TOP500
The TOP500 is a ranking of the 500 most powerful computer systems known to its compilers, published twice a year since June 1993 and ordered by each system's measured result on the LINPACK benchmark, in its High Performance Linpack (HPL) implementation, which times how fast a machine solves a large dense system of linear equations in 64-bit floating-point arithmetic.[1][2][3][15] The list is compiled by Erich Strohmaier, Jack Dongarra, Horst Simon and Martin Meuer, and from 1993 until his death in 2014 by Hans Meuer of the University of Mannheim.[4] The June edition is released at the ISC High Performance conference (held in Hamburg, Germany, in 2026) and the November edition at the SC conference in the United States.[5] As of the 68th edition, announced at SC26 in Chicago on November 16, 2026, the No. 1 system is LineShine at the National Supercomputing Centre in Shenzhen, China, at 2.198 exaflop/s, followed by El Capitan, Frontier, Aurora and JUPITER Booster, the five systems that have passed one exaflop/s on HPL.[6]
The list is relevant to artificial intelligence in several ways. More than half of the listed systems now use accelerators (291 of the 500 in November 2026, according to the project's press release), and the gap between HPL's 64-bit arithmetic and the lower precision used in machine learning led to a mixed-precision companion benchmark, HPL-MxP.[6][18] The TOP500 organizers also wrote in November 2026 that growth in computing hardware had shifted toward the large AI compute clusters often described as AI factories, rather than toward the systems reported to the list.[6]
History
Hans Meuer had published yearly counts of systems from the major vector computer makers since 1986, based mainly on those presented at the Mannheim Supercomputer Seminar. According to the project's own account, counts by manufacturer stopped being useful once massively parallel and workstation-derived machines spread, so in 1993 the authors decided to keep a list of the 500 most powerful systems, ranked by LINPACK performance.[2][1] The TOP500 project describes the first list as a one-time exercise for a small conference in Mannheim, followed later in 1993 by a second list for the SC93 conference in the United States; comparing the two showed the value of the statistics, and the June and November cycle has continued since.[6]
The first list, in June 1993, was led by a 1,024-processor Thinking Machines CM-5 at Los Alamos National Laboratory with 59.7 gigaflop/s.[7] Later lists recorded the main performance milestones:
| Date | Milestone | System | HPL result |
|---|---|---|---|
| 1993-06 | First list | CM-5/1024, Los Alamos National Laboratory | 59.7 Gflop/s [7] |
| 2008-06 | First system above 1 petaflop/s | Roadrunner (IBM), Los Alamos National Laboratory | 1.026 Pflop/s [8] |
| 2022-06 | First system above 1 exaflop/s | Frontier (HPE Cray EX), Oak Ridge National Laboratory | 1.102 Eflop/s [9] |
| 2025-11 | First exascale system outside the United States | JUPITER Booster (Eviden), Jülich, Germany | 1.000 Eflop/s [10] |
| 2026-06 | First CPU-only system above 2 exaflop/s | LineShine, National Supercomputing Centre in Shenzhen | 2.198 Eflop/s [11] |
The list has also tracked shifts between countries. In June 2016 China had more systems on the list than the United States (167 against 165), the first time since the list began that the US did not have the most entries.[12] Roadrunner in 2008 was built from IBM QS22 blades using advanced versions of the processor in the Sony PlayStation 3, and Supercomputer Fugaku in 2020 was the first No. 1 system powered by Arm processors.[8][13]
The list is published on top500.org, where readers can download each edition and build sublists. The site's operator is Prometeus GmbH, which reserves the right to include, rank or strike entries at its discretion.[2][5]
Methodology
The HPL benchmark
The TOP500 ranks systems by the best performance they achieve on the LINPACK benchmark, which Jack Dongarra introduced. The version used for the list lets the submitter scale the problem size and tune the software for the machine.[3] Its parallel reference implementation, HPL, solves a random dense linear system in double-precision (64-bit) arithmetic on distributed-memory computers; it needs an implementation of MPI and of the BLAS (or VSIPL) libraries.[14]
To keep results comparable, the solver must use LU factorization with partial pivoting, and the operation count is fixed at
