What Is Exascale Computing?

Exascale computing refers to computing systems capable of performing at least one exaflop, or 10¹⁸ floating-point operations per second, a performance threshold roughly ten times that of the petascale systems that preceded it. Achieving exascale performance requires faster processors alongside coordinated advances in parallel architecture, memory hierarchy, power management, and system software, because no single component improvement is sufficient to bridge the gap from petascale to exascale. The term is also used more broadly to describe the research programs, software ecosystems, and application codes developed in anticipation of and for deployment on these systems.

The United States Department of Energy (DOE) launched the Exascale Computing Project (ECP) in 2016 as a $1.8 billion joint effort between the Office of Science and the National Nuclear Security Administration. The ECP ran through 2024 and supported the development of three flagship systems: Frontier at Oak Ridge National Laboratory, Aurora at Argonne National Laboratory, and El Capitan at Lawrence Livermore National Laboratory. Frontier, deployed in 2022, was the first confirmed exascale system; El Capitan, completed in 2024, is currently the fastest.

Hardware Architecture

Exascale systems achieve their performance through massive parallelism rather than clock speed increases. Frontier, built by HPE on AMD EPYC processors and Radeon Instinct GPU accelerators, connects more than 9,400 nodes via a Slingshot high-speed interconnect and achieves 1.1 exaflops on the HPL (High Performance Linpack) benchmark. Aurora at Argonne National Laboratory employs Intel Xeon processors paired with Intel Ponte Vecchio GPU accelerators, with more than 60,000 GPUs contributing to its computing capability. The DOE's explanation of exascale computing describes how these heterogeneous node designs, combining CPU cores for general computation with GPU accelerators for throughput-intensive workloads, allow the systems to efficiently handle both simulation-heavy and data-intensive tasks.

Software and Programming Challenges

Developing applications for exascale systems requires rethinking software design at every layer of the stack. Traditional MPI-based parallelism is insufficient at this scale; applications must expose hierarchical parallelism that maps to GPU cores, CPU vector units, and distributed nodes simultaneously. The ECP invested heavily in programming model frameworks, including OpenMP for node-level threading, HIP and SYCL for GPU portability, and the Kokkos and RAJA abstraction libraries that allow a single source codebase to target different hardware architectures. Memory bandwidth, not raw arithmetic throughput, is often the limiting factor for scientific applications, leading to increased use of mixed-precision arithmetic and compressed data representations. The Exascale Computing Project at Oak Ridge National Laboratory provides details on the application codes ported to Frontier and the performance improvements achieved relative to the prior generation of systems.

Scientific Applications

Exascale systems were built to address specific scientific challenges that were computationally out of reach at petascale. Climate modeling at exascale permits global atmosphere-ocean simulations at kilometer-scale grid resolution, capturing mesoscale dynamics that coarser models parameterize. Fusion energy simulations using particle-in-cell codes can now model turbulent transport in tokamak plasmas at spatial and temporal scales relevant to confinement physics. Molecular dynamics simulations of protein folding and materials formation can evolve systems containing hundreds of millions of atoms over microsecond timescales. The Argonne Leadership Computing Facility's documentation on Aurora outlines the scientific and AI workloads the system is designed to support, including large-scale training of foundation models alongside traditional simulation codes.

Applications

Exascale computing has applications in a range of fields, including:

• Climate and weather modeling at fine spatial resolutions for improved forecast accuracy
• Nuclear stockpile stewardship simulations replacing physical testing
• Drug discovery through large-scale molecular dynamics and free energy calculations
• AI model training at scales that complement experimental scientific data collection
• Cosmological simulations tracing the formation of large-scale structure in the universe