The team led by Prof. Dr. Jürgen Gall, Principal Investigator at the Lamarr Institute for Machine Learning and Artificial Intelligence and head of the Computer Vision Research Group at the University of Bonn, has been awarded the MARVIN Cup. The award recognized the “RiverMamba” project, which develops approaches for the efficient modeling and processing of large, heterogeneous time-series data, with applications in flood forecasting. “The event has once again demonstrated that the HPC infrastructure and Marvin are essential for conducting excellent research at the University of Bonn. I really appreciated the transdisciplinary nature of the event,” says Prof. Dr. Jürgen Gall. The award recognizes work that makes particularly compelling use of the potential of modern AI infrastructures for scalable yet resource-efficient models.
MARVIN: Infrastructure and Scale for Scalable AI
MARVIN is the central high-performance computing infrastructure for AI research at the University of Bonn. It enables scalable experiments, as well as the development and evaluation of new modeling approaches under realistic conditions. As an HPC environment specifically designed for data- and computation-intensive machine learning processes, MARVIN supports, in particular, the training and evaluation of large models as well as the processing of extensive datasets. “For the past two years, Marvin has been a driving force behind innovation in our research at the University of Bonn. […] Marvin is opening up opportunities for our researchers that were unimaginable just two years ago,” says Prof. Dr. Maren Bennewitz.
As model complexity and data volumes grow, specialized computing infrastructure is becoming the limiting factor in scientific discovery. Conventional infrastructure quickly reaches its limits, particularly when training modern sequence models and processing continuous sensor data. “Marvin embodies our commitment to making high-performance computing truly accessible for research. […] These projects […] impressively demonstrate the scientific potential of this infrastructure,” adds Prof. Dr. Petra Mutzel. The MARVIN Award recognizes projects that make particularly innovative and effective use of this infrastructure.
Efficient model architectures for real-world applications
The award-winning “RiverMamba” project addresses a key challenge in current AI research: the efficient processing of large and complex time-series data. Building on novel model architectures—particularly those of the Mamba class—the approach enables the analysis of long data sequences with significantly reduced computational effort without sacrificing performance. Unlike traditional transformer architectures, this approach scales more effectively with sequence length, thereby opening up new possibilities for data-intensive applications. RiverMamba thus exemplifies a current shift in the field: rather than focusing primarily on ever-larger models, the emphasis is now on developing efficient, scalable architectures that are practical under real-world conditions. Applications such as flood forecasting benefit directly from this development, as they must continuously integrate large data streams. The project was first presented at NeurIPS last fall and combines methodological innovation with concrete practical applications. As part of the 2026 Bonn Science Night, RiverMamba will also be introduced to a wider audience in the Lamarr Institute’s AI tent.