Inside a functional building on Kirtland Air Force Base in New Mexico, liquid-cooled supercomputers quietly work through computational problems the U.S. government regards as highly sensitive and technically demanding. The machines at Sandia National Laboratories are used to model extreme physics scenarios - for example, how hypersonic nuclear devices travel through the atmosphere and the effects if one warhead were to detonate near another.
For many years the chips that powered those systems came from large semiconductor vendors such as Nvidia and Advanced Micro Devices. But as those firms increasingly tailor designs for artificial intelligence workloads and face supply tightness, Sandias high-performance computing managers are confronting uncertainty about where to procure processors that meet the precision and delivery requirements for scientific and national security applications.
"The pressure were feeling right now is on the computing front and also from the supply chain," said Steve Monk, who manages Sandias high-performance computing team, describing the challenge of obtaining chips that satisfy the labs technical needs. "Looking to the future, its a bit stressful in terms of our ability to deliver to the mission."
That strain on traditional supply channels has opened the door to smaller, less familiar entrants in the chip market. Sandia has begun testing hardware from NextSilicon, an Israeli startup whose processor architecture departs from conventional graphics processing units and central processing units. The labs work with outside firms has historically helped shape which technologies gain traction in high-performance computing contexts; Sandia collaborated closely with Nvidia as that company rose to prominence in supercomputing and continues to work with it on new memory technologies.
A core technical requirement for Sandias workloads is reliable double-precision floating point computation - the ability to compute both very large and very small numbers without losing accuracy to rounding errors. That capability is central to physics simulations used in national security work. Historically, Nvidia and AMD chased performance leadership in double-precision computing and won contracts at universities and government labs because of it.
But artificial intelligence training and inference typically do not demand double-precision capability to the same extent as complex scientific simulations. While AMD is releasing a variant of its chips targeted at scientific computing, some measures indicate a decline in double-precision performance for Nvidias forthcoming Rubin chips. Ian Cutress, chief analyst at chip consulting firm More Than Moore, said that decline has alarmed many scientists working in high-performance computing.
Nvidias senior director of supercomputing products, Daniel Ernst, said the company remains committed to scientific computing and aims to design balanced chips capable of running real-world scientific applications alongside AI workloads. Yet the market shift toward AI-optimized designs and the resulting supply dynamics have prompted Sandia to evaluate alternatives.
NextSilicons processor passed a key technical milestone this week in testing carried out in partnership with Sandia and Penguin Solutions, the firm that integrated the newcomers chips into a supercomputer. The systems successfully completed a suite of general supercomputing tests, qualifying NextSilicons chips for further consideration in government computing systems.
The milestone sets up a decision this fall on whether NextSilicons processors will progress to more demanding testing that more closely resembles the nuclear security workloads they would ultimately be expected to run. According to the companies and the lab, NextSilicons design supports double-precision computing and includes the capacity to reprogram itself dynamically to achieve greater efficiency.
One of the stated advantages of the NextSilicon architecture is energy efficiency. The chips use a data flow approach that reduces the time and energy spent moving data between the processor and memory, which helps lower electricity consumption for heavy computations.
Sandias role in evaluating new architectures is consistent with past work that helped normalize technologies once considered exotic. Liquid cooling, for example, was an uncommon approach when Sandia encouraged companies such as Intel, AMD and Nvidia to pursue it more than a decade ago; that technique has since become widely adopted in high-performance computing environments.
James Laros, a senior scientist at Sandia who oversees a program to test new computing architectures, said the collaboration with smaller chipmakers aims to preserve the labs ability to procure suitable computing hardware even if major vendors change focus. "We have to keep available options to complete our mission, because the mission is not optional," Laros said.
The current environment illustrates how the broader industrys pivot toward AI has altered priorities in the semiconductor market and created procurement and technical challenges for institutions that depend on high-precision computation. Sandias testing of alternative processors such as those from NextSilicon, and its history of nudging mainstream vendors toward technologies like liquid cooling, reflect an effort to maintain capability and choice in a shifting landscape.