Inside the historic Torre Girona chapel, the 19th‑century glass box hides MareNostrum V, one of the world’s top fifteen supercomputers. Built for €200 million, it sits beside a cold‑faced hall that once housed the original 2004 racks. The new machine pushes research limits in AI, physics, and climate science for researchers worldwide and industrial partners.

Unlike cloud instances, MareNostrum V distributes workloads across 8,000 nodes linked by an InfiniBand NDR200 fat‑tree network, ensuring non‑blocking bandwidth. Two partitions drive its power: a General Purpose Partition with 6,408 Intel Sapphire Rapids nodes delivering 45.9 PFlops, and an Accelerated Partition of 1,120 NVIDIA H100 GPUs that tops 260 PFlops for mixed‑precision training and inference.

Recent upgrades log quantum processors alongside classical hardware. Spain’s first gate‑based system and the MareNostrum‑Ona annealer now sit beside the 8,000‑node cluster. When a problem exceeds GPU capacity, the scheduler routes it to these QPUs, creating a hybrid environment that blends classical speed with quantum‑specific algorithms for chemistry, cryptography, and material science projects daily.

Researchers access the system via SSH login nodes that only handle file staging and job submission; compute nodes lack outbound internet, forcing all dependencies to be pre‑installed through a module system. SLURM enforces CPU‑hour budgets, concurrent‑job caps, and strict wall‑time limits, so any mis‑timed script is killed mid‑execution, leaving only log files for debugging today.