Key Points:
- Sandia National Laboratories tests new microchips from Israeli startup NextSilicon to run complex nuclear weapons simulations.
- Massive semiconductor companies like Nvidia and Advanced Micro Devices now focus their designs almost entirely on artificial intelligence.
- The sudden shift toward artificial intelligence leaves government labs struggling to find chips capable of high-precision math.
- NextSilicon chips passed initial general supercomputing tests and will face more demanding nuclear security trials this fall.
In a secure building located on the Kirtland Air Force Base in New Mexico, massive liquid-cooled supercomputers run constantly. These machines crunch some of the most complex math problems the United States government needs to solve. Scientists use these supercomputers to simulate exactly how a hypersonic nuclear weapon would move through the atmosphere. They also run highly classified models to see what happens when one nuclear warhead detonates near another.
For more than a decade, the government relied on mainstream semiconductor companies to supply the microchips for this secretive work. Tech giants like Nvidia and Advanced Micro Devices built their entire reputations by providing hardware for these massive government supercomputers. However, the technology market recently experienced a massive shift that threatens the future of these defense projects.
Major semiconductor firms now design their latest microchips almost exclusively for artificial intelligence. This massive shift in focus created severe global supply shortages for specialized hardware. Because of this artificial intelligence boom, the managers who run the systems at Sandia National Laboratories find themselves in a very difficult position. Sandia operates the machines at Kirtland and serves as one of three national labs tasked with developing and maintaining the American nuclear weapons arsenal.
Steve Monk works as the manager of the high-performance computing team at Sandia. He recently explained the massive challenge his team faces in securing the exact chips they need. Monk stated that his team feels intense pressure right now on both the computing front and the overall supply chain. He admitted that looking to the future feels quite stressful because the lab must ensure it can deliver on its national security mission.
This frustrating predicament at Sandia highlights an unintended consequence of the massive corporate race to build better artificial intelligence chips. As the big firms abandon traditional scientific computing to chase profits in artificial intelligence, they open the door for smaller players to enter the market.
NextSilicon, a young startup based in Israel, jumped at this new opportunity. The company builds specialized chips that perform complex math without focusing on artificial intelligence. Right now, a dedicated testing program at Sandia is running NextSilicon chips through a grueling series of tests. Sandia holds a long history of incubating and shaping new computing technologies. In the past, the lab worked extensively with Nvidia to help the company rise to prominence in supercomputing.
The biggest concern for officials at Sandia is a highly specific technical requirement called double-precision floating-point computation. This complex term simply describes the ability of a computer chip to calculate both very large and very small numbers simultaneously without incurring rounding errors. For years, Nvidia and Advanced Micro Devices fought hard to speed up this specific kind of computing so they could win lucrative contracts with universities and government defense labs.
However, modern artificial intelligence work does not need double-precision computing to function properly. Artificial intelligence models rely on a different type of math entirely. Because of this difference, the double-precision performance of Nvidia’s upcoming Rubin chips actually declined in some key measures. Ian Cutress, the chief analyst at the chip consulting firm More Than Moore, said this specific performance drop deeply worries scientists across the high-performance computing industry.
Nvidia defends its current business strategy. Daniel Ernst serves as the senior director of supercomputing products at Nvidia. He insisted that his company remains fully committed to traditional scientific computing. He explained that Nvidia aims to create a perfectly balanced chip that can run real-world scientific applications directly alongside modern artificial intelligence tasks.
Despite these corporate promises, the shifting market forced officials at Sandia to look elsewhere. The lab recently began testing NextSilicon products to see whether the startup could fill the performance gap. NextSilicon engineers built their chips using a completely different computing approach than the standard graphics processing units or central processing units sold by Nvidia and Advanced Micro Devices.
On Monday, Sandia, NextSilicon, and Penguin Solutions released a joint statement regarding their progress. Penguin Solutions helped weave the new NextSilicon chips directly into the government supercomputer network. The companies announced that the new systems officially passed a critical technical milestone. The chips survived a massive battery of general supercomputing tests, proving they have the raw power needed for use in strict government systems.
Passing this initial test sets up NextSilicon for a massive decision later this fall. The government will decide whether to start testing these chips on much more demanding computing problems. These upcoming tests closely resemble the highly classified nuclear-security math that the chips would eventually have to handle on a daily basis.
The NextSilicon chips stand out for easily handling double-precision computing. Engineers also designed the hardware to reprogram itself on the fly so it runs more efficiently during complex tasks. The chips save massive amounts of electricity by using a data-flow architecture. This specific design spends far less time and energy shuttling data back and forth between the computing system’s memory banks.
The work happening at Sandia often helps strange new technology become widespread across the commercial sector. For example, liquid cooling systems for microchips seemed like an exotic and crazy idea more than a decade ago. Sandia strongly urged companies like Intel, Advanced Micro Devices, and Nvidia to develop the technology anyway. Today, liquid cooling stands as a common feature in most major data centers.
James Laros works as a senior scientist at Sandia and oversees the program that tests new computing architectures. Laros explained that working with smaller players like NextSilicon guarantees that Sandia can always procure the hardware it desperately needs. He noted that the lab must keep its options open to complete its mission because national defense is never optional.