double-precision floating-point operations, where $n$ is the order of the matrix. The rules exclude fast matrix multiplication algorithms such as Strassen's method, algorithms that compute a lower-precision solution and refine it iteratively, and software emulation of floating-point arithmetic using integer arithmetic.[3] The TOP500 authors chose LINPACK because results were available for almost every relevant system and because the problem can scale across a performance range they describe as 12 orders of magnitude.[3][1]
What each entry reports
Each entry lists the manufacturer, system, installation site, country, year of installation or last major upgrade, application area and core count, along with four performance figures:[15]
| Field | Meaning |
|---|---|
| Rmax | Best LINPACK performance achieved |
| Rpeak | Theoretical peak performance |
| Nmax | Problem size at which Rmax was reached |
| N1/2 | Problem size at which half of Rmax is reached |
Rpeak is calculated from the advertised CPU clock rate.[24] The ratio $R_{max}/R_{peak}$ is a system's HPL efficiency: LineShine's 2.198 exaflop/s, for example, is about 80% of its 2.736 exaflop/s theoretical peak.[11] Systems are ordered by Rmax, with ties broken by Rpeak.[15] The list also gives power draw in kilowatts, though not for every system (Eagle's entry has none).[24]
Eligibility and verification
The list is meant to rank general-purpose systems, meaning machines that can be used to solve a range of scientific problems. A system designed specifically to solve the LINPACK problem, or built mainly to rank highly on the TOP500, is disqualified, and the authors reserve the right to verify submitted results independently.[2] They also check a random sample of entries by asking suppliers to put them in direct contact with the installation site, and they limit the systems listed at vendor sites to 5% of each vendor's total installed performance, so that the list reflects the market.[15]
Inclusion depends on submission. For the November list, new sites must be submitted by October 24 and installed by November 1; for the June list, the deadlines are June 5 and June 12.[5] Linpack numbers are collected by Jack Dongarra.[5]
Companion rankings
Three other rankings are published alongside the TOP500.
- Green500 ranks TOP500 systems by HPL performance per watt of electrical power, in gigaflops per watt. Because efficiency tends to fall slightly as systems grow, smaller machines of the same design have an advantage, and the order looks very different from the TOP500.[16] In November 2026 the top three were all BullSequana XH3000 systems with NVIDIA GH200 Grace Hopper superchips, led by KAIROS at CALMIP and the University of Toulouse (CNRS) in France at 73.28 gigaflops per watt; KAIROS ranked 472nd on the TOP500 itself.[16]
- HPCG (High Performance Conjugate Gradients) was designed as a complement to HPL. Its kernels (sparse matrix-vector multiplication, vector updates, global dot products and a symmetric Gauss-Seidel smoother inside a multigrid-preconditioned conjugate gradient solver) are meant to match the data access patterns of a broad set of applications that HPL does not represent.[17] LineShine led HPCG in November 2026 with 22.00 petaflop/s, ahead of El Capitan (17.41) and Fugaku (16).[16]
- HPL-MxP keeps the HPL problem but drops the requirement to compute in 64-bit arithmetic throughout. The LU factorization runs in low precision (its description says likely 16-bit), and an iterative method, GMRES preconditioned with the L and U factors, recovers a 64-bit-accurate solution. Its maintainers say it is meant to highlight the convergence of high-performance computing and AI workloads.[18] Sites have reported results twice a year since 2019, and results using the FP8 data type stopped being admissible in November 2025.[19]
Systems that have reached No. 1
The table lists each system that has led the TOP500, with the first edition in which it was No. 1.[20][11]
| First list at No. 1 | System | Site |
|---|---|---|
| 1993-06 | CM-5 | Los Alamos National Laboratory |
| 1993-11 | Numerical Wind Tunnel | National Aerospace Laboratory of Japan |
| 1994-06 | Intel XP/S 140 Paragon | Sandia National Laboratories |
| 1996-06 | Hitachi SR2201 | University of Tokyo |
| 1996-11 | CP-PACS | University of Tsukuba |
| 1997-06 | ASCI Red | Sandia National Laboratories |
| 2000-11 | ASCI White | Lawrence Livermore National Laboratory |
| 2002-06 | Earth Simulator | Earth Simulator Center |
