Known as Lingsheng, or LineShine, the system is designed to reach 2 exaflops, or two quintillion calculations per second, slightly above El Capitan’s 1.8 exaflops, according to the South China Morning Post.

Unlike most exascale supercomputers, which rely heavily on graphics processing units (GPUs), LineShine runs entirely on central processing units (CPUs). GPUs can handle thousands of similar calculations at the same time, making them useful for artificial intelligence and large-scale simulations. CPUs manage a wider range of tasks but usually need far more chips to reach the same speed.

The system uses 47,000 CPUs across 92 computing cabinets, according to chief designer Lu Yutong from the National Supercomputing Centre in Shenzhen, southern China.

A view of Frontier, an exascale-class machine at Oak Ridge National Laboratory, U.S. Photo courtesy of ORNL

At a conference in Shenzhen on April 24, deputy director Huang Xiaohui said LineShine had achieved full-stack independence, from hardware to core software, making it a fully domestic supercomputer.

"Its performance has already surpassed that of the U.S.’ El Capitan, returning China to the world’s No. 1 position," she said.

LineShine is expected to run on a new CPU called LX2, with each chip containing more than 300 computing cores. In its initial phase, the project used servers powered by Huawei’s Kunpeng processors, according to details released by the center.

Questions remain about how the system will perform in real-world use. The claimed 2-exaflop computing power is still a design target, and no timeline has been announced for when the full system will begin operating, according to Chinese media reports.

So far, testing has only been carried out on a much smaller version using around 100 Huawei Kunpeng servers with 12,800 processing cores, far below the scale planned for the final system.

It may also be difficult for outside experts to verify the system’s true performance. China has not submitted data to the global TOP500 supercomputer ranking since 2019, limiting independent comparisons with the world’s fastest machines.