Inside an unremarkable facility at Kirtland Air Force Base in New Mexico’s high desert, sophisticated liquid-cooled computing systems quietly work through some of America’s most challenging mathematical calculations: modeling hypersonic nuclear weapons traveling through Earth’s atmosphere and simulating nuclear warhead interactions.
For over ten years, the processors powering this classified and intensive work originated from major semiconductor companies such as Nvidia or Advanced Micro Devices.
However, as these corporations focus more on creating processors for artificial intelligence applications while dealing with supply constraints, administrators overseeing the systems at Sandia National Laboratories – which runs the computers at Kirtland and serves as one of three U.S. facilities responsible for creating and maintaining America’s nuclear weapons stockpile – face growing uncertainty about securing computing resources for their high-precision scientific calculations.
“The pressure we’re feeling right now is on the computing front and also from the supply chain,” said Steve Monk, the manager of Sandia’s high-performance computing team, explaining the challenge of getting chips that meet his needs. “Looking to the future, it’s a bit stressful in terms of our ability to deliver to the mission.”
The laboratory’s situation demonstrates how the competition for superior AI processors has unexpectedly created opportunities for smaller companies like NextSilicon, an Israeli startup whose processors are undergoing evaluation through a Sandia program, to enter markets previously controlled by industry giants. This also highlights Sandia’s role in nurturing and developing new computing technologies, having previously collaborated extensively with Nvidia during the company’s rise in supercomputing and continuing to work with Nvidia on innovative memory solutions.
A primary worry for Sandia officials involves double-precision floating point computation, a technical concept referring to the ability to calculate extremely large and small numbers while maintaining accuracy and avoiding rounding mistakes. For years, Nvidia and AMD competed to advance this type of computing speed, securing supercomputing agreements with universities and government laboratories.
However, AI applications don’t require double-precision computing to the same extent as physics simulations. Although AMD is developing a chip version targeted at scientific computing, the double-precision capabilities of Nvidia’s upcoming Rubin processors have decreased by certain standards, causing concern among numerous scientists in the high-performance computing field, according to Ian Cutress, chief analyst at More Than Moore, a chip consulting firm.
Daniel Ernst, senior director of supercomputing products at Nvidia, said the company remains committed to scientific computing, aiming to create a balanced chip that can run real-world scientific applications alongside AI work.
The evolving processor market has led Sandia officials to evaluate products from newcomers like NextSilicon, whose processor employs an entirely different computational method than graphics processing units (GPUs) or central processing units (CPUs) from Nvidia and AMD.
On Monday, Sandia, NextSilicon and Penguin Solutions, the company that integrated NextSilicon’s processors into a supercomputer, announced the systems have achieved an important technical benchmark using comprehensive supercomputing evaluations that qualify the processors for potential government system deployment.
This achievement positions NextSilicon’s processors for an autumn decision regarding whether to begin testing them with more challenging computational problems that closely mirror the nuclear security work they would ultimately need to perform.
NextSilicon’s processors can execute double-precision computing and are engineered to reconfigure themselves dynamically for improved efficiency. The company’s chips conserve power by utilizing a data flow architecture that reduces the time and energy spent moving data between the computing system’s memory.
Sandia’s collaboration with chip companies frequently helps technologies gain widespread adoption. Liquid cooling systems for processors were considered unusual when Sandia began encouraging Intel, AMD and Nvidia to develop the technology over ten years ago, and they are now standard.
James Laros, a senior scientist at Sandia who oversees a program to test new computing architectures at Sandia, said the work with smaller players like NextSilicon is aimed at ensuring Sandia can always procure the chips it needs, even if major chip firms shift focus.
“We have to keep available options to complete our mission, because the mission is not optional,” Laros said.
