Summary

  • DFKI operates a documented AI high-performance computing cluster named Pegasus at Kaiserslautern, with published specifications (100+ PFLOPs for AI workloads, 300+ GPUs, 5,500+ CPU cores, 40+ TiB RAM), a Slurm-based access workflow and a public per-partition GPU table.
  • The institute's GPU capacity has a dated procurement trail: DFKI's press release of 27 July 2020 records its first NVIDIA DGX A100 going into operation, lifting cluster performance from 20 to 45 petaFLOPS; a later announcement records a seventh DGX A100 and a rise to more than 60 petaFLOPS.
  • In Bremen, DFKI's Robotics Innovation Center runs the 1,300 m² Maritime Exploration Hall (MarEH), including a 23 m × 19 m × 8 m basin holding 3.4 million liters of saltwater and a pressure chamber rated to 660 bar — though the source page itself is inconsistent on this figure.
  • SmartFactory-KL, co-anchored by DFKI, was announced as a Gaia-X testbed in a DFKI press release dated 26 May 2021, with the smartMA-X sub-project connecting a Production Level 4 demonstrator to the Gaia-X network.
  • Dated, funded projects give the deployment narrative substance: LUMINOUS (EU grant 101135724, January 2024 – March 2027), LIEREx (July 2024 – December 2026), SiSWiss (June 2025 – May 2028), TSS-MoVe and Genki4Media.
  • Against this, DFKI's self-reported scale counts conflict across its own pages: 29 research departments versus 28, ten competence centers versus 13.

The German Research Center for Artificial Intelligence (Deutsches Forschungszentrum für Künstliche Intelligenz GmbH) presents itself as a non-profit public-private partnership founded in 1988, with research facilities in Kaiserslautern, Saarbrücken, Bremen, Oldenburg, Osnabrück and Darmstadt, laboratories in Berlin and Lübeck, and a branch office in Trier. Prior BTW reporting has examined how DFKI describes its money (its public funding is described in at least two different accounting languages by the ministry that provides it) and how its governance leaves a dated record (a supervisory board chair succession on 20 November 2025).

Both articles stopped short of a third question that matters to anyone assessing an applied AI institute: what does it actually operate?

The compute base: Pegasus and the DGX A100 timeline

DFKI's most concrete infrastructure artifact is not a press release but a documentation site. Pegasus, described as DFKI's AI high-performance computing cluster hosted at Kaiserslautern, is documented as offering more than 100 PFLOPs for AI workloads, more than 300 GPUs, more than 5,500 CPU cores and more than 40 TiB of RAM, managed with the Slurm workload manager, Enroot container environments and a custom monitoring suite [https://pegasus.dfki.de/]. The documentation goes further than a marketing page would: it publishes a partition table broken out by GPU hardware type — A100 (40 GB and 80 GB variants), B200 (Blackwell), H100, H200, L40S, RTX 3090, RTX A6000, RTX Pro 6000 Blackwell and V100 — with nodes typically of eight GPUs interconnected by NVLink/NVSwitch internally and InfiniBand across nodes, job time limits per partition (one to three days for older hardware, one day for the newest Blackwell and Hopper parts), and published ResNet and Transformer training benchmarks per partition using NGC PyTorch containers [https://pegasus.dfki.de/docs/slurm-cluster/partitions/].

That level of disclosure matters because it is the kind of thing an institution publishes for its users, not for its funders. It describes a system someone has to actually schedule jobs on. Independently of DFKI's own marketing, the NLP group's public repository on HPC infrastructure access documents the practical access procedures — Slurm, partitions, containers — as an operational workflow rather than a claim [https://github.com/MBAZA-NLP/hpc-infrastructure-access/blob/main/README.md].

