Training large-scale AI models presents major challenges, particularly regarding computational cost and energy efficiency. This project addresses these issues by developing a new software stack for large-scale deep learning, based on a close integration of algorithmic advances, systems-level optimization, and concrete application use cases. It directly supports the priorities of the NumPEx PEPR program on HPC for AI.

The SAGE-HPC project aims to develop a scalable, open, and interoperable software platform for multifidelity optimization of complex physical problems in exascale high-performance computing (HPC) environments. Solving such optimization problems poses a major scientific challenge due to the complexity of the physical phenomena involved and the computational cost associated with high-fidelity simulations. To overcome this challenge, the project leverages both the coordinated use of variable-fidelity models — where simplified, low-cost models guide the exploration of the solution space, and high-fidelity models are used selectively to refine the results — and the massive exploitation of exascale HPC resources, enabling large-scale parallel processing of these approaches.

The KOKTAILS project aims to enhance the portability of simulation software on Exascale computing architectures, by contributing to the development of a sovereign software stack adapted to GPU-based supercomputers. It is part of the NumPEx PEPR strategy and contributes to French digital sovereignty in high-performance computing (HPC). It includes the development of scalable middleware to guarantee performance portability on various GPU architectures, including European processors such as SiPearl Rhea. The project thus contributes to the transition of existing applications to Exascale computing, through the creation of an open-source ecosystem in line with European sovereignty policy.

This research project addresses the critical transformation underway in radio astronomy, driven by next-generation observatories such as LOFAR2.0 and the SKA. These instruments are producing massive, heterogeneous datasets distributed across multiple sites, which cannot be efficiently handled using legacy data processing approaches. The project aims to overcome these structural bottlenecks by developing a unified, scalable digital platform that federates HPC, cloud, and object storage resources. It will support the execution of complex workflows (including AI-based processing) across heterogeneous infrastructures through modern containerization technologies. Key principles such as data provenance, reproducibility, and energy-aware computing will be integrated to support both interactive and automated scientific workflows.