This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| NERSC Perlmutter | |
|---|---|
| Name | Perlmutter |
| Location | Berkeley, California |
| Operator | Lawrence Berkeley National Laboratory |
| Sponsor | United States Department of Energy |
| Date operational | 2021 |
| Peak performance | 70 PF/s (mixed precision) |
| Architecture | HPE Cray Shasta chassis with NVIDIA Ampere GPUs and AMD EPYC CPUs |
| Memory | large GPU and CPU memory pools |
| Interconnect | NVIDIA Mellanox Slingshot interconnect |
| File system | Hewlett Packard Enterprise Lustre or DAOS tiering |
NERSC Perlmutter
Perlmutter is a leadership-class supercomputer hosted at Lawrence Berkeley National Laboratory and funded by the United States Department of Energy, designed to accelerate computational science for projects ranging from astrophysics to climate modeling. It succeeds prior systems at the National Energy Research Scientific Computing Center and combines technologies from HPE Cray, NVIDIA, and AMD to deliver heterogeneous CPU–GPU workflows for workflows used by researchers at Oak Ridge National Laboratory, Argonne National Laboratory, and university consortia. The system supports DOE missions coordinated with programs such as the Advanced Scientific Computing Research office and integrates with national facilities including the Energy Sciences Network.
Perlmutter was procured through a collaboration among Lawrence Berkeley National Laboratory, Hewlett Packard Enterprise, and NVIDIA Corporation, with system architecture shaped by requirements from the Exascale Computing Project and the DOE Office of Science. The system delivers mixed-precision performance intended for applications developed at institutions like Massachusetts Institute of Technology, Stanford University, University of California, Berkeley, and Princeton University. Perlmutter serves user communities affiliated with centers such as the National Renewable Energy Laboratory, Brookhaven National Laboratory, Fermi National Accelerator Laboratory, and international partners including CERN and Max Planck Society.
Perlmutter’s design emphasizes heterogeneous compute nodes that combine AMD EPYC CPU sockets with NVIDIA A100 or Ampere-series GPUs in an HPE Cray Shasta chassis, reflecting architecture choices similar to machines at Oak Ridge Leadership Computing Facility and Argonne Leadership Computing Facility. The network fabric leverages Mellanox Technologies Slingshot technology interoperable with fabrics used by Frontera (supercomputer) and Summit (supercomputer) design studies. Storage architecture integrates parallel file systems and object stores comparable to deployments at Oak Ridge National Laboratory and National Institute for Computational Sciences, enabling data-staging workflows used by scientists at University of Chicago and California Institute of Technology.
Compute nodes combine AMD EPYC processors with multiple NVIDIA A100 GPUs employing high-bandwidth NVLink and PCIe interconnect paradigms similar to those in Hopper (supercomputer) proposals. The chassis and rack-level integration are provided by HPE under contracts with DOE procurement offices and include management firmware informed by vendors like Intel Corporation for control-plane components. The high-performance network uses Mellanox Slingshot switches and cables also found in deployments at TACC and CSCS (Swiss National Supercomputing Centre), while storage tiers incorporate technologies from DataDirect Networks, NetApp, and object storage patterns analogous to Amazon S3 architectures (as studied by researchers at Carnegie Mellon University).
The software environment supports programming models common to DOE leadership systems: CUDA for GPU kernels, OpenMP for shared-memory parallelism, MPI for distributed workloads, and newer paradigms like Kokkos and RAJA for performance portability used by teams at Sandia National Laboratories and Los Alamos National Laboratory. Container runtimes such as Singularity and Docker-compatible OCI images are supported in curated modules maintained by NERSC operations staff and research software engineers from University of California, Berkeley and Lawrence Livermore National Laboratory. Resource managers and schedulers integrate components like Slurm and telemetry stacks referenced in deployments at NERSC predecessors and at National Center for Supercomputing Applications.
Perlmutter’s mixed-precision peak was reported in procurement benchmarks aligning with LINPACK-style measurements used by the TOP500 list and with application-level benchmarks such as those used by HPL-AI and the Graph500 community. Comparative performance studies reference machines like Summit (supercomputer), Fugaku, and Frontier (supercomputer) for scaling characteristics across domains pursued by researchers at Columbia University and Yale University. Performance engineering efforts involve teams from NVIDIA Research, AMD Research, and academic centers such as University of Illinois Urbana-Champaign to optimize kernels in libraries like cuBLAS, cuDNN, and vendor-tuned BLAS implementations.
The procurement followed DOE acquisition processes with awardees including Hewlett Packard Enterprise and NVIDIA Corporation, coordinated by procurement officials at Lawrence Berkeley National Laboratory and informed by roadmap discussions with the Exascale Project Office and vendor engagement from Cray Inc.. Deployment activities drew on site planning teams with expertise from Pacific Gas and Electric Company for power provisioning and from Schweitzer Engineering Laboratories for electrical distribution modeling used also at Argonne and Oak Ridge. Operational teams include systems engineers and user support staff who coordinate allocations with programs like the ALCC and institutional collaborators at University of California campuses.
Perlmutter targets high-impact science in areas such as cosmology simulations used by groups at Lawrence Berkeley National Laboratory, Princeton University, and Flatiron Institute, climate and earth system modeling pursued at NOAA and NASA Goddard Space Flight Center, materials science workflows from SLAC National Accelerator Laboratory and Brookhaven National Laboratory, and fusion science simulations supported by Princeton Plasma Physics Laboratory and General Atomics. Application teams leverage community codes and frameworks including GROMACS for molecular dynamics, LAMMPS for materials modeling, WRF for atmospheric modeling, and machine-learning stacks from TensorFlow and PyTorch developed in collaboration with researchers at Google and Facebook AI Research.
Category:Supercomputers Category:Lawrence Berkeley National Laboratory