The First Joint Solution of Sugon DCU and Scientific Large Models Debuts: Propelling World-Class Application Innovation to the Forefront

The 10th Scientific Data Conference, themed “Scientific Data and Sustainable Development,” recently concluded, showcasing a range of technological and research innovations powered by Hygon DCUs (Deep Computing Units). The event highlighted collaborative efforts with leading institutions like the Institute of Automation of the Chinese Academy of Sciences (CAS), the Institute of High Energy Physics (IHEP), and the National Astronomical Observatories of China (NAOC), demonstrating the burgeoning ecosystem driving intelligent advancements across numerous scientific disciplines.

At the conference, Hygon Information and the IHEP jointly unveiled a collaborative solution centered around scientific large language models built on Hygon’s DCU architecture. This partnership creates a potent synergy, leveraging Hygon’s domestically produced computing power alongside IHEP’s proprietary models and extensive scientific data sets, aiming to accelerate AI for Science research initiatives.

Multimodal Models Usher in a New Era for Automated Research

As large language models (LLMs) reshape research paradigms, automated research is venturing into uncharted territory, integrating cross-modal and cross-domain approaches. Dr. Chen Yingying, Associate Researcher at the Institute of Automation of CAS, presented the “Zidong Taichu” large model’s research applications, illustrating how multimodal AI empowers researchers by facilitating interdisciplinary data fusion, knowledge association, and automated reasoning.

Focusing on the systematic development of domestic toolchains, the Institute of Automation of CAS and Hygon Information have jointly developed seven comprehensive high-performance toolchain solutions, including Hygon DCU driver components and runtime components. These solutions offer integrated intelligent support for complex tasks like neural network model training, image-text model training, and LLM training and inference.

AI Accelerates High Energy Physics Towards Human-Machine Interaction

Highlighting the increasing reliance on big data and AI in high energy physics research, Dr. Zhang Zhengde, a distinguished young researcher at IHEP, emphasized that traditional analysis methods are no longer sufficient to handle the massive datasets generated by large-scale scientific instruments like electron-positron colliders, hindering the efficiency and depth of theoretical investigations.

Addressing challenges like limited training resources, complex model ecosystems, and stringent regulatory compliance requirements, IHEP ultimately selected Hygon DCUs as its underlying computing platform. This led to the creation of “Xi Wu,” the world’s first L2-level high energy physics large model. This model anchors the “Sai Doctor” high energy physics scientific intelligence system, demonstrating world-leading technical specifications.

Hygon Information’s Head of Intelligent Computing R&D elaborated on the technical strategies and organizational mechanisms behind the collaborative effort. He highlighted that IHEP’s choice of Hygon DCUs as its computing foundation reflects a high level of trust in domestic, secure, and reliable technological capabilities. Hygon Information is also providing comprehensive support to the “Sai Doctor” project, spanning computing deployment, algorithm adaptation, and model training, establishing a repeatable “methodology” for helping clients cultivate world-class research achievements.

Synergistic Optimization of Scientific Research Software

In the field of astronomy, data is exploding exponentially as instrument sensitivity increases and observation methods expand. Dr. Zhang Yanxia, Researcher at the National Astronomical Observatories of China, mentioned that large-scale scientific projects like China’s FAST telescope generate 96 PB of data annually, with projections exceeding 1000 EB in the future, placing unprecedented strain on storage, analysis, and computation resources.

The Head of Intelligent Computing Products at Hygon Information pointed out that Hygon DCUs are compatible with mainstream global AI architectures. The supporting development toolkit DTK, the AI basic software stack DAS, and the AI application platform DAP form a synergistic “troika” for scientific research software, driving computing power leaps in various scenarios, from celestial object identification to weather simulation and cryo-electron microscopy reconstruction.

For example, in carbon star identification, the computational efficiency of Hygon DCUs is more than 107 times that of a single CPU core. In terms of parallel efficiency of cryo-electron microscopy algorithms, Hygon DCUs have achieved a high standard of 91.7%.

Going beyond straightforward collaboration to encompass technological advancement, research output, and industrial application, the value generated by Hygon Information and its numerous research clients is evident. Hygon Information is actively collaborating with leading scientific research organizations to explore a preferred domestic computing technology pathway through synergistic hardware and software integration. This approach is gradually establishing a uniquely Chinese technological innovation incubator system positioning itself for a head start in global AI development.