China has built a powerful new supercomputer, LineShine, that for the first time since 2017 pushed the United States out of first place on the Top500 list of the world’s fastest systems. The machine, running at the National Supercomputing Center in Shenzhen, delivers 2.198 exaFLOP of computing power, compared with 1.809 exaFLOP for the previous leader, El Capitan at Lawrence Livermore National Laboratory.
What makes the result notable is that LineShine was built mainly with central processing units, not graphics processing units. GPUs are the usual core of modern supercomputers, but their export to China has been tightly restricted by the United States. Instead of relying on GPUs, the Chinese system routes GPU-type workloads to dedicated circuits designed to speed up those calculations.
LineShine has already been used for sophisticated Earth simulations, including the atmosphere, oceans, continents, and ice, as well as complex simulations of the human brain. Jack Dongarra of the University of Tennessee, a Top500 contributor, said the architecture could point to a better way to combine AI and traditional scientific computing. “It’s an impressive system. They got around us by developing a system that is not so dependent on graphics processors,” he said.
The achievement also highlights the limits of the U.S. chip embargo, which has focused mainly on banning exports of high-performance GPUs, also known as AI chips. Washington has repeatedly raised the performance threshold for banned chips as Chinese firms adapted, but experts now think the success of LineShine may force the U.S. to broaden export controls. Jimmy Goodrich of the University of California’s Institute on Global Conflict and Cooperation told the New York Times that the U.S. should have stronger controls on the export and production of CPUs. Even so, he said China still lags in AI training performance because GPUs can process many simple calculations in parallel, something LineShine may not match despite its higher theoretical output.
