09:35 - 09:55
Automated Design Space Exploration of ARM CMN-Based HPC Architectures: A Case Study for the European Rhea Family Processors

Lilia Zaourar, Mohamed Benazouz
French Alternative Energies and Atomic Energy Commission (CEA), France

Valentin Dubois-Fournet, Eric Mercier
Sipearl

The Network-on-Chip (NoC) is a major determinant of memory-system performance in modern high-performance computing (HPC) processors. Configurable coherent interconnects such as Arm Coherent Mesh Networks (CMNs) represent large design spaces involving request and home nodes, memory interfaces, queue resources, and node placement. This paper presents an automated multi-objective design-space exploration methodology combining automated Design Space Exploration (DSE) framework, Arm Performance Models, and industrial architecture-generation tools. The methodology is evaluated on a representative CMN-S3 subsystem defined in the context of future processors of the European Rhea family, exploring the organization of the coherent interconnect and memory subsystem while keeping the compute architecture fixed. Results highlight the importance of balancing request and home/cache resources to sustain DDR bandwidth while controlling implementation cost. The automated exploration identifies the same Pareto optimal configurations as exhaustive evaluation using only 26 of 343 valid configurations, reducing total exploration time from approximately 140 h to 2 h.