China's Supercomputer Returns to World First: What It Means, Victory of Full-Stack Independence
On June 23, the International Supercomputing Conference in Hamburg, Germany, released the global TOP500 supercomputer list. The domestically produced supercomputer "Lingsheng" in the Guangming Science City of Shenzhen, Guangdong, successfully topped the global list. This is the first time in nine years that China's supercomputer has ranked first in the world.
As a major scientific and technological infrastructure in Guangming Science City, the completion and operation of "Lingsheng" will provide strong computing power support for major scientific and technological breakthroughs and cutting-edge interdisciplinary research in the Guangdong-Hong Kong-Macao Greater Bay Area. "Lingsheng" is the world's first supercomputing system to continuously exceed 2 EFLOPS (exaflops, or 10^18 floating-point operations per second), with a measured sustained performance of 2.198 EFLOPS. This computing power is equivalent to 7 billion people performing one operation per second continuously for nearly 10 years.
Looking back at the development of China's supercomputing, from "Tianhe-2" to "Sunway TaihuLight," Chinese supercomputers have topped the list for many years. Facing challenges from changes in the external environment, China's supercomputing has been continuously refined and iterated in scientific research, engineering, industry, and artificial intelligence scenarios. Through long-term investment and persistent efforts, it has reconstructed a technological development path that is autonomous, controllable, safe, reliable, and sustainable.
Supercomputing is a highly complex system engineering, from chips, networks, storage, system software to application ecosystems. Any weakness in one link affects the overall capability. The key to "Lingsheng" topping the list is systematic innovation. It uses full-stack domestic software and hardware, building autonomous and controllable systematic capabilities in key chips, high-speed interconnection, system software, green energy saving, and application coordination. This full-stack independence not only ensures safety and reliability but also gives us the ability to master the initiative in supercomputing technology iteration and the right to define technical standards.
A major difficulty in the R&D process was the choice of technical route and the implementation of a large-scale system. After multiple rounds of research and repeated demonstrations, the team innovatively adopted an all-CPU hyper-intelligent fusion architecture. This is not a passive alternative but an active choice after in-depth study. Although the traditional CPU+GPU heterogeneous route is powerful, it has problems such as high data migration overhead, complex programming, and high application migration costs. "Lingsheng" introduces matrix acceleration computing units into the CPU, giving the CPU its own acceleration capability, and together with domestic high-bandwidth memory, self-developed high-speed interconnection networks, and full-stack software, it achieves a deep integration of supercomputing and intelligent computing at the architectural level.
