3, Qwen 3, Gemma 3

2 Nodes · RTX 6000 GPUs

Mid-Performance Nodes

Two mid-performance worker nodes hosting up to 8 NVIDIA RTX PRO 6000 Blackwell Max-Q GPUs. Tailored for resource-efficient training and vision models.

Blackwell Max-Q GPU architecture
Optimized for training models < 3B
Supports vision models (SAM, CLIP)

Fileserver · ≥ 150 TB

Storage & Master Node

A master node and fileserver equipped with ≥ 150 TB of high-speed persistent storage to manage large datasets, model checkpoints, and evaluation runs.

≥ 150 TB persistent storage capacity
Centralized checkpoint registry
Secure, high-availability backups

Up to 200 Gbit/s

InfiniBand Network Interconnect

High-speed InfiniBand network interface running up to 200 Gbit/s to minimize latency and maximize transfer rates during distributed training operations.

200 Gbit/s maximum throughput
Low latency multi-GPU sync
Fast dataset streaming from fileserver

Cluster Queues

Partition layouts and nodes

Slurm uses partitions to manage resource allocations and group nodes.

All Cluster Nodes

batch (Default)

Non-interactive partition

Default queue for sbatch jobs
Supports long-running training tasks
Max duration up to 36 hours (llm-research)

RTX 6000 Pro Nodes Only

interactive

Interactive partition

Enables srun --pty bash or salloc sessions
Strict 6-hour time limit per session
Ideal for debugging and GPU checking

Gateway Node (Rechenzentrum)

Login / Frontend

cloud-201.rz.tu-clausthal.de

Located at the Computing Center of TU Clausthal
For editing code, data upload, and scheduling
Strictly no compute workloads allowed

Submit

Submit jobs via the Slurm scheduler using sbatch (batch) or srun (interactive).

Monitor

Monitor active jobs and resource allocations with squeue and sinfo.

Execute

Jobs run in isolated environments with resources allocated automatically.

Policies

Access, queue, and usage policies

Use of the Kaiser I GPU Cluster is governed by account groups and limits designed to share compute resources fairly.

Workshops & Open Lab

Regular workshops presenting the infrastructure to science, start-ups, and industry.
Active initiation of collaborative research and engineering projects.
Workshops registration opens after cluster commissioning (by Sep 1, 2025).

Innovation Hub

Connects regional universities and enterprises to strengthen technology transfer.
Directly supports the Lower Saxony RIS3 innovation strategy.
Enables resource-efficient design, circular production, and cognitive robotics.

User-Friendly Access

Training courses, direct hotline, and on-site support in German.
Full maintenance and replacement part service guaranteed for five years.
For collaboration inquiries, contact bartelt@isse.tu-clausthal.de.

Need detailed usage instructions?

Learn how to connect via SSH, run interactive pseudo-terminals, submit batch jobs, and manage your conda/singularity environments in our comprehensive cluster documentation.

Access Cluster Documentation

Publications

Research outputs from the network.

Papers, preprints, and workshop contributions produced using CORE Network infrastructure.

View full publication archive

ECML PKDD 2026 2026

Beyond Either-Or Reasoning: Transduction and Induction as Cooperative Problem-Solving Paradigms

ICML 2026 (Combining Theory and Benchmarks Workshop) 2026

BARISTA: A Multi-Task Egocentric Benchmark for Compositional Visual Understanding

CVPR 2026 (XAI4CV Workshop) and ICML 2026 (CompLearn Workshop) 2026

Concepts in Motion: Temporal Concept Bottleneck Model for Interpretable Video Classification

Transactions on Machine Learning Research 2026

What’s in the Bottle? A Survey and Roadmap of Concept Bottleneck Models

International Journal of Computer Vision 2026

Investigating Uncertainty Weighting for Multi-Task Learning: Insights and Analytical Alternative

