America’s Supercomputers Are Starting to Bet on AI Chip Newcomers

The supercomputers at Sandia National Laboratories run simulations of nuclear warhead behavior and hypersonic weapons trajectories. For more than a decade, that work ran on chips from NvidiaNVDA

and Advanced Micro DevicesAMD

. Now, as both companies redirect their hardware roadmaps toward artificial intelligence, Sandia is evaluating chips from an Israeli startup for that mission-critical work instead.

Why mainstream chips are falling short

The core issue is a technical specification called double-precision floating point computation. It's the ability to handle both very large and very small numbers in the same calculation without losing accuracy to rounding errors.

For years, Nvidia and AMD competed to lead in this capability, landing supercomputing contracts with universities and government labs. AI workloads don't require the same level of double-precision performance, and the industry has shifted accordingly.

Per Reuters, the double-precision performance of Nvidia's forthcoming Rubin chips has declined by some measures. Ian Cutress, chief analyst at chip consulting firm More Than Moore, said the decline has worried many scientists across the high-performance computing industry.

AMD is releasing a version of its chips aimed at scientific computing, but supply chain pressures are compounding the problem for labs like Sandia.

Steve Monk, manager of Sandia's high-performance computing team, put the situation directly. "The pressure we're feeling right now is on the computing front and also from the supply chain," he said.

How NextSilicon's chips work differently

NextSilicon uses an architecture that's fundamentally different from the graphics processing units and central processing units that dominate the market. The chips are designed to reprogram themselves in real time to run more efficiently on whatever task they're processing.

The design uses what's known as a data flow architecture. It reduces the time and energy spent moving data back and forth between the processor and system memory, according to Reuters.

NextSilicon's chips also support double-precision computation, the capability that mainstream AI-focused chips have begun to deprioritize. That combination of energy efficiency and scientific precision is what put the company on Sandia's radar.

The high-stakes trial that comes next

This week, Sandia, NextSilicon, and Penguin Solutions, the integrator that built NextSilicon's chips into a working supercomputer, announced the systems passed a key technical milestone. The chips cleared a battery of general supercomputing benchmarks, putting them in contention for government use.

That result sets up a decision this fall on whether to advance to more demanding tests, ones that would closely resemble the nuclear security calculations the chips would eventually handle.

James Laros, a senior scientist at Sandia who oversees the program evaluating new computing architectures, described the rationale for testing smaller suppliers in straightforward terms. "We have to keep available options to complete our mission, because the mission is not optional," he said.

Sandia's track record in shaping chip technology

Sandia's decisions carry weight beyond its own systems. The lab began pushing Intel, AMD, and Nvidia to develop liquid cooling for chips more than a decade ago, when the technology was considered experimental. Liquid cooling is now standard across the industry.

That history gives Sandia's engagement with NextSilicon more significance than a routine procurement test. If the chips advance through Sandia's evaluation process, the institutional endorsement could accelerate adoption beyond government labs.

Nvidia has maintained that it isn't walking away from scientific computing. Daniel Ernst, the company's senior director of supercomputing products, said Nvidia is working toward a chip capable of handling both AI and real-world scientific applications.

The market dynamic, however, is already shifting. As the two dominant chip companies orient their development cycles toward AI, niche markets with demanding scientific requirements are opening up to competitors that couldn't gain entry before.