China overtakes USA with fastest Top500 supercomputer: 2.2 Exaflops | heise online

heise+ entdecken

SuchenAbo

Suchen

Alle Magazine im Browser lesen<br>AnzeigeSpecial: Collaboration im KI-Zeitalter

Newsletter<br>heise-Bot<br>Push-Nachrichten

${lead}

${lead}

${content}

${content}

${content}

${content}

Anzeige<br>Special: Collaboration im KI-Zeitalter

Advertisement

Advertisement

Continue after ad

The 67th list of the fastest supercomputers, also known as Top500, has it all once again. Now, a Chinese native system is the first supercomputer to break the prestigious mark of 2 trillion calculations per second (2 EFlops) in the High Performance Linpack (HPL) benchmark. This measures the computing power in solving huge linear equation systems with double-precision floating-point accuracy (FP64). While export restrictions on US hardware have posed some challenges for China, they have ultimately not hindered it permanently.

A year ago, the Jupiter Booster from Jülich was one of the highlights as the fastest European system and later as the first European Exascale system. Exascale systems are supercomputers that can achieve at least one trillion (1018) calculations per second in HPL.

Also new in the Top 10 of the list is a second commercial system from the Italian oil and energy group Eni in sixth place. The Finnish LUMI and the Italian Leonardo have fallen out of the Top 10 list.

In total, 44 new systems with an average of 91 PFlops and a total of 3994 PFlops made it into the 67th Top500 list. The new leader from China thus contributes more than half of the new computing power. If you want to try your luck yourself, for entry at rank 500, your PC needs more than 2661.56 TFlops.

Much Ado About Nothing

After China had become one of the leading countries on the Top500 list, it stopped publishing figures for its supercomputers from 2019 onwards, partly due to US export restrictions. Before that, the Sunway TaihuLight in 2017 was the last Chinese system at the Top500 top spot. However, it was an open secret that several Exascale systems were already running in China.

Continue after ad

Configuration of the LX2 processor with ARMv9 cores of the Chinese supercomputer Lineshine.

(Image: NSCCSZ)

The new system is called LineShine and was already presented in April 2026. However, concrete performance data for the machine, also known as Ling Chen or Ling Shen (灵晟), was missing.

Lineshine relies entirely on Chinese hardware developments. The platform is called LingKun and does not use accelerators, which eight of the other Top10 systems use. The computing power comes from Lineshine-LX2 processors with ARMv9 architecture, each with 304 cores running at 1.55 Gigahertz. They are strongly based on the Fujitsu A64FX of the Japanese supercomputer Fugaku, with an integrated interconnect called LinqQi and bundled High Bandwidth Memory (HBM). Like Fugaku's ARM cores, the LX2's cores have the ARM vector extension SVE but also Scalable Matrix Extensions (SME).

In total, 45,360 LX2 processors with a combined 13,789,440 cores ran in the HPL benchmark. Their theoretical computing power of 2.736 EFLops is just under that of the previous leader, El Capitan (2.821 EFlops), but Lineshine manages to translate the raw power into performance slightly better. The proprietary interconnect called LingQi delivers 1.6 Tbit/s (200 MByte/s) per CPU, and the Linux derivative Kylin OS runs as the operating system.

Not only in HPL but also in the more memory- and interconnect-intensive HP Conjugate Gradient (HPCG), Lineshine climbs to the top spot, reaching 22 PFlops. Only in HPL-MxP, where calculations are performed in mixed precision friendly to AI accelerators, it falls short. At 7.92 EFlops, Lineshine only ranks 4th, as in CPU-only systems like Lineshine, the speed gain from reduced precision is typically lower than in supercomputers whose computing power is largely provided by accelerators.

67th Top500 Supercomputer List (June 2026): Top 10

Rank

Name

Country

CPU Type

Accelerator

RMax*

Efficiency

Lineshine (Ling Cheng)

China

LX2

2198 PFlops

52.1 GFlops/W

El Capitan

USA

MI300A

AMD Instinct MI300A

1809 PFlops

60.9 GFlops/W

Frontier

USA

Epyc

AMD Instinct MI 250X

1353 PFlops

55.0 GFlops/W

Aurora

USA

Xeon

Intel Xeon GPU Max

1012 PFlops

26.2 GFlops/W

Jupiter Booster

Germany

GH200

Nvidia GH200

1000 PFlops

63.3 GFlops/W

Eni HPC7

Italy

Epyc

AMD Instinct MI300A

572 PFlops

65.4 GFlops/W

Eagle (MS Azure)

USA

Xeon

Nvidia H100

561 PFlops

n.a.

Eni HPC6

Italy

Epyc

AMD Instinct MI 250X

478 PFlops

56.5 GFlops/W

Fugaku

Japan

A64FX

442 PFlops

15.4 GFlops/W

10

Alps

Switzerland

GH200

Nvidia GH200

435 PFlops

61.1 GFlops/W

*RMax is the floating-point performance (FP64) measured by Linpack

Getting Greener?

In the Top 10 of the Green500, the list of the most efficient supercomputers, nothing has changed. Around 73.3 GFlops per watt is still the best that systems can achieve in optimized runs....