China's LineShine Tops Supercomputer List, Built Without GPUs

For the first time since 2017, the world's fastest supercomputer is Chinese. A machine called LineShine, built at the National Supercomputing Centre in Shenzhen, entered the TOP500 ranking straight at No. 1 on Tuesday, knocking the United States off a list governments treat as a scoreboard for technological prestige.

The numbers are real and they are large. LineShine clocked 2.198 exaflops on the High Performance Linpack benchmark, more than two quintillion calculations a second, a roughly 20 percent lead over the previous champion, El Capitan, the Lawrence Livermore machine the U.S. uses to model its nuclear stockpile, which now sits second at 1.809 exaflops.

How it got there is the more interesting story. LineShine carries no GPUs at all. Where almost every leading system leans on graphics processors, China's machine runs entirely on conventional CPUs — Armv9-based LX2 chips wired together with a proprietary fabric called LingQi, with domestic storage underneath, according to analysis by Jon Peddie Research. It is the first computer to sustain more than 2 exaflops on CPUs alone, running at roughly 80 percent of its 2.736-exaflop theoretical peak. Shenzhen officials have framed the build as self-reliance “across the entire stack,” and after years of U.S. export controls on advanced chips, the all-domestic design is the headline Beijing wanted.

That claim needed checking. China had not submitted a Linpack result to the TOP500 since 2019, so the 2-exaflop figure had circulated for weeks as an assertion rather than a verified number until the 67th list landed at the ISC High Performance conference in Hamburg. An exascale machine is not an abstraction: this is the class of hardware used to model nuclear weapons, simulate climate systems and fold proteins, work that scales with raw double-precision throughput. On that measure, validated by an outside body, LineShine is now the leader.

Fast, but not built for AI

Here the boast needs an asterisk. Linpack rewards a particular kind of double-precision math that suits weather models and physics simulations, not the lower-precision work that trains modern AI. On a separate benchmark built to mimic AI-style computing, LineShine ranked only fourth, and technology and policy experts told Reuters the result says more about Beijing's appetite to prove self-sufficiency than about its standing in the AI race. The same week, President Trump signed an executive order aimed at keeping the U.S. ahead of China in quantum computing.

The five publicly verified exascale systems still split four-to-one in America's favor: El Capitan, Tennessee's Frontier and Illinois' Aurora trail LineShine, with Germany's Jupiter in fifth. That backdrop is the same rivalry now playing out in competing export controls on critical materials. A ranking is a snapshot, and this one was taken on a benchmark tuned to China's strengths. It is a genuine engineering feat. It is also a reminder that “fastest” depends entirely on which race you are timing.