Aurora Reaches Exascale, Leads in AI Performance

In 2024, the system surpassed the exascale barrier, demonstrated its world-class AI capabilities, and completed critical preparations for its release to the research community.

Built in partnership with Intel and HPE, Aurora is equipped with 63,744 GPUs and 84,992 network endpoints, making it one of the largest supercomputer installations to date.

Over the past year, the ALCF made significant progress in preparing its Aurora exascale supercomputer for deployment in January 2025. In close collaboration with Intel and Hewlett Packard Enterprise, the team completed extensive system validation, verification, and scale-up efforts to pave the way for system acceptance.

Dedicated to open science, Aurora will provide the research community with powerful simulation, AI, and data analysis capabilities to drive breakthroughs in physics, engineering, materials science, and other domains.

Breaking the Exascale Barrier

Aurora demonstrated its capabilities in various performance benchmarks revealed at the ISC and SC conferences, further cementing its place among the world’s most powerful supercomputers. The ALCF system officially broke the exascale barrier in June, achieving 1.012 exaflops on the High Performance LINPACK (HPL) benchmark. Aurora also set a new record for AI performance, registering 11.6 exaflops on the HPL-MxP mixed-precision benchmark. Its strengths in data-intensive applications were further highlighted with leading results on the Graph500 and HPCG benchmarks, while its storage system, DAOS, retained the top ranking on the IO500 production list.

Together with Oak Ridge National Laboratory’s Frontier and Lawrence Livermore National Laboratory’s El Capitan, DOE is now home to the world’s first three exascale systems. These machines not only mark the first to reach exascale but are also the three fastest supercomputers on the TOP500 List.

Built in partnership with Intel and HPE, Aurora’s architecture represents a first-of-its-kind deployment, integrating cutting-edge technologies at an unprecedented scale. Equipped with 63,744 GPUs and 84,992 network endpoints, the system is designed to tackle complex computational challenges in ways that were previously unimaginable.

Argonne's Rick Stevens discusses how researchers will use Aurora to revolutionize the use of AI for science.

World-Class Simulation, AI, and Data Capabilities

Aurora’s computing power and advanced capabilities are expected to transform research across a wide range of scientific domains. Ahead of the system’s deployment, teams participating in DOE’s Exascale Computing Project (ECP) the ALCF’s Aurora Early Science Program (ESP) have demonstrated its potential in training large-scale AI models and carrying out extreme-scale modeling and simulation campaigns.

One key target involves the development of AI-driven scientific models that can accelerate discovery across multiple disciplines, including materials design, drug development, and energy research. The system is also being prepared to support high-fidelity simulations of complex systems, such as the human circulatory system, nuclear reactors, and supernovae, to gain new insights into their behavior. Additionally, its capacity to process massive datasets will be critical for analyzing the growing data streams from large-scale research facilities such as Argonne’s Advanced Photon Source and CERN’s Large Hadron Collider.

Preparing for Science on Day One

Bringing a system of this scale online has required close collaboration among the ALCF, Intel, HPE, and researchers from the DOE’s Exascale Computing Project and Aurora Early Science Program. Throughout 2024, these teams worked to optimize codes and stress-test the system, ensuring it would be ready for science from day one of production.

A co-design approach was essential in this effort, with hardware and software developed in tandem to maximize performance and usability. As part of this process, researchers ran early science applications to fine-tune their software for Aurora’s architecture, resulting in a suite of computational tools that will be ready to accelerate discoveries as soon as the system becomes fully operational.

With acceptance testing and final preparations completed in December, Aurora is poised to drive significant advancements in scientific computing. Its advanced capabilities are empowering researchers to tackle some of the most challenging problems in science and engineering at an unprecedented scale and speed, unlocking discoveries and insights more quickly than ever before.

ALCF-4: Planning for Argonne’s Next-Generation Supercomputer

Q3 2024 Q2 2025 Q3 2025 Q4 2025 Q4 2028
Q3 2024 Published draft technical specifications/requirements

Technical design review
Q2 2025 CD-1 review Q3 2025 RFP release Q4 2025 Vendor selection Q4 2028 System delivery

Looking beyond Aurora, the ALCF continued the efforts to prepare for its next-generation leadership supercomputer. Known as ALCF-4, the project released the draft technical specifications for the system in June 2024, soliciting comments and questions from the HPC community and potential vendors. They will use this input to prepare the request for proposals, which will be issued in 2025.

The ALCF-4 team is targeting 2028–2029 for the deployment of the facility’s next production system. The project’s goals include enabling a significant improvement in application performance over Aurora, continuing to support traditional HPC workloads alongside AI and data-intensive computations, and investigating the potential to accelerate the deployment and realization of new technologies.