AI Exascale

China Retains TOP500 Lead in Hyper-Scale Datacenter Ranking With Fewest Debuts in 33 Years

Shenzhen's LineShine sustains 2,198 ExaFLOPS with 13.7 million cores, while the TOP500 logs the fewest new machines in 33 years

• Updated 21 hours ago
A LineShine, no Centro Nacional de Supercomputação de Shenzhen, lidera a TOP500 com 2,198 ExaFLOPS em uma arquitetura baseada exclusivamente em CPUs
A LineShine, no Centro Nacional de Supercomputação de Shenzhen, lidera a TOP500 com 2,198 ExaFLOPS em uma arquitetura baseada exclusivamente em CPUs (Imagem conceitual editorial / NEWS by Box4Dev)

Spec Sheet

  • LineShine totals 13,789,440 cores across 304-core LX2 processors, with no dedicated accelerators
  • El Capitan reaches 1,809 ExaFLOPS at 29.69 MW and is more energy-efficient than the Chinese leader
  • Germany's Jupiter Booster leads on energy efficiency with 1.0 ExaFLOPS at 15.79 MW

China remains at the top of global scientific computing. LineShine, installed at the Shenzhen National Supercomputing Center, sustains 2,198 ExaFLOPS of FP64 performance on the High Performance Linpack (HPL) benchmark while consuming 42.22 MW — a position that keeps the country atop the TOP500 ranking.

The detail that sets the tone for this edition, however, is another one: the list records the fewest debut systems in 33 years. In practice, hyperscalers' demand for artificial intelligence hardware is draining the supply of new machines, and traditional supercomputers have lost priority in supply queues.

A CPU-only architecture

LineShine runs on the LingKun platform, with 304-core Arm-based LX2 processors at 1.55 GHz, totaling 13,789,440 cores, plus the proprietary LingQi interconnect.

It is the only one among the five exascale-class machines to operate without dedicated compute accelerators. That choice explains the gain of only 3.6x when moving from FP64 to mixed precision — a result below that of systems equipped with low-precision accelerators.

The American trio and the race for watts

The United States holds the following positions with three exascale systems. El Capitan, at Lawrence Livermore National Laboratory, uses AMD Instinct MI300A APUs on the HPE Cray EX255a platform and delivers 1,809 ExaFLOPS at 29.69 MW — far more energy-efficient than the Chinese leader.

Frontier, at Oak Ridge National Laboratory, combines 3rd-generation AMD EPYC processors and Instinct MI250X GPUs to reach 1,353 ExaFLOPS at 24.61 MW. Aurora, at Argonne National Laboratory, backed by Xeon CPU Max 9470 processors and Data Center GPU Max accelerators, reaches 1,012 ExaFLOPS at 38.70 MW.

All three American systems share the same foundation: HPE Cray EX platforms and the Slingshot-11 interconnect.

German efficiency

Jupiter Booster, at the Jülich Supercomputing Center in Germany, uses Nvidia GH200 Grace Hopper superchips, the BullSequana XH3000 platform and Nvidia's quad-rail InfiniBand NDR200 interconnect. That yields 1.0 ExaFLOPS at 15.79 MW.

With that figure, the German machine remains the most energy-efficient exascale supercomputer on the entire list — a direct contrast with the CPU-dense model adopted by LineShine.

A frozen ranking

Neither the ten highest-performing systems nor the ten most energy-efficient have changed position since June. The stagnation, combined with the scarcity of newcomers, points to a slower renewal cycle than usual in high-performance computing.

The stated reason is pressure from AI datacenters, which absorb accelerators and high-bandwidth memory before those components reach national laboratories. Without accelerators, machines like LineShine gain relative ground — but lose energy efficiency.

Practical Impact

On the technical front, leadership based solely on CPUs exposes an architectural fork: systems without accelerators scale in cores but convert performance worse to reduced precisions — precisely the regime used in the training and inference of AI models.

On the operational front, the scarcity of new systems makes laboratory renewal more expensive and extends the useful life of already-installed equipment, raising the cost per ExaFLOPS delivered and widening the gap between scientific infrastructure and commercial demand for accelerated computing.