China unveiled a comprehensive software platform in Beijing on Monday aimed at overcoming a long-standing bottleneck in its drive for self-reliant high-performance computing, promising to make it much easier for scientists to run complex research programs on domestically developed chips.
The platform, called Yisuanfangzhou, was jointly developed by the Computer Network Information Center of the Chinese Academy of Sciences, the University of Science and Technology of China, the CAS Institute of Mechanics and supercomputing company Sugon.
"Open-source software is the public infrastructure of the digital world," said Wang Yangang, a researcher at the Computer Network Information Center who led the platform's development.
"From operating systems like Linux and Android to the tools used in scientific research, almost everything we rely on is built on open source. It lowers costs, speeds up innovation, and empowers breakthroughs in cutting-edge technologies," Wang said.
However, much of today's scientific software remains deeply dependent on NVIDIA's CUDA ecosystem, which has grown over nearly two decades into a vast code base maintained by millions of developers.
"That ecosystem has become a deep moat that is very hard to cross. If you want to move workloads to domestic hardware, you have to deal with all the code those millions of developers have produced over the years," Wang said.
He said that although China has made rapid progress in building world-class supercomputers using its own CPUs and GPUs, software remains the weak link.
"If you don't have software that can run on the hardware, it's like trying to cook a meal with no rice in the pantry," he said. "Even if your machine ranks first in the world, you won't be able to unlock its true performance."
To bridge that gap, Yisuanfangzhou provides an integrated toolkit designed to migrate scientific software to domestic chips with minimal manual modification. It offers end-to-end support for algorithm adaptation, code translation, and application deployment.
Wang described the platform as being structured like a building: algorithms provide the computing power that forms the foundation, code translation builds the framework by ensuring compatibility, and intelligent agents complete the structure by turning complex workflows into simple, repeatable tasks.
At its foundation is Jiuyanshu, an algorithm library that packages 16 high-performance solvers covering a range of widely used scientific computing fields, including linear algebra, parallel computing, fluid dynamics, biological computing and deep learning. Each solver is optimized for domestic computing architectures, with key modules delivering more than 10 times the performance of generic implementations.
Another core component, BoundX is an AI-driven code translation engine designed to address one of the biggest challenges in migrating legacy CUDA code.
"Code that used to take an engineer about 10 hours to migrate line by line can now be processed by the platform in roughly 30 minutes, with the converted program submitted directly to a domestic supercomputing environment," Wang said.
According to the development team, it has an automated conversion success rate of 71 percent, outperforming mainstream domestic tools by more than 20 percentage points.
A third component, Agent-HiReFlow, is an intelligent simulation agent for engineering applications such as aerospace fluid dynamics, allowing users to describe simulation tasks in natural language.
"The agent translates human language into physical models, and then into software instructions," said Cheng Wan, a professor at the University of Science and Technology of China who led development of the component.
"It configures parameters, launches solvers, monitors faults, and delivers visualization-ready results, without users writing any code themselves," Cheng said.
In one test case, the system completed a 10-million-cell hypersonic flow simulation in about an hour, matching performance previously achievable only on top-end NVIDIA hardware.
"Our goal is to move from hardware leadership to software maturity, giving researchers an environment where development is convenient and applications are easy to use," Wang said.
He said the platform will be integrated with national computing infrastructure, including the Orient supercomputing system and the National Supercomputing Internet, allowing approved researchers to access domestic chips without rebuilding their software stacks from scratch.
The development team plans to support more domestic chip architectures, deepen collaboration with open-source communities and expand into fields such as digital twins and AI for Science.
Nie Hua, chairman of Sugon's subsidiary Zhongke Controllable Information Industry Co, said the launch highlights that building a software ecosystem is even more challenging than achieving breakthroughs in hardware.
"China's computing industry has largely solved the problem of 'having nothing', but now faces a new industrial security challenge — how to integrate chips into real-world applications," Nie said.
"Software-hardware co-design is the next big hurdle. If we succeed, we will have a fully self-reliant, world-class computing platform that serves both scientific discovery and national needs," he added.
京公网安备 11010202009201号