The Bulgarian AI Factory BRAIN++ continues to strengthen its integration into the European scientific and infrastructure ecosystem. From 26 to 28 May 2026, a leading AI expert from our team took part in the HACE 2026 conference – “HPC/AI Hybridization” – hosted by IRIT, the Institut de Recherche en Informatique de Toulouse, in Toulouse, France.
The event was a key component of the thematic trimester organized by AISSAI and PEPR NumPEx, dedicated to the convergence of high-performance computing and artificial intelligence in the era of Exascale computing systems.
Key technological directions with direct relevance to BRAIN++
The conference programme addressed some of the most advanced software and mathematical approaches for accelerating computations through AI. The methodologies and technologies explored during the event will be directly applied across two core areas of BRAIN++ activity: the sizing and design of the supercomputing infrastructure under the EuroHPC JU framework, including LOT 1 and LOT 2, and the further development of CHAIN++, our federated AI data management platform.
Based on the topics discussed, several follow-up actions are already being considered.
Within BRAIN++, we are launching a review of whether MLIR-based compiler tools, including SYCL support for heterogeneous environments, perform as expected within the technical proposals submitted by vendors for our HPC/AI procurement procedure.
In the upcoming version of the CHAIN++ White Paper, we will add a dedicated workflow management scenario for coupled AI-physics applications, presented as a proof-of-concept.
Laure Raynaud from Météo-France demonstrated how AI surrogate models can be embedded in real time into operational pipelines alongside classical numerical weather prediction systems. BRAIN++ will study this architecture, including data frequency, retraining requirements and HPC resource needs, as a reference model for our forecasting activities.
Alban Farchi, representing ECMWF/CEREA, presented approaches for accelerating computationally intensive components of atmospheric modelling through AI on GPU infrastructure, within Category B use cases.
These examples confirm that the “AI for HPC” concept has reached operational maturity. They also support the need to size the BRAIN++ GPU infrastructure for workloads beyond traditional LLM tasks, which require different memory profiles and batch-size configurations.
Building partnerships with leading European supercomputing centres
During the event, the BRAIN++ team also held working meetings with representatives of leading European and international supercomputing organizations, including Forschungszentrum Jülich, represented by Andreas Lintermann; Barcelona Supercomputing Center, represented by Oriol Lehmkuhl; and LLNL/CASC, represented by Youngsoo Choi, with discussions focused on his pioneering DD-FEM approach for foundation-style PDE solvers.
These organizations are natural partners for future benchmarking activities, enabling BRAIN++ to compare its infrastructure design against operational EuroHPC installations.
BRAIN++ will continue to bring the latest European innovations into Bulgaria’s AI and supercomputing infrastructure. Follow us for the next steps in implementing these technologies at national level.