Scientific computing and artificial intelligence were once separate worlds, using different kinds of calculations on distinctly different hardware. But the two fields are steadily merging, as shown by a massive new machine coming to Berkeley, Calif.
On Thursday, the Department of Energy’s laboratory near the University of California, Berkeley, said it had selected Dell Technologies to deliver its next flagship supercomputer in 2026. The system will use Nvidia chips tailored for A.I. calculations and the simulations common to energy research and other scientific fields.
Lawrence Berkeley National Laboratory expects the new machine — to be named for Jennifer Doudna, a Berkeley biochemist who shared the 2020 Nobel Prize for chemistry — to offer more than a tenfold speed boost over the lab’s most powerful current system. If fully outfitted, the machine could be the Energy Department’s biggest resource for tasks like training A.I. models, said Jonathan Carter, associate laboratory director for computing sciences at the Berkeley center.
The supercomputer stands out for its technology choices, which indicate the growing desire for government labs to adopt more technologies from commercial A.I. systems. Nvidia chips, though widely used by big cloud companies as well as in supercomputers, were passed over by the Energy Department for three previous record-setting machines that were assembled by Hewlett Packard Enterprise. Dell has hardly been a player in the highest end of the supercomputer market, but it has had success in large commercial A.I. installations.
“HPE has been sweeping the D.O.E. space,” said Addison Snell, the chief executive of Intersect360 Research, which tracks the supercomputer market. “This is a big win for Dell.”
Supercomputers — which are computing systems that take up entire rooms first used for jobs like designing weapons and cracking codes — have long been symbols for national prowess in technology.
The Energy Department, which typically purchases the government’s biggest computers, devoted $1.8 billion over eight years to reach what the industry calls “exascale” performance, topped by a $600 million supercomputer at Lawrence Livermore National Laboratory called El Capitan. The Trump administration has not set a comparable funding target, nor has a price for the Berkeley system been released.
But Chris Wright, secretary of energy, who has compared A.I.’s development to the Manhattan Project, called the Doudna machine a key tool for winning the global A.I. race in remarks prepared for a Thursday event in Berkeley to announce the system.
Supercomputers have historically relied on precise calculations, handling data in what are known as 64-bit chunks. Commercial A.I. systems often use simpler 16-bit or 8-bit instructions, sacrificing accuracy for higher speed. Being able to use a mix of calculations in supercomputers powered by graphics processing units, the chips Nvidia sells for A.I., opens up many new kinds of computing jobs, said Dion Harris, Nvidia’s head of data center product marketing.
The National Energy Research Scientific Computing Center at Berkeley already used Nvidia’s GPUs in a system called Perlmutter. Doudna will use a future version called Rubin but, in a key technical departure, will also use a general-purpose Nvidia processor based on technology from the British company Arm rather than Intel and Advanced Micro Devices chips.
Mr. Carter, the laboratory director, said Doudna’s design was partly driven by the need to serve various tasks used by the center’s 11,000 users. Besides historical jobs like modeling how fusion reactors work, researchers are increasingly using A.I. to improve simulations of phenomena such as how water removes heat from geothermal fields, he said.
Other attractions of using Nvidia included its array of A.I. software, tailored for tasks like modeling future quantum computers, Mr. Carter said.
Executives at Dell, which outbid other vendors for the Berkeley system, said the contest offered a chance to design systems that could be adapted to serve many customers — breaking from a tradition of customizing creations for individual labs.
“This market had shifted into some form of autopilot,” said Paul Perez, a senior vice president and senior technology fellow at Dell. “What we did was disengage the autopilot.”
The post Energy Dept. Unveils Supercomputer That Merges With A.I. appeared first on New York Times.
