Frontier Supercomputer — Oak Ridge National Laboratory

Frontier, housed at Oak Ridge National Laboratory in Tennessee, became the world’s first confirmed exascale supercomputer when it achieved 1.1 exaflops (1.1 quintillion floating-point operations per second) in May 2022, topping the TOP500 list and dethroning Japan’s Fugaku. It remains one of the fastest supercomputers ever built and a central pillar of US AI and scientific computing infrastructure.

Hardware: Frontier runs on 9,408 AMD EPYC 64-core CPUs paired with 37,632 AMD Instinct MI250X GPU accelerators. The AMD MI250X units are purpose-built for high-performance computing and deep learning training at scale. The system has 4.6 petabytes of total storage through a Lustre-based parallel filesystem.

Power: The system draws approximately 21 megawatts at peak operation, powered through the Tennessee Valley Authority grid. ORNL is pursuing renewable power agreements and efficiency upgrades. The facility uses a direct water-cooled architecture that substantially reduces the cooling overhead compared to traditional air-cooled systems.

AI workloads: Frontier is used for foundation model training, protein folding simulations (including AlphaFold-class workloads), climate modeling, materials science, and nuclear physics research. The US Department of Energy’s INCITE (Innovative and Novel Computational Impact on Theory and Experiment) program allocates compute time to research teams globally, including AI labs working on scientific applications of large language models.

Operator structure: ORNL is a US Department of Energy national laboratory managed by UT-Battelle LLC — a partnership between the University of Tennessee and Battelle Memorial Institute. HPE built and supplies the Cray EX supercomputer infrastructure. Frontier cost approximately $600 million to design, build, and deploy.

What comes next: ORNL’s next system, Frontier’s successor (“Centennial”), is in planning for the early 2030s and expected to reach zettascale or beyond. In the near term, Frontier remains the reference architecture for US AI training infrastructure in the scientific computing domain.