Scaling MPI Applications on Aurora Supercomputer at Argonne

Huda Ibeid and co-authors present a technical and performance study of the Aurora exascale supercomputer at Argonne National Laboratory.

• Main announcement/action: The paper provides system design details and validated performance results for Aurora, deployed in 2024 at Argonne National Laboratory, describing a system of over ten thousand nodes with six Intel Data Center Max Series GPUs and two Intel Xeon Max Series CPUs per node, connected via HPE Slingshot fabric with nearly 85,000 Cassini NICs and 5,600 Rosetta switches in a dragonfly topology, and demonstrates MPI benchmarks and application scaling on a large fraction of the machine.
• Background and details: The manuscript (submitted 3 Dec 2025 to arXiv) focuses on network fabric validation and presents results for performance benchmarks HPL, HPL-MxP, Graph500, HPCG and applications HACC, AMR-Wind, LAMMPS, FMM, reporting throughput, latency, and bandwidth measurements across large node counts; PDF, HTML, and TeX source links are provided.