The 68th edition of the TOP500 supercomputer rankings, released at the SC26 conference in Chicago on November 16, 2026, revealed unprecedented stability at the summit of global high-performance computing, keeping the entire top 10 unchanged from June 2026 while registering a historic low of just 27 new entrants. China’s LineShine retained the world No. 1 ranking, while institutional computing centres navigate rising entry thresholds, shifting capital allocation toward commercial AI clusters, and mounting software portability demands across heterogeneous architectures.
Leadership stability and record-low turnover at SC26
The release of the 68th TOP500 list at the SC26 supercomputing conference confirmed that the upper tier of international capability computing has entered an era of consolidation. For the first time in the ranking’s 33-year history, all ten leading systems held their exact placements from the preceding June edition, and only 27 new machines entered the overall 500-system roster. TOP500 project organizers noted that this represents the lowest turnover ever recorded across the biannual list, falling well below historical averages where dozens of new installations routinely displaced older clusters.
China’s LineShine supercomputer, hosted at the National Supercomputing Centre in Shenzhen, retained the global crown it captured earlier in the year. Delivering a measured High Performance Linpack (HPL) benchmark score of 2.198 exaflops, LineShine remains the world’s fastest publicly benchmarked system. Independent reporting from Network World highlighted that LineShine achieves this performance without discrete accelerators, relying instead on custom 304-core LX2 processors connected via a proprietary LingQi interconnect fabric.
The United States Department of Energy (DOE) systems maintained their presence directly behind the leader. El Capitan, deployed at Lawrence Livermore National Laboratory (LLNL), retained second position with an HPL mark of 1.809 exaflops, followed by Oak Ridge National Laboratory’s Frontier at 1.353 exaflops and Argonne National Laboratory’s Aurora at 1.012 exaflops. Germany’s JUPITER Booster module, stationed at Forschungszentrum Jülich, held the fifth position. Together, these five installations represent the complete global tier of verified exascale systems operating above 1.0 exaflops on double-precision Linpack.
Comparative performance benchmarks across leading systems
While the top tier remained locked, the overall computational floor of the list continued to rise. The minimum performance threshold required to enter the TOP500 climbed 7.3% compared to June 2026, reaching 2.857 petaflops. Combined HPL performance across all 500 installations grew by 4.4% to reach an aggregate 19.57 exaflops. Accelerated hardware configurations continue to dominate the broader cohort, with 291 systems (58.2% of the list) utilizing specialized coprocessors or GPUs.
| Rank | System | Institution / Country | Architecture | HPL (Exaflops) | HPCG (Petaflops) | HPL-MxP (Exaflops) |
|---|---|---|---|---|---|---|
| 1 | LineShine | NSCS Shenzhen, China | Custom 304-core LX2 CPU (LingQi) | 2.198 | 22.00 | N/A |
| 2 | El Capitan | LLNL, United States | AMD 4th Gen EPYC + MI300A (Slingshot-11) | 1.809 | 17.41 | 16.70 |
| 3 | Frontier | ORNL, United States | AMD 3rd Gen EPYC + MI250X (Slingshot-11) | 1.353 | 14.05 | 11.40 |
| 4 | Aurora | ANL, United States | Intel Xeon Max CPU + Data Center GPU Max (Slingshot-11) | 1.012 | 5.60 | 11.60 |
| 5 | JUPITER Booster | FZ Jülich, Germany | Eviden BullSequana + NVIDIA GH200 (InfiniBand) | 1.000 | N/A | N/A |
| 14 | Lux | ORNL, United States | AMD EPYC + Instinct MI355X (Pensando) | 0.221 | N/A | N/A |
Beyond traditional raw Linpack metrics, secondary benchmarks illustrated distinct architectural strengths. LineShine retained the top spot on the High Performance Conjugate Gradient (HPCG) benchmark with 22.00 petaflops, reflecting exceptional memory subsystem bandwidth, followed by El Capitan at 17.41 petaflops and Japan’s Fugaku at 16.00 petaflops. On the HPL-MxP mixed-precision benchmark, which evaluates lower-precision arithmetic commonly leveraged in scientific artificial intelligence, El Capitan led the field at 16.70 exaflops, outpacing Aurora (11.60 exaflops) and Frontier (11.40 exaflops).
The shift from capability computing to AI factories
The record-low influx of only 27 new systems reflects an industry-wide reorientation of computing hardware investments. Rather than funding monolithic national capability systems designed primarily for double-precision scientific simulation, major technology vendors and government sponsors have shifted capital toward hyperscale AI factories and capacity clusters. TOP500 organizers observed that advanced accelerator manufacturing capacity has been largely absorbed by commercial infrastructure operators, citing research from Epoch AI indicating that five United States hyperscalers controlled roughly 71% of global AI compute capacity by late 2025.
Within this constrained environment, AMD established a notable milestone by expanding its footprint across the list. According to detailed industry analysis from Tom’s Hardware, AMD-powered systems captured over 40% of the TOP500 rankings, representing more than 200 active installations and reflecting a 14% year-over-year increase. AMD technology currently underpins two of the world’s three fastest supercomputers as well as high-profile newcomers.
