China Reclaims the Supercomputer Crown: All-Domestic 'LineShine' Tops the TOP500

China Reclaims the Supercomputer Crown: All-Domestic 'LineShine' Tops the TOP500

China is back on top of supercomputing. In the June 2026 TOP500 list, a system named LineShine at the National Supercomputing Center in Shenzhen clocked 2.198 Exaflops on the High Performance Linpack benchmark — the first machine ever to break 2 EFlops — and reclaimed the number-one spot for the first time since 2017. The twist: it’s built on a fully domestic stack — Chinese CPUs, interconnect, and operating system — and notably uses no GPUs.

Fast-moving story. Figures below reflect the June 2026 TOP500 edition and official system disclosures. Rankings and specs change between editions — treat them as point-in-time.

At a glance

	Detail
System	LineShine (a.k.a. Ling Chen / Ling Shen)
Location	National Supercomputing Center, Shenzhen
HPL performance	2.198 EFlops — first system past 2 EFlops
CPU	LineShine-LX2, ARMv9, 304 cores each, 1.55 GHz
Scale	45,360 CPUs = ~13.79 million cores
Interconnect / OS	LinqQi fabric · Kylin OS (Linux)
Power	42.22 MW at 52.1 GFlops/W
Beats	US El Capitan (1.809 EFlops, AMD MI300A)

What LineShine actually is

LineShine isn’t just fast — it’s all-Chinese, end to end. The compute comes from LineShine-LX2 processors built on the ARMv9 architecture, each packing 304 cores at 1.55 GHz with ARM SVE and Scalable Matrix Extensions (SME). Across 45,360 processors, that’s roughly 13.79 million cores.

Tying it together is a homegrown LinqQi interconnect (≈1.6 Tbit/s per CPU) and the domestic Kylin OS, a Linux derivative. It draws 42.22 megawatts and lands at 52.1 GFlops per watt on efficiency.

The headline number: 2.198 Exaflops on HPL — over 2 quintillion floating-point operations per second, and the first time any system has crossed the 2-EFlops line.

How it beat El Capitan

The previous champion, the US El Capitan at Lawrence Livermore (which the US uses for its nuclear-stockpile work), posts 1.809 EFlops on HPL using AMD MI300A accelerators. LineShine edges it by nearly 400 PFlops on the benchmark — even though El Capitan has a higher theoretical peak (~2.821 EFlops). In other words, China’s machine converts more of its raw potential into measured Linpack performance here.

It’s also a comeback: the last Chinese system to hold #1 was Sunway TaihuLight in 2017. China then stopped submitting systems to the TOP500 in 2023 amid US chip-export controls — so this isn’t just a ranking, it’s a statement that the country can build a world-leading machine without Western silicon.

The catch: it’s not built for AI

Topping the TOP500 measures 64-bit (FP64) Linpack throughput — classic scientific-computing horsepower. Modern AI training lives on lower-precision math (FP16/FP8), where GPUs shine.

That’s where LineShine’s CPU-only design shows its limits: on the HPL-MxP mixed-precision benchmark it ranks only 4th at 7.92 EFlops, because CPU-only systems get far less speedup from reduced precision than GPU-accelerated platforms. So while LineShine is the world’s fastest at traditional HPC, it likely won’t lead the AI race, where US GPU clusters still dominate.

Why it matters

1. Self-sufficiency is the real story. A #1 supercomputer with a 100% domestic CPU, interconnect, and OS is a direct answer to export controls — proof that China’s chip ecosystem can produce frontier-class hardware on its own terms.

2. HPC leadership ≠ AI leadership. The result is a useful reminder that “world’s fastest” depends on the benchmark. For nuclear simulation, climate, materials, and physics, FP64 is king — but the AI frontier is a different contest, still tilted toward GPUs.

3. The benchmark game is geopolitical. China returning to the list — loudly, with domestic parts — reframes the TOP500 as much a statement of technological independence as a performance chart.

Bottom line

LineShine’s 2.198 EFlops makes China the fastest in the world on the TOP500 for the first time since 2017 — and it did it with an entirely domestic stack, no GPUs required. It’s a genuine engineering and geopolitical milestone. But the CPU-only architecture that wins at FP64 Linpack is not optimized for AI, where the US still holds the edge. Two different races — and right now, each side leads one.

---

Sources: June 2026 TOP500 list and system disclosures, reporting from heise online, Reuters, Business Standard, SCMP, and technology.org. Rankings and specifications are point-in-time and may change in future TOP500 editions.