Challenge:
BSC’s MareNostrum 5 is a EuroHPC pre-exascale supercomputer with a combined peak performance of over 314 PFlops. Managing energy efficiency at this scale — across two highly diverse partitions — is critical for sustainability, system stability, and operational costs.
Solution:
EAR was deployed on both the general-purpose and accelerated partitions of MareNostrum 5. Its full-stack runtime capabilities allow real-time optimization, cluster-wide power capping, and detailed energy monitoring across CPUs and GPUs.
Results:
EAR ensures that BSC can run cutting-edge research workloads within strict energy constraints, while preserving performance and system availability. It empowers BSC to scale sustainably and maintain its leadership in European high-performance computing.
About BSC:
Barcelona Supercomputing Center (BSC) is Spain’s national supercomputing facility and a global leader in scientific computing. It hosts MareNostrum 5, a key component of the European EuroHPC pre-exascale initiative.
Using EAR since: May 2024
Systems: General-purpose partition: 6408 Intel Sapphire Rapids nodes + 72 HBM nodes; Accelerated partition: 1120 nodes with 4× NVIDIA Hopper GPUs each