China's LineShine took the No. 1 spot on the TOP500 using nothing but CPUs

A Chinese system called LineShine took the top spot on the 67th TOP500 list, announced June 23 at the ISC conference in Hamburg. It posted 2.198 exaflops on the High Performance Linpack benchmark, more than 20 percent ahead of the U.S. machine El Capitan, which drops to second. It is the first system in the list's history to clear two sustained exaflops of double-precision math using CPUs only, with no GPUs or other accelerators doing the heavy lifting. That last detail is the whole story.

This is China's first number one since Sunway TaihuLight in 2017, so the symbolism is real. But a TOP500 ranking is a single number on a single benchmark, and it is worth knowing exactly what LineShine won before deciding what it means.

What the number measures

High Performance Linpack, usually just called Linpack or HPL, solves a large dense system of linear equations. It is the benchmark the TOP500 has used since 1993, which is why it is the one everyone quotes. It is also a friendly test. Dense linear algebra keeps the processors fed and busy, so HPL tends to report a machine near its theoretical ceiling.

LineShine's 2.198 exaflops is its Rmax, the measured score. Its Rpeak, the theoretical maximum, is 2.736 exaflops. So the machine hit about 80 percent of its paper peak on HPL, which is a healthy efficiency for a system this size.

The catch is that very little real science looks like Linpack. The TOP500 organizers know this, which is why a companion benchmark, HPCG, exists to model the sparse, memory-bound work that actual simulations do. Machines routinely score one or two percent of their HPL number on HPCG. LineShine topping the Linpack chart tells you it can run one specific kind of math very fast. It does not tell you it is the best machine for weather models, fusion codes, or training runs.

The engineering is the interesting part

Here is what makes LineShine worth a column rather than a headline. The fastest U.S. systems get their flops from accelerators. El Capitan pairs AMD EPYC CPUs with AMD Instinct MI300A chips. Frontier and Aurora do the same with AMD and Intel parts. Accelerators win because they deliver far more math per watt than a general-purpose CPU.

LineShine has none of that. It is built on a domestic platform called LingKun, using LX2 processors with 304 cores each running at 1.55 GHz, tied together by a proprietary interconnect called LingQi and running the Kylin operating system. To reach 2.198 exaflops without accelerators, it uses roughly 13.8 million CPU cores. El Capitan reaches its 1.809 exaflops with about 11.3 million cores plus its GPUs.

The reason for the all-CPU design is not hard to guess. U.S. export controls have cut China off from Nvidia, AMD, and Intel's top accelerators for years. LineShine is what happens when you cannot buy the efficient parts and decide to win on Linpack anyway: you throw an enormous number of domestic cores at the problem. As an answer to "can export controls keep China off the top of this list," the system is a fairly direct no.

The part the ranking does not show

Brute force has a bill. A CPU-only machine at this scale draws a large amount of power and floor space to do what an accelerator-dense system does in less, and the all-CPU approach is the least efficient route to a high Linpack score. The TOP500 publishes a separate efficiency ranking, the Green500, precisely because raw speed and performance-per-watt are different competitions. A first-place HPL result built on 13.8 million CPU cores is not the kind of design that tends to lead on efficiency.

None of that erases the result. LineShine is genuinely the fastest machine on the benchmark that defines the list, China built it without Western accelerators, and it is one of now five exascale systems in operation alongside El Capitan, Frontier, Aurora, and Germany's JUPITER Booster. Total computing power across the whole TOP500 climbed to 18.74 exaflops, up from 14.99 six months ago. The field is moving fast.

The honest summary is narrower than the headline. China has the fastest supercomputer in the world as Linpack measures it, built the hard way because the easy parts were off the table, and the cost of doing it that way is the figure nobody put on the slide in Hamburg.