The highest-ranking new entrant in the 68th list was the Lux supercomputer, which debuted at No. 14. Detailed specifications published by the Oak Ridge Leadership Computing Facility confirm that Lux achieved 220.79 petaflops on HPL. Built in a public-private partnership between AMD, HPE, and the Department of Energy as part of a $1 billion initiative, Lux incorporates AMD EPYC host processors, Instinct MI355X accelerators with 288 GB of HBM3E memory, and AMD Pensando networking inside HPE ProLiant Compute XD685 chassis. An official release from the AMD Newsroom outlined Lux’s role as the first dedicated AI factory supporting the DOE’s Genesis Mission, targeted specifically at foundation model training for fusion reactor design, materials discovery, and climate modeling.
Energy efficiency and Green500 rankings
Energy constraints continue to dictate high-performance infrastructure design, as demonstrated by the corresponding Green500 rankings. Much like the TOP500, the Green500 top 10 exhibited complete stability, with leading European installations maintaining their benchmarks from June 2026. Official results on Green500 confirmed that France’s KAIROS system remained the world’s most energy-efficient supercomputer, achieving a leading efficiency rating of 73.28 gigaflops per watt.
KAIROS, deployed at the CALMIP supercomputing centre operated by the University of Toulouse and the CNRS, is built on Eviden BullSequana XH3000 architecture utilizing NVIDIA Grace Hopper GH200 Superchips. It was followed by France’s ROMEO-2025 at the University of Reims Champagne-Ardenne at 70.91 gigaflops per watt, and Germany’s Levante GPU extension at the German Climate Computing Centre (DKRZ) at 69.43 gigaflops per watt. For academic institutions confronting municipal power caps and escalating electricity tariffs, monitoring institutional power efficiency and compute carbon emissions through tools like CodeCarbon has become an operational necessity alongside raw floating-point throughput.
What the 68th rankings mean for researchers and university computing
The trends showcased at SC26 carry direct operational implications for university research offices, campus high-performance computing directors, and computational investigators:
- Intensified allocation competition: With system turnover at historic lows and fewer newly commissioned national capability machines coming online, open-access allocation queues on national lab facilities (such as DOE INCITE and ALCC programs) face heightened oversubscription. Principal investigators must budget for extended wait times or pursue hybrid campus allocations.
- Software portability imperatives: The architectural divergence across the rankings—ranging from China’s CPU-only LX2 processors to AMD MI300A/MI355X and NVIDIA Grace Hopper GH200 systems—precludes reliance on single-vendor programming models. Scientific research groups must increasingly adopt vendor-neutral abstraction frameworks such as Kokkos to maintain performance portability across diverse heterogeneous clusters without extensive code rewrites.
- Computational reproducibility standards: As leadership supercomputers adopt specialized hardware configurations and complex container stacks, reproducing simulation results across differing machine environments grows more difficult. Implementing robust package management and deterministic environment definitions via tools like GNU Guix ensures computational experiments remain verifiable when migrated between university clusters and national exascale resources.
- Institutional data infrastructure alignment: High-throughput AI factories require massive parallel storage fabrics capable of feeding multi-gigabyte bandwidth per node. Aligning campus data workflows with national storage platforms and institutional standards cataloged under CASRAI’s Research data infrastructure domain ensures research data remains accessible throughout extended simulation and training campaigns.
Roadmap and upcoming exascale deployments
The current plateau in top-tier ranking turnover is expected to persist until the next generation of planned leadership supercomputers reaches final acceptance. In the United States, the DOE and Oak Ridge National Laboratory are preparing for Discovery, a next-generation flagship exascale system developed jointly with AMD and HPE that is scheduled to debut around 2028. In Europe, the full commissioning of the JUPITER exascale system in Germany will further expand scientific capacity during 2027.
The 69th edition of the TOP500 rankings will be unveiled at the ISC High Performance conference in Hamburg, Germany, in June 2027, where researchers and facility managers will track whether the entry threshold continues to rise and whether new national capability systems emerge to break the current top-10 stability.
Sources
- TOP500 Project, TOP500 List: November 2026, November 16, 2026. Checked 2026-11-16.
- TOP500 Project, Green500 List: November 2026, November 16, 2026. Checked 2026-11-16.
- Network World, China’s LineShine remains world’s fastest supercomputer, November 16, 2026. Checked 2026-11-16.
- Tom’s Hardware, Anton Shilov, AMD grabs over 40% of new TOP500 list for supercomputers, up 14% since last year — new Lux system developed between AMD and Department of Energy debuts at no. 14 with EPYC CPUs and Instinct MI355X GPUs, November 16, 2026. Checked 2026-11-16.
- Oak Ridge Leadership Computing Facility, Lux – Oak Ridge Leadership Computing Facility, November 16, 2026. Checked 2026-11-16.
- AMD Newsroom, AMD Expands Frontier of Scientific Computing at SC26, November 16, 2026. Checked 2026-11-16.