The cluster also has a dated growth history. On 27 July 2020, DFKI announced that its first NVIDIA DGX A100 had gone into operation at its machine learning computing center in Kaiserslautern, raising cluster processing power from 20 petaFLOPS to 45 petaFLOPS. Each DGX A100 provides five petaFLOPS through eight A100 Tensor Core GPUs, at roughly 1.2 kilowatts per petaFLOPS — a notable efficiency claim at a time when DFKI described prior systems as consuming three to five kW per petaFLOPS. The announcement also noted use of NGC-optimized TensorFlow and PyTorch containers [https://www.dfki.de/en/web/news/nvidia-dgx-a100-20]. A later DFKI announcement recorded the extension of the ML Computing Center with a seventh DGX A100 system, with performance described as rising from the initial 20 petaFLOPS to more than 60 petaFLOPS; that system is tied to the Interactive Machine Learning department and to foundation-model research for health, with Prof. Daniel Sonntag quoted on the goals of interactive deep learning and large-scale pre-trained models [https://www.dfki.de/en/web/news/rechenzentrum]. The IML lab's own page describes the same class of system — eight A100 GPUs, up to 640 GB of total GPU memory, 1.5 petaFLOPS — hosted at the Kaiserslautern deep learning center [https://www.dfki.de/en/web/news/rechenzentrum].

The reconciliation between the 60+ petaFLOPS figure of the DGX A100 era and the 100+ PFLOPs figure in the current Pegasus documentation is not dated anywhere in the public record retrieved for this article. The two figures describe the same Kaiserslautern lineage at different points in its expansion, but no public document marks when the cluster crossed from one description to the other. That gap is worth stating precisely: the numbers are consistent with growth, but the growth itself is not independently dated.

The physical testbeds: MarEH and SmartFactory-KL

Not all of DFKI's infrastructure is computational. In Bremen, the institute's Robotics Innovation Center operates the Maritime Exploration Hall (MarEH), a 1,300 m² facility for testing robotic technologies on and under water, simulating offshore-industry missions. Its main test basin measures 23 m × 19 m × 8 m and holds 3.4 million liters of saltwater (at 18 g of salt per liter). Named equipment includes crane systems rated at 12.5 t and 250 kg, a 500 kg gantry crane, a 5 m × 4 m × 2.2 m glass tank holding 40,000 liters with three glass walls, a light-controlled black tank (3.4 m × 2.6 m × 2.2 m), a VR laboratory with 21.6 m² of projection across seven elements, and a Qualisys 12-camera underwater motion-capture system capturing at 300 Hz with roughly 4 mm accuracy [https://robotik.dfki-bremen.de/en/research/research-facilities-labs/maritime-infrastructure].

One figure requires careful attribution. The MarEH page describes its pressure chamber inconsistently: one section states a limit of up to 6,000 meters (600 bar), another states up to 660 bar — roughly 6,600 meters of water depth. Both descriptions appear on the same DFKI page, so any use of the pressure-chamber specification has to acknowledge that the institute's own documentation does not reconcile it.

The second major physical testbed is SmartFactory-KL, a technology initiative founded in 2005 in which DFKI is a central partner. A DFKI press release dated 26 May 2021 announced that SmartFactory-KL had been selected as a testbed for production within Gaia-X, the European data-infrastructure project. The networked-production setup spans three locations: SmartFactory-KL itself, DFKI's Innovative Factory Systems department, and the WSKL department of TU Kaiserslautern. The chronology given in the release is specific: the Production Level 4 (PL4) vision — which extended the Industrie 4.0 concept of 2011 — dates to 2019, and a PL4 demonstrator presented a possible Gaia-X use case in 2020. That demonstrator already ran in the SmartFactory-KL Innovation Lab; a second demonstrator with a new transport system was under construction at the DFKI building. Keran Sivalingam is named project leader of smartMA-X, the SmartFactory-KL sub-project within the GAIA-X project [https://www.dfki.de/en/web/news/sfkl-testbed-gaiax0]. A companion DFKI page on smartMA-X adds the connection requirement in Sivalingam's own words — "first, we need to connect our demonstrator to the GAIA-X network" — and describes the research into compound versus atomic skills [https://www.dfki.de/en/web/news/gaia-x-projekt-smartma-x-startet-in-kaiserslautern0].

Dated projects: where the deployment story has receipts

The third layer of evidence is the project portfolio, where funding records and hardware specifications give the transfer narrative specific, datable content.

