AMD has released scientific-computing chips, but there are concerns AI-optimized architectures and memory models may not meet Sandia’s future double-precision needs.

James Laros emphasizes keeping options open to ensure mission needs never risked, given the non-optional nature of nuclear security work.

Sandia is collaborating with Penguin Solutions to integrate NextSilicon’s chips into a supercomputer, with a milestone indicating readiness to test on demanding nuclear-security simulations later this year.

The program will evaluate this fall whether NextSilicon’s chips pass more demanding tests resembling real nuclear-security workloads.

Double-precision floating-point performance remains a critical concern, as it’s essential for certain physics simulations and may not be optimally served by AI-optimized chips.

Sandia National Laboratories at Kirtland Air Force Base relies on liquid-cooled high-performance supercomputers to simulate nuclear weapons physics, but faces supply challenges as AI-focused chipmakers shift priorities.

Nvidia remains involved with Sandia for scientific computing, indicating a mixed ecosystem rather than a full replacement of traditional suppliers.

Officials caution about double-precision performance and supply reliability, underscoring the need for multiple procurement options to ensure mission readiness.

NextSilicon’s approach features double-precision capable chips, a data-flow architecture to reduce memory shuffling, and on-the-fly reprogrammability to improve efficiency.

NextSilicon’s data-flow architecture reduces memory movement and energy use, with self-reprogrammability to boost efficiency, aiming to balance AI and scientific workloads.

A key technical milestone was reached, moving NextSilicon’s chips closer to potential use in government systems for nuclear-security work.

There is ongoing concern about double-precision performance, a strength of Nvidia and AMD in scientific computing, for future workloads.