On November 16, 2026, NextSilicon of the United States announced that it had delivered four of its Maverick-2 accelerators to Los Alamos National Laboratory's Darwin evaluation environment and brought them into operation. The laboratory will use scientific simulation codes related to national security to evaluate the new computing approach and its development tools.
Darwin is a group of small-scale systems where new hardware is run in practice and refined in cooperation with vendors. Maverick-2 works alongside CPUs and handles the most compute-intensive portions of an application. The laboratory will examine how well this approach suits multiphysics simulations, which combine multiple physical phenomena.
NextSilicon and Los Alamos National Laboratory have collaborated over multiple years across several generations of technology development. The announcement reports the installation of the hardware and the start of evaluation; no measured performance results have been published. Meanwhile, Sandia National Laboratories has already run existing scientific codes on an earlier NextSilicon system. On Darwin, researchers will analyze the results of running existing codes in detail and use them to optimize applications and to inform hardware co-design.
A new computing approach that changes hardware configuration at runtime
The scientific simulation codes that Los Alamos National Laboratory works with face long-standing challenges: irregular memory access and complex control flow. Irregular memory access refers to processing in which the required data is not laid out contiguously in memory and must be read from scattered locations. Complex control flow, on the other hand, means that the path of execution changes frequently because of conditional branches and similar constructs.
Scientific simulations must not only execute calculations quickly but also read the necessary data efficiently and cope with intricate processing flows. For that reason, when evaluating new computing devices, it matters not just how fast they perform raw arithmetic but how efficiently they can run real scientific codes.
NextSilicon positions Maverick-2 as an accelerator that works with CPUs to speed up the heavy portions of applications. The company's "Intelligent Compute Architecture" changes the hardware configuration while a program is running and allocates compute resources to match the processing needed at that moment.
According to the company, computation is assigned to a reconfigurable hardware execution platform. It also analyzes application behavior using data collected at runtime and accelerates the code sections that carry the heaviest load. The distinguishing feature of this approach is the combination of software learning algorithms with hardware whose configuration can be changed flexibly.
Supported languages and porting challenges
When moving scientific simulation code to a new computing environment, support for the programming languages and parallel-processing frameworks in use is important. The official NextSilicon Maverick-2 product page lists C/C++, Fortran, OpenMP and Kokkos as natively supported. CUDA and HIP/ROCm, by contrast, are distinguished as planned future integrations.
| Support status | Languages / frameworks | Description on the official page |
|---|---|---|
| Supported | C/C++, Fortran, OpenMP, Kokkos | Native support |
| Planned integration | CUDA, HIP/ROCm | Future integration planned |
Note: Based on the "Experience effortless porting" section of the official product page as of November 17, 2026. This table classifies support status and does not compare performance.
The page also lists integration with major AI frameworks as a future plan. The conditions for porting code differ between languages and frameworks that are already supported and those whose integration is only planned. For CUDA and HIP/ROCm, no specifics such as timing or the extent of compatibility are given.
NextSilicon also says that many applications can run on Maverick-2 without modification. However, the product page does not make it possible to confirm support for individual language features or libraries, or the operating conditions for each application. In actual porting, verification tailored to the combination of code and software in use will be necessary.
Once code is running, the next step is to analyze and optimize performance. Los Alamos researchers plan to use profiling and visualization tools provided by NextSilicon to identify where processing stalls, and to apply the resulting information to application tuning and future design. The aim of the evaluation is not only to see whether existing code runs, but to understand how compute resources are used and where there is room for improvement.
Sandia ran existing scientific codes
An earlier example of existing code running on a NextSilicon computing system comes from Sandia National Laboratories. In an article dated January 29, 2026, the laboratory described "Spectra," a prototype system installed in cooperation with NextSilicon. Spectra has 64 compute nodes and is the second system in the laboratory's "Vanguard" program.
According to Sandia's James Laros, Spectra was able to run the supercomputing benchmark "HPCG" on its first day of installation. It also ran "LAMMPS," a molecular dynamics simulator, and "SPARTA," a Monte Carlo code for rarefied gases. Being able to run three codes of different character is a significant result for evaluating support for existing software.
Based on these results, Sandia says it will evaluate in detail how far backward compatibility with existing code is ensured. The fact that particular codes ran is not enough to judge support for all the software used at the laboratory. The laboratory will expand the types and features of supported code and examine the range that can be applied to real research use.
Vanguard is an effort to evaluate whether new computing technologies can be applied to the National Nuclear Security Administration (NNSA)'s Advanced Simulation and Computing (ASC) program. Under this program, Sandia conducts joint research with Lawrence Livermore National Laboratory and Los Alamos National Laboratory. The targets include advanced hydrodynamics simulations that contribute to assessing the safety and reliability of the nuclear deterrent.
Vanguard's earlier systems include "Astra," an Arm-based supercomputer installed in 2018. Sandia positioned Astra as an experiment in applying the Arm architecture to compute-intensive modeling and simulation. The laboratory's article also recounts how, in 2024, Los Alamos National Laboratory installed "Venado," a large supercomputer that uses Arm in part. The pattern is to try new computing technologies in small-scale environments and then consider adopting them in future large-scale systems.
Darwin will evaluate fit with real scientific codes
The evaluation on Darwin goes beyond whether scientific codes can be run. The laboratory will examine how each part of an application is assigned to the compute platform and where processing time and load concentrate. Such analysis offers clues to which areas should be tuned to improve performance.
Galen Shipman, chief architect for Advanced Technology Systems at Los Alamos National Laboratory, explained in NextSilicon's announcement that the information obtained from tools that analyze code allocation will help with performance analysis and application co-design. Gary Grider, senior director for Computing Technologies at the laboratory, said that working directly with technology developers is intended to let the laboratory evaluate new computing architectures suited to its mission and to influence their design.
By bringing in the scientific codes it actually uses and analyzing their behavior and performance together with the developer, the laboratory can reflect software requirements in hardware improvements. Darwin will serve as an environment where researchers and developers carry out verification on real hardware.
The November 16 announcement did not disclose the names of the codes to be evaluated, the execution conditions, or the comparison targets. Speedup ratios for each code, measured power consumption, and the timing of completion of the evaluation have not been published either. If specific scientific codes, execution conditions and comparisons with existing systems are presented in the future, it will be possible to judge Maverick-2's applicability in more detail. How issues found during performance analysis are reflected in application optimization and future hardware design will also be a measure of what this deployment achieves.
