Summit (supercomputer)
| Industry | Supercomputing |
|---|---|
| Founded | November 2014 |
| Founder | IBM, Nvidia, Mellanox |
| Defunct | November 15, 2024 |
| Fate | Retired |
| Headquarters | Oak Ridge, Tennessee , United States |
Area served | Scientific research |
| Products | Supercomputer |
| Owner | Oak Ridge National Laboratory |
| Parent | United States Department of Energy |
Summit, also known as OLCF-4, was a high-performance supercomputer engineered by IBM for the Oak Ridge Leadership Computing Facility (OLCF) at the Oak Ridge National Laboratory in the United States. It maintained the top position on the TOP500 rankings from June 2018 until June 2020. As of June 2024, its performance on the LINPACK benchmark reached 148.6 petaFLOPS. Summit was officially retired on November 15, 2024.
By November 2019, Summit was ranked fifth worldwide in terms of energy efficiency among supercomputers, achieving a power efficiency rating of 14.668 gigaFLOPS per watt. It was the pioneering supercomputer to surpass exaflop speeds based on a non-traditional metric, attaining 1.88 exaflops during genomic data processing and projected to reach up to 3.3 exaflops through mixed-precision computational methods.
History
In November 2014, IBM, Nvidia, and Mellanox received a $325 million contract from the United States Department of Energy to create the Summit and Sierra supercomputers. Summit, stationed at Oak Ridge National Laboratory in Tennessee, is dedicated to civilian scientific investigations, whereas Sierra, located at Lawrence Livermore National Laboratory in California, is primarily used for nuclear weapons simulation.
The Summit system occupies about 5,600 square feet and incorporates approximately 219 kilometers of cabling. It facilitates research across diverse disciplines including cosmology, medicine, and climate science.
The 2015 Collaboration of Oak Ridge, Argonne and Lawrence Livermore (CORAL) initiative also included a third supercomputer named Aurora, initially planned for Argonne National Laboratory. By 2018, Aurora was re-envisioned as an exascale computing project alongside Frontier and El Capitan, with an anticipated completion date near 2021. Aurora reached completion in late 2022.
Uses
Summit serves as a computational platform for investigations in energy research, artificial intelligence development, human health studies, among other areas. Its applications extend to simulating earthquakes and extreme weather events, advancing materials science, genomic analysis, and estimating neutrino lifetimes.
Design
Each of Summit’s 4,608 compute nodes is equipped with two IBM POWER9 processors and six Nvidia Tesla GPUs. The system provides over 600 GB of shared memory (comprising 96 GB HBM2 plus 512 GB DDR4) accessible across all processors, along with 800 GB of non-volatile RAM that can function as either a burst buffer or extended memory. The POWER9 CPUs are linked to Nvidia Volta GPUs through Nvidia’s high-bandwidth NVLink technology, enabling a heterogeneous computing architecture.
Nodes are interconnected via a non-blocking fat-tree network topology utilizing dual-rail Mellanox EDR InfiniBand connections that deliver up to 200 Gbit/s bandwidth between nodes. This infrastructure supports in-network computing capabilities for MPI and PGAS communication protocols.
Summit’s storage subsystem includes a rapid in-node SSD layer complemented by a central parallel filesystem (IBM GPFS) with a total capacity of 250 petabytes. This design offers seamless data access with peak throughput rates reaching 2.5 terabytes per second for single-stream reads and one terabyte per second for one million file operations. It was among the earliest supercomputers optimized for the intensive metadata demands characteristic of AI and machine learning workloads, achieving up to 2.6 million file creations per second involving files sized at 32k.
See also
- Titan (supercomputer) – OLCF-3
- Frontier (supercomputer) – OLCF-5
- TOP500
- OpenBMC
- Red Hat Enterprise Linux