Milestone in domestic AI computing power: Deployment of 3 scaleX superclusters with tens of thousands of cards simultaneously implemented

On February 5th, domestic AI computing power reached a milestone as the national supercomputing internet core node was launched for trial operation in Zhengzhou. The node’s computing resources are supported by three sets of Sugon scaleX ten-thousand-card supercluster systems, capable of providing over 30,000 cards of domestically produced AI computing power, making it the largest single domestic AI computing resource pool connected to the national supercomputing internet platform since its launch.

In the first year of the 14th Five-Year Plan, the strategy to build a nationwide integrated computing power system has accelerated again, and China’s infrastructure for computing power has entered a new stage of “building and using in tandem, promoting construction through use.” After the supercomputing internet core node goes online, it will, on one hand, create a highland of computing power in the central region, fully leveraging the clustering effects of talent, data, and application scenarios; on the other hand, it will become the largest computing power scheduling hub in the country, effectively linking China’s and even global computing resources and application demands.

Cao Zhenan, Deputy Director of the National High-Performance Computing Center, explained in detail, “The supercomputing internet is becoming a computing power application marketplace similar to JD.com or Taobao, aiming to provide efficient and inclusive computing services for various cutting-edge application scenarios, making computing power as freely flowing, efficient, and easy to use as water and electricity.”

Data shows that by the end of 2025, the national supercomputing internet platform will have served over one million users, with more than 7,200 application products, a peak of 1.03 million jobs processed in a single day, and a total of 196 million jobs supported and operated so far. This integrated “computing power + application” service model significantly reduces costs and usage barriers, effectively empowering AI innovation acceleration.

Among them, the scaleX ten-thousand-card supercluster system from Sugon provides a critical computing foundation for this core node.

It is introduced that the scaleX supercluster, based on an open architecture design for AI computing, fully supports mainstream software ecosystems like CUDA, supports mixed deployment of multiple brands of domestically produced acceleration cards, not only reducing the difficulty for developers to migrate and adapt but also possessing flexible scalability to hundreds of thousands or even millions of cards, better meeting the needs of industrial-scale development.

Currently, the scaleX supercluster has completed adaptation and optimization for over 400 mainstream large models and world models, and through the national supercomputing internet, it can connect to over a thousand applications, linking more AI industry ecosystem partners, providing efficient and inclusive computing power support for large-scale AI scenarios such as trillion-parameter model training, high-throughput inference, and AI for Science.

According to industry reports, the scaleX supercluster made its first appearance on real machines at the HAIC conference in December last year, and within two months, it was deployed with a computing power scale of over 30,000 cards. Some experts believe this marks the official move of domestically produced multi-card clusters into a new stage of large-scale deployment and practical application. As the “China Computing Network” is fully rolled out, domestic clusters are expected to open up broader market opportunities.

(Source: Caixin)