| 2004-11 | BlueGene/L | Lawrence Livermore National Laboratory |
| 2008-06 | Roadrunner | Los Alamos National Laboratory |
| 2009-11 | Jaguar | Oak Ridge National Laboratory |
| 2010-11 | Tianhe-1A | National Supercomputing Center in Tianjin |
| 2011-06 | K computer | RIKEN Advanced Institute for Computational Science |
| 2012-06 | Sequoia | Lawrence Livermore National Laboratory |
| 2012-11 | Titan | Oak Ridge National Laboratory |
| 2013-06 | Tianhe-2 (MilkyWay-2) | National University of Defense Technology |
| 2016-06 | Sunway TaihuLight | National Supercomputing Center in Wuxi |
| 2018-06 | Summit | Oak Ridge National Laboratory |
| 2020-06 | Supercomputer Fugaku | RIKEN Center for Computational Science |
| 2022-06 | Frontier | Oak Ridge National Laboratory |
| 2024-11 | El Capitan | Lawrence Livermore National Laboratory |
| 2026-06 | LineShine | National Supercomputing Centre in Shenzhen |
Summit, an IBM system at Oak Ridge with six NVIDIA Tesla V100 GPUs in each of its 4,356 nodes, took first place in June 2018 with 122.3 petaflop/s, the first US No. 1 since November 2012.[21] (See Volta for the GPU.) Fugaku, built around Fujitsu's 48-core A64FX processor, led from June 2020 until November 2021.[13][22] Frontier was still the only exascale system on the list in November 2023.[22] El Capitan, with AMD Instinct MI300A accelerators, took the top spot in November 2024 at 1.742 exaflop/s as the third system to pass one exaflop/s, after Frontier and Aurora;[23] a later remeasurement raised it to 1.809 exaflop/s.[10]
The November 2026 list
The 68th list kept the same top ten as June 2026. Only 27 of the 500 systems were new (5.4%), which the organizers called the lowest new-system turnover in the list's history.[6] The top ten were as follows.[24][16]
| Rank | System | Site | Country | Processors and accelerators | Rmax (Pflop/s) | Power (kW) |
|---|---|---|---|---|---|---|
| 1 | LineShine | National Supercomputing Centre in Shenzhen | China | LX2 304-core CPUs, no accelerators | 2,198.40 | 42,220 |
| 2 | El Capitan | Lawrence Livermore National Laboratory | United States | AMD EPYC, AMD Instinct MI300A | 1,809.00 | 29,685 |
| 3 | Frontier | Oak Ridge National Laboratory | United States | AMD EPYC, AMD Instinct MI250X | 1,353.00 | 24,607 |
| 4 | Aurora | Argonne National Laboratory | United States | Intel Xeon CPU Max, Intel Data Center GPU Max | 1,012.00 | 38,698 |
| 5 | JUPITER Booster | EuroHPC / Jülich Supercomputing Centre | Germany | NVIDIA GH200 | 1,000.00 | 15,794 |
| 6 | HPC7 | Eni S.p.A. | Italy | AMD EPYC, AMD Instinct MI300A | 571.50 | 8,735 |
| 7 | Eagle | Microsoft Azure | United States | Intel Xeon Platinum, NVIDIA H100 | 561.20 | not listed |
| 8 | HPC6 | Eni S.p.A. | Italy | AMD EPYC, AMD Instinct MI250X | 477.90 | 8,461 |
| 9 | Supercomputer Fugaku | RIKEN Center for Computational Science | Japan | Fujitsu A64FX | 442.01 | 29,899 |
| 10 | Alps | Swiss National Supercomputing Centre (CSCS) | Switzerland | NVIDIA Grace, NVIDIA GH200 | 434.90 | 7,124 |
Lower down the list, performance kept rising. According to the project's press release, the 500th system reached 2.857 petaflop/s, up 7.3% from June; 27 systems delivered at least 100 petaflop/s; combined HPL performance rose 4.4% to 19.57 exaflop/s; and 291 systems (58.2%) used accelerators. The highest-ranked newcomer was HANGANG, an HPE Cray EX254n system with NVIDIA GH200 superchips at the KISTI Supercomputing Center in South Korea, at No. 12 with 342.6 petaflop/s.[6] The list's highlights page gives slightly different totals for the same edition: 294 systems using accelerators, 19.71 exaflop/s of combined performance and an entry level of 2.87 petaflop/s.[16] In June 2026 the TOP500 project counted HPE as the integrator of six of the top ten, including Aurora, which runs on the HPE Cray EX platform but is credited to Intel as prime contractor, and it counted NVIDIA technology in three (JUPITER Booster, Eagle and Alps).[11]
China and the list
Which Chinese systems appear on the list has depended on what Chinese centers submit. In a rule that took effect on April 8, 2021 and was published the next day, the US Bureau of Industry and Security added seven Chinese entities to the Entity List: the National Supercomputing Centers in Jinan, Shenzhen and Wuxi, the National Supercomputer Center in Zhengzhou, and three chip designers (the Shanghai High-Performance Integrated Circuit Design Center, Sunway Microelectronics and Tianjin Phytium Information Technology).[25] In November 2021 The Next Platform reported that China had not published results for two exascale-class systems that could have made it the first country with exascale machines on the list, and that sources in US, Japanese and European laboratories called the silence "purely political".[26] Tom's Hardware later wrote that China stopped submitting its fastest systems around 2021, and that Jack Dongarra had said Chinese researchers told him they were not permitted to submit.[27]
