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.
| Hopper (supercomputer) | |
|---|---|
| Name | Hopper |
| Manufacturer | NVIDIA |
| Release | 2022 |
| Type | Supercomputer |
| Flops | 70e15 |
| Memory | 80 GB HBM3 per GPU |
| Nodes | 80 |
| Location | Lawrence Berkeley National Laboratory |
Hopper (supercomputer) is an exascale-class supercomputer deployed at Lawrence Berkeley National Laboratory to accelerate research across artificial intelligence, computational fluid dynamics, climate science, and molecular dynamics. Built by NVIDIA with system integration by Hewlett Packard Enterprise and collaboration from National Energy Research Scientific Computing Center engineers, Hopper pairs advanced GPU microarchitecture with high-speed interconnects to deliver peak performance for mixed-precision workloads. The system supports multidisciplinary projects led by researchers from institutions such as University of California, Berkeley, Massachusetts Institute of Technology, Stanford University, Oak Ridge National Laboratory, and Lawrence Livermore National Laboratory.
Hopper is designed to serve national-scale science programs funded by United States Department of Energy offices including Office of Science and Advanced Scientific Computing Research. Announced following milestones set by Exascale Computing Project roadmaps, Hopper aims to complement capacity systems like Summit (supercomputer), Frontier (supercomputer), and Perlmutter (supercomputer). The platform emphasizes tensor core acceleration, mixed-precision arithmetic, and large model training for projects affiliated with Human Brain Project, National Institutes of Health, NASA, and industrial partners such as Intel Corporation and Google research groups.
Hopper is built around NVIDIA Hopper GPU microarchitecture accelerators featuring Hopper H100 or successor dies employing memory, offering up to 80 GB per GPU. Nodes integrate NVIDIA Mellanox HDR and NDR InfiniBand interconnects for low-latency, high-bandwidth communication comparable to networks in Frontera (supercomputer), Diamond (supercomputer), and Tianhe-2. The system uses dense compute cabinets from Hewlett Packard Enterprise, draws power and cooling similar to designs at Argonne National Laboratory and Los Alamos National Laboratory, and leverages liquid cooling technologies pioneered in projects at CERN and IBM Research. Storage and parallel I/O are provided via BeeGFS, Lustre, and burst-buffer tiers modeled after deployments on Summit (supercomputer) and Perlmutter (supercomputer). System management integrates tools from OpenHPC, Slurm Workload Manager, and NVIDIA NGC catalogs.
Hopper achieves multi-petaflop to tens-of-petaflop sustained performance on double-precision workloads and substantially higher performance on mixed-precision and AI kernels, rivaling early exascale-class systems such as Frontier (supercomputer) and research prototypes from Oak Ridge National Laboratory. It posts leading scores on LINPACK-style benchmarks adapted for GPU-heavy systems and on AI-centric benchmarks developed by MLPerf and consortiums that include teams from Facebook AI Research, DeepMind, and OpenAI. Benchmarking uses scientific applications from Argonne Leadership Computing Facility, NERSC workload suites, and community codes like LAMMPS, GROMACS, NWChem, WRF, CESM, and AMBER to characterize performance across physics, chemistry, and climate workloads.
The software stack supports NVIDIA CUDA, Rocm-compatible middleware from AMD collaborators in mixed environments, and high-level frameworks including TensorFlow, PyTorch, JAX, Hugging Face, and MXNet. Scientific libraries encompass cuBLAS, cuDNN, cuFFT, MAGMA, and community projects such as PETSc, Trilinos, SciPy, and Numba. Compiler toolchains incorporate GCC, LLVM, and vendor compilers from Intel Corporation for CPU-GPU co-optimization. Workflow and data management integrate Dask, Kubernetes, Singularity, and Docker images distributed via NERSC Perlmutter-style registries. The environment supports reproducible pipelines linked to initiatives at DOE Office of Science and collaborations with European Centre for Medium-Range Weather Forecasts teams.
Hopper is housed at the National Energy Research Scientific Computing Center facility within Lawrence Berkeley National Laboratory, with facility engineering influenced by designs at Oak Ridge National Laboratory and Argonne National Laboratory. The deployment required electrical upgrades resembling those for Frontier (supercomputer) and advanced cooling plants inspired by installations at CERN and high-density datacenters operated by Microsoft Azure and Amazon Web Services. Security and user access policies align with protocols used by DOE National Laboratories and research access programs managed by XSEDE and Science Gateways. The system supports multi-tenant scheduling and allocations coordinated with national user programs and collaborations with universities including California Institute of Technology, Columbia University, Princeton University, and Yale University.
Hopper accelerates machine learning research for large language models developed by groups such as OpenAI, Google DeepMind, and Meta AI; supports climate modeling collaborations with National Oceanic and Atmospheric Administration, Intergovernmental Panel on Climate Change, and NOAA researchers; enables computational chemistry projects from Pfizer, BASF, and academic consortia; and advances astrophysics simulations used by teams at Space Telescope Science Institute and Jet Propulsion Laboratory. Use cases include high-resolution earth-system modeling for IPCC assessments, vaccine and drug discovery projects partnered with National Institutes of Health, and materials discovery programs linked to Materials Project and DOE Office of Technology Transitions.
Development of Hopper followed NVIDIA’s roadmap after GPUs such as Volta and Ampere generations; the microarchitecture name honors computer scientist Grace Hopper, whose legacy includes work with UNIVAC I and contributions to COBOL. The system’s procurement and integration involved collaborations among DOE, NVIDIA, Hewlett Packard Enterprise, and regional research consortia, reflecting precedents set by procurement of systems like Perlmutter (supercomputer), Summit (supercomputer), and Frontera (supercomputer). Early deployments supported pilot projects from Berkeley Lab scientists and partner institutions including Lawrence Livermore National Laboratory and Sandia National Laboratories.