Chips & Compute

China's LineShine Retains TOP500 Lead as Newcomer Count Hits Record Low

LineShine held the No. 1 spot at 2.198 exaflops in the 68th TOP500 list, while only 27 new systems joined — the fewest newcomers in the ranking's history.

China’s LineShine remains world’s fastest supercomputer
China’s LineShine remains world’s fastest supercomputerfrabre / pixabay
By Marcus Bennett6 min read

Why it matters

  • LineShine at China's National Supercomputing Centre in Shenzhen retained the No. 1 spot at 2.198 exaflops HPL in the 68th TOP500, announced at SC26 in Chicago.
  • Only 27 new systems joined the November 2026 TOP500, the fewest newcomers in the list's history.
  • El Capitan held second at 1.809 exaflops and led the HPL-MxP mixed-precision benchmark at 16.7 exaflops.
  • Combined performance across all 500 systems rose 4.4% to 19.57 exaflops; the entry threshold rose 7.3% to 2.857 petaflops.
  • TOP500 organizers, citing Epoch AI's estimate that five U.S. hyperscalers owned 71% of global AI compute capacity in Q4 2025, said growth has shifted from capability systems to AI capacity systems.

China's LineShine retained its position as the world's fastest supercomputer in the 68th edition of the TOP500, announced this week at SC26 in Chicago, posting 2.198 exaflops on the High Performance Linpack benchmark (HPL, the standard measure of a machine's peak floating-point throughput). The more significant number sat elsewhere in the list: only 27 new systems joined the ranking, the fewest newcomers in the TOP500's history, according to the list's organizers.

All ten of the leading systems held their June positions. That static top tier, combined with record-low turnover across the full list, points to a structural shift in where high-performance computing hardware now gets installed — and it is not in traditional capability-class supercomputers.

What does the November 2026 list show?

LineShine, installed at the National Supercomputing Centre in Shenzhen, is a CPU-only system built on the LingKun platform. It runs custom LX2 processors — 304 cores per chip, clocked at 1.55GHz — for a total of 13,789,440 cores, linked by the proprietary LingQi interconnect (the internal network that connects processors across a supercomputer's nodes). Its 2.198 exaflops on HPL kept it comfortably ahead of the field.

El Capitan, the HPE Cray EX255a at Lawrence Livermore National Laboratory in California, held second place at 1.809 exaflops. Frontier at Oak Ridge National Laboratory took third at 1.353 exaflops, Aurora at Argonne's Leadership Computing Facility in Illinois took fourth at 1.012 exaflops, and Germany's JUPITER Booster rounded out the top five at 1.000 exaflops. Those five machines are the only systems on the November 2026 list to cross the exaflop threshold on HPL.

LineShine's lead carries weight beyond raw speed. China built the system on domestic processors and a domestic interconnect, without accelerators from Nvidia or AMD, and it outperforms every U.S. national-lab machine on the linpack benchmark. The result extends the competitive gap the system opened when it first dethroned El Capitan.

How competitive is the rest of the top 10?

The bottom half of the top 10 spans three countries and two corporate operators, with scores clustered far below the exaflop leaders:

  • HPC7 (Italy, Eni S.p.A.): 571.5 petaflops; HPE Cray EX255a; AMD 4th Gen EPYC 24-core 1.8GHz processors with AMD Instinct MI300A accelerators; 3,461,472 cores; Cray Slingshot-11 interconnect.
  • Eagle (U.S., Microsoft Azure): 561.2 petaflops; Microsoft NDv5; Intel Xeon Platinum 8480C 48-core 2GHz processors with Nvidia H100 accelerators; 2,073,600 cores; Nvidia InfiniBand NDR.
  • HPC6 (Italy, Eni S.p.A., Ferrera Erbognone): 477.9 petaflops; HPE Cray EX235a; AMD Optimized 3rd Gen EPYC 64-core 2GHz processors with AMD Instinct MI250X accelerators; 3,143,520 cores; Cray Slingshot-11.
  • Fugaku (Japan, RIKEN Center for Computational Science, Kobe): 442.01 petaflops; Fujitsu A64FX 48-core 2.2GHz processors; 7,630,848 cores; Tofu Interconnect D.
  • Alps (Switzerland, Swiss National Supercomputing Centre): 434.9 petaflops; HPE Cray EX254n; Nvidia Grace 72C 3.1GHz with the GH200 Superchip; 2,121,600 cores; Cray Slingshot-11.

The interconnect picture is concentrated. Six of the top ten systems — El Capitan, Frontier, Aurora, HPC7, HPC6 and Alps — run HPE's Slingshot-11. JUPITER Booster and Eagle rely on Nvidia InfiniBand, while LineShine and Fugaku use their own proprietary networks: LingQi and Tofu Interconnect D, respectively.

