Jennifer Doudna, Ph.D. [Credit: Berkeley Lab]
NVIDIA CEO Jensen Huang called the forthcoming Doudna supercomputer “a time machine for science,” in that it can compress the pace of scientific discovery. Michael Dell, Chairman and CEO of Dell Technologies, added that Doudna represents a “shared vision” with the Department of Energy that could “redefine the limits of high-performance computing and drive innovation that accelerates human progress.” Meanwhile, U.S. Secretary of Energy Chris Wright said “The Doudna system represents DOE’s commitment to advancing American leadership in science, AI, and high-performance computing.” He went on to say “It will be a powerhouse for rapid innovation that will transform our efforts to develop abundant, affordable energy supplies and advance breakthroughs in quantum computing. AI is the Manhattan Project of our time, and Doudna will help ensure America’s scientists have the tools they need to win the global race for AI dominance.”
The supercomputer was named after Jennifer Doudna, a professor at the University of California, Berkeley who won the 2020 Nobel Prize in Chemistry for her work on CRISPR. The Berkeley Lab named the machine after her “because of the impact of her work,” said Nick Wright, advanced technologies group lead and Doudna chief architect at National Energy Research Scientific Computing Center (NERSC). “She’s really an inspirational figure. And the practical applications of CRISPR are only just beginning to be understood,” he said.
U.S. Energy Secretary Chris Wright (left) speaks at the announcement of the Doudna supercomputer at Lawrence Berkeley National Laboratory, joined by NVIDIA CEO Jensen Huang (center) and Dell Technologies Senior VP Paul Perez (right).
Enabling integrated, real-time science
The DOE continues to invest in building out progressively more advanced integrated research infrastructure that links instruments, supercomputers, and AI code in one loop. At its center sits NERSC. “We are the mission supercomputing center for the Department of Energy, Office of Science,” Wright said. Researchers who hold a DOE grant that needs heavy computation can apply for time on NERSC’s systems. “As part of that, we tend to buy a supercomputer roughly every five years,” Wright said. Doudna represents the next in line.
With Doudna and prior systems, the DOE is championing an “integrated research infrastructure,” linking experimental facilities directly with supercomputing power. Elizabeth Ball, communication lead at NERSC, described the investment in “integrated research infrastructure.” She continued: “That is a method of connecting experimental facilities like telescopes and fusion research locations to high-performance computing. In this context, experimental locations are connected to us via ESnet, which is the DOE’s High-Performance Network for science.
The system enables the researchers doing experiments to, say, rapidly do simulations or analyze “data very rapidly,” Ball said. “They can get answers back very quickly and see if their experiments and data are valid.” The approach can sometimes accelerate the pace of scientific discovery from weeks or months “down to minutes or hours,” she said.
Wright shared specific scenarios where rapid turnaround time matters: “Essentially, the problem we’re trying to solve is that if you’re running a tokamak experiment, you have a fixed time between shots. Or you’re using a telescope, you get time on that telescope, and you want to turn around the data analysis almost in real-time to point at a different area of the sky, or focus in more detail on the area of the sky you’re looking at currently.”
Supercomputers have shifted from being passive participants to integral parts of the workflow
Under the hood
Doudna will be based on NVIDIA’s forthcoming Vera Rubin architecture, which couples high‑performance CPUs and coherent GPUs so all processors share data directly. The Rubin microarchitecture, named after astrophysicist Vera Rubin, will consist of a GPU named Rubin and a CPU named Vera, manufactured using a 3 nm process with HBM4 memory.
Dell’s liquid‑cooled ORv3 racks and PowerEdge servers supply the plumbing. An announcement notes that Dell Integrated Rack Scalable Systems and PowerEdge servers with NVIDIA accelerators for AI-optimized and compute-optimized workloads will be part of the computer. Also in the mix are the NVIDIA Quantum‑X800 InfiniBand, which accelerates in‑network computing.
“For Doudna, it will be 10 times better at doing science applications than the current machine we have, Perlmutter,” Wright said.
Doudna will also balance traditional HPC and AI needs, Wright explained: “Traditional HPC is usually focused on floating-point 64. AI is FP16 or lower. One of the things we are doing in this machine is trying to make sure we balance those competing needs.”
From quantum simulation and beyond
The center is also significantly ramping up its quantum‑information‑science support. Doudna will support NVIDIA’s CUDA-Q platform, which enables scalable quantum algorithm development, classical modeling and verification of quantum computers at scale, and co-design of future integrated quantum-classical hybrid systems.
NERSC has a relatively nascent quantum information sciences research program that Doudna will significantly expand. Doudna will provide the computational muscle for advanced quantum simulation tools. “We’re supporting quantum computing on the machine through software and policies and training and tools,” Wright said. “The idea there is to use the classical supercomputer to simulate the quantum computer of the future and figure out how to optimally use it and accelerate the time in which a quantum computer will be available.”
The second component focuses on practical integration. “We’re also engaging with the quantum computing research community, with some of our staff ourselves, to understand more deeply how our user community could use a quantum computing resource in the future,” Wright added. “So, modeling quantum computers, we are also looking at how much classical compute resource it needs to ensure a quantum computer can work effectively and efficiently.”