EACL 2026 2026

Is This LLM Library Learning? Evaluation Must Account For Compute and Behaviour

ECAI 2025 2025

Beyond Pixels: Enhancing LIME with Hierarchical Features and Segmentation Foundation Models

ECAI 2025 2025

Disentangling Exploration of Large Language Models by Optimal Exploitation

ICML 2025 (AI4MATH Workshop) 2025

A Compute-Matched Re-Evaluation of TroVE on MATH

ICSE 2025 2025

GUIDE: LLM-Driven GUI Generation Decomposition for Automated Prototyping

NAACL 2025 2025

Large Language Models Share Representations of Latent Grammatical Concepts Across Typologically Diverse Languages

CHI 2025 2025

Closing the Loop Between User Stories and GUI Prototypes: An LLM-based Assistant for Cross-Functional Integration in Software Development

ICLR 2025 2025

Mitigating Information Loss in Tree-Based Reinforcement Learning via Direct Optimization

ICLR 2025 (DL4C Workshop) 2025

Shedding Light on Task Decomposition in Program Synthesis: The Driving Force of the Synthesizer Model

xAI 2025 2025

Which LIME should I trust? Concepts, Challenges, and Solutions

ASE 2024 2024

Self-Elicitation of Requirements with Automated GUI Prototyping

RE 2024 2024

Interlinking User Stories and GUI Prototyping: A Semi-Automatic LLM-based Approach

AAMAS 2024 2024

RAISE the Bar: Restriction of Action Spaces for Improved Social Welfare and Equity in Traffic Management

HICSS 2024 2024

DRAMA at the PettingZoo: Dynamically Restricted Action Spaces for Multi-Agent Reinforcement Learning Frameworks

ECML PKDD 2026 2026

Beyond Either-Or Reasoning: Transduction and Induction as Cooperative Problem-Solving Paradigms

ICML 2026 (Combining Theory and Benchmarks Workshop) 2026

BARISTA: A Multi-Task Egocentric Benchmark for Compositional Visual Understanding

CVPR 2026 (XAI4CV Workshop) and ICML 2026 (CompLearn Workshop) 2026

Concepts in Motion: Temporal Concept Bottleneck Model for Interpretable Video Classification

Transactions on Machine Learning Research 2026

What’s in the Bottle? A Survey and Roadmap of Concept Bottleneck Models

International Journal of Computer Vision 2026

Investigating Uncertainty Weighting for Multi-Task Learning: Insights and Analytical Alternative

EACL 2026 2026

Is This LLM Library Learning? Evaluation Must Account For Compute and Behaviour

ECAI 2025 2025

Beyond Pixels: Enhancing LIME with Hierarchical Features and Segmentation Foundation Models

ECAI 2025 2025

Disentangling Exploration of Large Language Models by Optimal Exploitation

ICML 2025 (AI4MATH Workshop) 2025

A Compute-Matched Re-Evaluation of TroVE on MATH

ICSE 2025 2025

GUIDE: LLM-Driven GUI Generation Decomposition for Automated Prototyping

NAACL 2025 2025

Large Language Models Share Representations of Latent Grammatical Concepts Across Typologically Diverse Languages

CHI 2025 2025

Closing the Loop Between User Stories and GUI Prototypes: An LLM-based Assistant for Cross-Functional Integration in Software Development

ICLR 2025 2025

Mitigating Information Loss in Tree-Based Reinforcement Learning via Direct Optimization

ICLR 2025 (DL4C Workshop) 2025

Shedding Light on Task Decomposition in Program Synthesis: The Driving Force of the Synthesizer Model

xAI 2025 2025

Which LIME should I trust? Concepts, Challenges, and Solutions

ASE 2024 2024

Self-Elicitation of Requirements with Automated GUI Prototyping

RE 2024 2024

Interlinking User Stories and GUI Prototyping: A Semi-Automatic LLM-based Approach

AAMAS 2024 2024

RAISE the Bar: Restriction of Action Spaces for Improved Social Welfare and Equity in Traffic Management

HICSS 2024 2024