07,15,2026 7views
【Smart CityOnline Brand Column】Recently, Zhongke Shuguang, together with Longxun Kuangteng, Suzhou Laboratory, and Professor Fan Zheyong's team, relied on its self-developed super intelligent fusion computing power base: Shuguang 8000 (Ascending Peak) and MatPL machine learning force field software to achieve quantum mechanical accuracy in molecular dynamics simulations of "trillions of atoms" for the first time worldwide, breaking the world record for multi-scale simulations.
Compared with similar international technology solutions, the single atom single update time for this simulation is as low as 5.0 × 10 ⁻¹ ³ seconds, and the computing speed is 100 times faster than the world's top achievements, which is a thousand times faster than the previous Gordon Bell Award winning solutions; At the level of parallel expansion of computing power, from single card expansion to large-scale parallel clusters, the overall weak expansion efficiency is as high as 86.9%, and the communication ratio is only 4.26%. This means that we can almost linearly utilize any scale of computing power to challenge any massive atomic system.
Deep collaboration of software and hardware: Super intelligent fusion architecture unleashes ultimate simulation efficiency
This world-class achievement is derived from the self-developed super intelligent fusion underlying computing power base of Zhongke Shuguang: Shuguang 8000 (Peak) and the bidirectional deep tuning of domestic simulation software. It is also a typical implementation practice of Zhongke Shuguang's software and hardware collaboration solution for the AI for Science track.
Chip level deep adaptation optimization
For domestically produced acceleration chips, deep optimization of chip microarchitecture has been completed, with intermediate data residing in chip registers, significantly reducing performance loss caused by back and forth data transmission; Handing over AI inference tasks to tensor core for parallel computing, achieving nearly 10 times performance improvement on a single card; On the premise of not losing physical simulation accuracy, adopting a partially half precision (FP16) storage scheme directly reduces the memory bandwidth occupation by half, and can carry atomic systems of twice the size under the same hardware conditions.
Integrated base compatible with two types of core computing
Relying on the native super intelligent fusion design logic of Zhongke Shuguang, the same computing power base can simultaneously carry differentiated tasks such as high-precision first principles calculations, machine learning force field training and inference, without the need for cross system data migration. It completely solves the pain points of traditional scientific research platform process fragmentation and resource waste, and perfectly matches the requirements of new scientific research models such as NEP machine learning force field, which are "scientific computing+AI coupling".
Balancing Extreme Exploration and Inclusive Scientific Research: Lowering the Threshold for High Precision Simulation Use
This trillion atom simulation is a computational boundary limit exploration completed by Zhongke Shuguang based on its self-developed computing power base. Zhongke Shuguang, in collaboration with partners, has completed strong scalability testing for six major systems including metals, aqueous solutions, oxides, and two-dimensional materials. The atomic scale covers 1.8-2.8 billion, and the parallel efficiency remains above 80% during the expansion of computing resources, meeting the needs of cutting-edge scientific research.
At the same time, the architecture optimization of the entire base is also taking into account the routine research and development needs of various small and medium-sized scientific research teams, helping to achieve the universal implementation of scientific research computing power. Based on the computing power base of Zhongke Shuguang, small and medium-sized teams do not need to build their own super large clusters. With the integrated computing power services of Zhongke Shuguang, they can carry out high-precision simulations at the level of billions of atomic nanoseconds. Longxun Kuangteng synchronously opens up the full set of MatPL product manuals, open source code, and LAMMPS adaptation interfaces. Domestic research institutes and industrial R&D institutions can rely on the computing power base of Zhongke Shuguang to quickly deploy and implement large-scale material simulation projects.
Full stack domestic collaboration,Building a solid foundation for AI4S scientific research and innovation
In the era of AI for Science, the localization and collaboration of computing power bases and material simulation software is the core direction for seizing the global research and innovation highland. Relying on its self-developed super intelligent fusion computing power system, Zhongke Shuguang continues to connect the underlying hardware, scheduling platform, and industry software full stack adaptation links, and collaborates with ecological partners to continuously break through the upper limit of multi-scale material simulation scale.
In the future, Zhongke Shuguang will continue to collaborate with ecological partners in the fields of materials, energy, and advanced semiconductors, transforming trillion atomic level high-precision simulation computing power into innovative tools that can be easily accessed by industries and scientific research. With an independent integrated domestic computing power base, it will help achieve a leapfrog improvement in domestic basic scientific innovation capabilities.