Energy and specialized benchmark leaders also held steady. On the Green500, which ranks systems by gigaflops per watt, France's KAIROS stayed at No. 1 with 73.28 gigaflops per watt. France's ROMEO-2025 followed at 70.91, and Germany's Levante GPU extension took third at 69.43. The entire Green500 top ten held its June positions.

What do the other benchmarks say?

LineShine dominated the High Performance Conjugate Gradient benchmark (HPCG, which stresses memory bandwidth and interconnect rather than raw compute and typically yields far lower numbers than HPL). It led at 22 petaflops, ahead of El Capitan at 17.41 petaflops and Japan's Fugaku at 16 petaflops.

El Capitan reclaimed the top spot on the HPL-MxP mixed-precision benchmark, which measures performance on lower-precision arithmetic of the kind used in AI training. It scored 16.7 exaflops, ahead of Aurora at 11.6 exaflops and Frontier at 11.4 exaflops. That result keeps the U.S. flagship ahead on AI-relevant workloads even as LineShine leads on double-precision linpack.

Why did only 27 systems join the list?

The entry threshold for the TOP500 rose 7.3 percent from June, to 2.857 petaflops, even as newcomer count hit its historical floor. Aggregate HPL performance across all 500 systems climbed 4.4 percent to 19.57 exaflops, and 291 systems — 58.2 percent of the list — now use accelerators such as GPUs alongside CPUs.

TOP500 organizers attributed the shrinking influx of new systems to a redirection of hardware purchases. They cited an estimate from Epoch AI, a research organization tracking AI compute, that five U.S. hyperscalers owned 71 percent of global AI compute capacity as of the fourth quarter of 2025.

"This suggests that growth has indeed shifted toward AI factories, and from capability systems to capacity systems," TOP500 organizers wrote in their press release for the November list.

The distinction matters for how the industry reads these rankings. Capability systems — tightly integrated machines like LineShine and El Capitan, built to maximize peak performance on scientific workloads — compete for the top of the TOP500. Capacity systems are sprawling clusters, increasingly owned by cloud providers, that prioritize total throughput for AI training and inference. Eagle, the Microsoft Azure system at No. 7 with Nvidia H100 accelerators, is the top-10 representative of that category.

The 27-newcomer figure, the organizers' argument implies, does not mean the hardware pipeline is drying up. It means the newest silicon is landing in AI data centers that either do not submit linpack results or do not compete for traditional supercomputing prestige — a trend the Epoch AI concentration figure quantifies.

What is at stake in the U.S.–China supercomputer race?

LineShine's second consecutive list at No. 1 extends China's lead in exascale-class capability computing on a fully domestic technology stack. The system's 13.8 million LX2 cores and proprietary LingQi interconnect demonstrate that a CPU-only architecture, without U.S.-export-controlled accelerators, can outperform the flagship machines of the U.S. Department of Energy — El Capitan, Frontier and Aurora, which together occupy positions two through four.

The U.S. retains depth. American systems hold four of the top ten positions if Eagle is counted, and El Capitan's 16.7-exaflop lead on the mixed-precision HPL-MxP benchmark signals stronger performance on the lower-precision arithmetic that dominates AI workloads. Germany's JUPITER Booster, at exactly 1.000 exaflops on Nvidia GH200 Grace Hopper Superchips with 4,801,344 cores, keeps Europe on the exascale list, while Italy's Eni operates two of the top ten and Japan and Switzerland each field one.

For the TOP500 itself, the November 2026 edition marks an inflection point. Aggregate list performance grew just 4.4 percent in six months, the top ten froze completely, and newcomer volume fell to a record low — all while, per Epoch AI's estimate, the majority of the world's fastest silicon flows into five U.S. companies' AI factories. Whether the June 2027 list records a rebound in traditional supercomputing installations or confirms the permanent migration of compute buildout toward AI capacity will be the number to watch.

Source: networkworld.com

Share this article:

More from Marcus Bennett

Marcus Bennett

Show full bio

Senior reporter covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

718 articles

Related articles

  1. Supermicro Ships 20-Plus H15 Servers Built on AMD EPYC 9006 Chips
  2. AMD's 6th Gen EPYC claims up to 2.2x over Intel's flagship Xeon
  3. AMD Ships 256-Core Zen 6 EPYC 'Venice' Chips as AI Agents Drive CPU Demand
  4. Huawei and Alibaba Chip Blitz Arms Xi With Leverage Ahead of Trump Summit
  5. UK's 'Largest Supercomputer' Site Slips Past 2027 Launch

« Previous articleNext article »