Exaflop/s
CPU cores
all-domestic silicon
on the TOP500
For the first time since 2017, the world’s fastest supercomputer is Chinese — and it didn’t use a single GPU to get there.
A crown changes continents
On June 23, 2026, in a conference hall in Hamburg, the rankings that the high-performance computing world watches twice a year were rewritten. A machine almost nobody had heard of the week before — LineShine — appeared at the very top of the 67th TOP500 list, and in doing so ended a streak that had defined an era.
Installed at the National Supercomputing Centre in Shenzhen and built by the Shenzhen Cloud Computing Center, LineShine clocked 2.198 exaflops on the High Performance Linpack benchmark — roughly 2.2 quintillion calculations every second. That figure pushed past the United States’ reigning champion, El Capitan at Lawrence Livermore National Laboratory, which holds steady at 1.809 exaflops. It is the first time a China-based system has led the list since Sunway TaihuLight in 2017, and the symbolism was impossible to miss: the supercomputing crown had crossed an ocean.
Built from homegrown silicon
What makes LineShine remarkable isn’t only its speed — it’s how it got there. Every leading supercomputer of the past decade has leaned on graphics processors, the same accelerators that power modern AI. LineShine threw that playbook out. It is the first system on the TOP500 to break two exaflops of sustained double-precision performance using CPUs only, with no GPUs anywhere in the rack.
LineShine · Linpack sustained
The system is built on the custom Chinese “LingKun” platform: 13.79 million cores spread across 304-core LX2 processors running at 1.55 GHz, stitched together by a proprietary LingQi interconnect and running the domestic Kylin operating system. Its 2.198-exaflop result represents about 80 percent of a 2.736-exaflop theoretical peak — and it drew 42.2 megawatts to do it, for an efficiency of 52.07 gigaflops per watt.
That all-domestic design is the quiet story underneath the headline. China achieved this despite years of U.S. export controls aimed at cutting off access to the most advanced GPUs. By engineering around the embargo rather than through it — building a CPU-centric architecture on indigenous chips — Shenzhen’s engineers turned a constraint into a statement.
Five machines past the exascale line
LineShine’s debut did more than swap the name at the top. It lifted the number of systems sustaining more than one exaflop from four to five — and, for the first time, placed exascale machines on three continents at once: Asia, North America, and Europe.
The United States still owns three of the five exascale systems — Frontier at Oak Ridge and Aurora at Argonne join El Capitan — while Germany’s JUPITER Booster, operated under the EuroHPC Joint Undertaking, sits at exactly one exaflop as Europe’s sole representative above the line. The depth of the American fleet is undiminished. But the single number everyone remembers, the one at rank No. 1, now belongs to Shenzhen.
The figures that tell the tale
Raw Linpack speed is only one lens. LineShine also seized the top spot on the HPCG benchmark — a test built to reflect messy, data-intensive real-world workloads — with 22.00 HPCG-petaflops, ahead of El Capitan and Japan’s Fugaku. On the measures that reward sheer numerical muscle and irregular memory access, the new champion is convincing.
A win with a caveat
For all its dominance on classic benchmarks, LineShine carries a meaningful footnote. The same CPU-only design that let it dodge the GPU embargo also leaves it comparatively weak at the workload that matters most in 2026: artificial intelligence.
On the mixed-precision HPL-MxP benchmark — the test that best mirrors AI training — LineShine lands only fourth, at 7.92 exaflops, a modest 3.6× speedup over its standard score. El Capitan keeps that crown at 16.7 exaflops. The CPU architecture has no dedicated low-precision accelerators to lean on.
In other words, China now holds the title for the fastest traditional supercomputer — the kind used for climate modeling, physics simulation, and engineering — while the United States retains the lead in the precise area driving the global tech race. The crown is real, but it is not the AI crown. Both facts can be true at once, and both are.
How we got here
The pattern beneath the timeline is a contest that no single technology path now controls. The June 2026 top ten spans custom Chinese processors, AMD-powered exascale machines, Intel’s Aurora, NVIDIA’s Grace Hopper systems, cloud-based clusters from Microsoft, and Japan’s Arm-based Fugaku. There is no longer one road to leadership-class computing — and that diversity is exactly why the lead can change hands so suddenly.
What it really means
LineShine is, all at once, an engineering triumph, a geopolitical signal, and a reminder of how the supercomputing map is being redrawn. It proves that sanctions can reshape a strategy without stopping it, that a CPU-only machine can still touch the summit, and that the gap between the world’s computing superpowers is now measured in fractions of an exaflop rather than generations.
The next list arrives in November. The American fleet runs deep, new European and Italian systems are climbing, and the AI-optimized machines keep multiplying. For now, though, the fastest supercomputer on Earth answers to Shenzhen — and the race has never looked closer.
Leave a Reply