LUMINOUS (Language Augmentation for Humanverse) is the clearest case. The EU grant record under Horizon Europe topic HORIZON-CL4-2023-HUMAN-01-21 (Next Generation eXtended Reality, a Research and Innovation Action) shows EC signature on 23 October 2023 and a run from 1 January 2024 to 31 March 2027 under grant agreement 101135724 [https://cordis.europa.eu/project/id/101135724]. The project combines generative and multimodal large language models with extended reality and zero-shot recognition of unknown objects and scenarios, with three demonstration domains: neurorehabilitation for speech-impaired stroke survivors, immersive industrial safety training, and 3D architectural design review. DFKI's own transfer story on the project is dated 19 August 2024 [https://www.dfki.de/en/web/news/project-luminous-the-next-level-of-augmented-reality]. One discrepancy is on the record: a third-party observatory page lists LUMINOUS as ending in December 2026, while CORDIS — the authoritative grant registry — states 31 March 2027.

LIEREx (Language-Image Embeddings for Robotic Exploration) runs from 1 July 2024 to 31 December 2026. It builds vision-language models such as CLIP into a semantic map that supports queries for arbitrary objects, implemented on a mobile robot and evaluated through goal-oriented indoor exploration. The supporting technical reports document the actual platform: a TIAGo 2 robot with an Ouster OS0 LiDAR and a Femto Bolt ToF RGB-D camera, and a Habitat simulator environment built on Matterport3D and HM3D scenes [https://www.dfki.de/en/web/research/projects-and-publications/project/lierex].

SiSWiss (Secure language models for knowledge management) runs from 1 June 2025 to 31 May 2028, funded under the Federal Ministry of Research, Technology and Space (BMFTR) programme "Sichere Zukunftstechnologien in einer hypervernetzten Welt: Künstliche Intelligenz," coordinated by the L3S Research Center with DFKI, CISPA and Laverana as partners. Its technical premise is specific: multi-persona large language models with a Differential Sensitivity Awareness framework that controls access to sensitive information according to user permissions, trained with reinforcement learning from human feedback and released as open-source methods optimized for the German-speaking market [https://www-live.dfki.de/en/web/research/projects-and-publications/project/siswiss]. The coordinator's own project page and a June 2025 launch announcement confirm the same window and the BMFTR funding line [https://www.l3s.de/research-at-l3s/all-projects/siswiss/].

Two further named projects round out the picture. TSS-MoVe develops a pilot system for automated technology research across all modes of transport, documented on DFKI's project pages and by the federal funding side [https://dfki-nlp.github.io/project/tss-move/]. Genki4Media is a DFKI NLP project under the federal "Gen-KI für den Mittelstand" programme, addressing generative AI for media applications with Fraunhofer FOKUS as a partner [https://dfki-nlp.github.io/project/genki4media/].

Where the record does not reconcile

The same institute whose cluster documentation publishes per-partition benchmarks also publishes scale counts about itself that its own pages do not reconcile. DFKI's Company Profile page reports 29 research departments, 13 competence centers and 8 living labs, with approximately 960 researchers and administrators and 600 graduate students from more than 76 countries working on more than 560 projects [https://www.dfki.de/en/web/about-us/dfki-at-a-glance/company-profile]. The institute's main About page, however, states "ten competence centers" — not 13 — and a mirrored version of that same page reportedly says 28 research departments and ten competence centers. These are not third-party disputes; they are conflicts within DFKI's own self-description, visible in simultaneously published material.

The pattern is worth naming because it defines the evidentiary boundary of everything above. Where DFKI documents a system for its users — a partition table, a benchmark, an access workflow — the record is specific, datable and internally coherent. Where DFKI describes its aggregate scale for funders and the public, the record shifts between pages and versions. Prior BTW coverage found the same asymmetry in its funding language and its governance statistics.

The technical-execution picture assembled here is therefore not a rebuttal of the institute's self-description; it is a demonstration of which parts of that description can be independently carried by named, dated documents — and which parts, so far, cannot.