Department of Energy Announces AI-Integrated Supercomputer for 2026

Historically distinct, scientific computing and artificial intelligence are increasingly converging, a trend exemplified by a powerful new supercomputer planned for Berkeley, California.

The Department of Energy's laboratory near the University of California, Berkeley, announced that Dell Technologies has been chosen to build its next flagship supercomputer, expected to be operational by 2026. This system will incorporate Nvidia processors specifically designed for AI workloads alongside traditional simulations used in energy and scientific research.

Named after Jennifer Doudna, the Berkeley biochemist and 2020 Nobel Prize laureate in chemistry, the new supercomputer is projected to deliver more than ten times the performance of the laboratory's current top system. When fully equipped, it could become the department’s premier resource for tasks such as AI model training, according to Jonathan Carter, associate laboratory director for computing sciences.

This supercomputer distinguishes itself through its technology choices, reflecting a growing government interest in adopting commercial AI innovations. Nvidia chips, prevalent in cloud computing and AI but previously not selected for the department’s three prior leading machines assembled by Hewlett Packard Enterprise, will now be central. Dell, relatively new to the elite supercomputing arena, has demonstrated strength in large-scale commercial AI deployments.

Industry analysts note this selection marks a notable shift, with Dell securing a major contract traditionally dominated by Hewlett Packard Enterprise in the Department of Energy space.

Supercomputers, massive computing installations historically used for tasks like weapons design and code breaking, continue to symbolize technological leadership on the national stage.

The Department of Energy has invested approximately $1.8 billion over eight years to achieve 'exascale' computing performance, exemplified by the $600 million El Capitan supercomputer at Lawrence Livermore National Laboratory. While funding goals for the Berkeley system remain undisclosed, it is poised to be a significant addition to the department’s high-performance computing portfolio.

Energy Secretary Chris Wright, comparing AI’s development to the Manhattan Project, described the forthcoming Doudna supercomputer as a crucial asset in advancing the global AI competition during an announcement event in Berkeley.

Traditionally, supercomputers rely on highly precise 64-bit calculations. However, commercial AI systems often use lower-precision 16-bit or 8-bit operations to boost speed at the expense of some accuracy. The ability to integrate these mixed-precision calculations using Nvidia’s graphics processing units expands the range of computational tasks supercomputers can perform, explained Dion Harris, Nvidia’s head of data center product marketing.

Berkeley’s National Energy Research Scientific Computing Center has previously employed Nvidia GPUs in its Perlmutter system. The Doudna supercomputer will utilize an upcoming version named Rubin and will also incorporate general-purpose Nvidia processors based on British company Arm’s technology, departing from the Intel and AMD chips used in past systems.

Jonathan Carter highlighted that the design of Doudna addresses the diverse needs of the center’s 11,000 users. Beyond traditional applications like fusion reactor modeling, researchers increasingly leverage AI to enhance simulations, such as analyzing how water cools geothermal fields.

Additional appeal for Nvidia lies in its extensive AI software ecosystem, tailored for tasks including modeling future quantum computing architectures, Carter noted.

Dell executives emphasized that winning the Berkeley contract allowed them to create adaptable systems designed to serve a broader range of clients, moving away from the conventional approach of crafting supercomputers tailored to individual laboratories.

Paul Perez, Dell’s senior vice president and senior technology fellow, remarked that the supercomputing market had been on 'autopilot,' and their strategy was to break that cycle with innovative approaches.