LineShine's debut at No. 1 in June 2026 was the first time since Sunway TaihuLight in 2017 that a Chinese system led the list.[11][24] It is installed at the National Supercomputing Centre in Shenzhen, one of the centers on the 2021 Entity List, whose entry gives "Shenzhen Cloud Computing Center" as one of its aliases.[25] It was built by the Shenzhen Cloud Computing Center on a custom "LingKun" platform with 304-core LX2 processors at 1.55 GHz, the proprietary LingQi interconnect and the Kylin operating system.[11] Tom's Hardware reported that the center did not name the processor's designer, and that Jon Peddie Research has attributed the chip to Huawei.[27] The Associated Press described the five systems above one exaflop/s as the only publicly verified exascale computers in the world.[28]
Relevance to AI
Precision
HPL measures 64-bit arithmetic, which most scientific simulations need, while machine learning methods reach the results they need at 32-bit and lower precision.[18] When Fugaku took first place in June 2020 with 415.5 petaflop/s on HPL, the TOP500 project noted that its peak in single or further reduced precision, often used in machine learning and AI applications, was over one exaflop/s.[13]
HPL-MxP was created to measure that low-precision capability in a solver that still returns a 64-bit-accurate answer.[18] On the November 2026 results the accelerated systems gained most from it: El Capitan reached 16.68 exaflop/s (9.2 times its HPL score), Aurora 11.64 (11.5 times) and Frontier 11.39 (8.4 times). LineShine reached 7.92 exaflop/s, a 3.6 times speedup.[19] The TOP500 project attributed LineShine's modest gain to a CPU-only design without dedicated low-precision accelerators.[11]
Comparisons with AI systems
Comparisons between AI systems and the TOP500 leader can mix precisions. In April 2025 Google said a 9,216-chip pod of its Ironwood TPU delivered 42.5 exaflops, more than 24 times El Capitan's 1.7 exaflops.[29] The Register pointed out that Google's figure was an FP8 peak while El Capitan's was a measured FP64 LINPACK result, and estimated El Capitan's own dense FP8 peak at a little over 87 exaflops; Google told the publication it did not have information on El Capitan's sustained FP8 performance.[30] The FLOPS article discusses these unit differences in more detail.
AI clusters on and off the list
Some systems built for AI work do appear on the TOP500. Eagle, a Microsoft Azure NDv5 system with Intel Xeon Platinum 8480C processors and NVIDIA H100 accelerators, entered at No. 3 in November 2023, the highest rank a cloud system had reached, with 561.2 petaflop/s.[22] NVIDIA's in-house Selene, a DGX SuperPOD built from DGX A100 systems, was No. 8 in June 2022,[9] and it still appears in the HPL-MxP results.[19] JUPITER Booster and the Alps system at CSCS in Switzerland are both built on NVIDIA GH200 Grace Hopper superchips.[24]
The list only includes machines whose operators send in HPL results, so it is not a census of high-performance computing. The Next Platform's Timothy Prickett Morgan noted in November 2026 that many organizations with very large machines do not submit results.[31] In its November 2026 release, the TOP500 project cited Epoch AI estimates that quarterly additions to global AI computing capacity have grown about 2.5 times per year, that cumulative AI compute exceeded the equivalent of 15 million H100 GPUs, and that five large US hyperscalers owned about 71% of it as of the fourth quarter of 2025. The organizers concluded that "growth has indeed shifted toward AI factories, and from capability systems to capacity systems."[6]
Criticism and limitations
The TOP500 authors acknowledge that they have been criticized for relying on LINPACK since the list began. They write that HPL produces performance figures at the upper end of what any application achieves, that a TOP500 rank says nothing about performance on other applications, and that the list is not a tool for choosing a system: centers need to run benchmarks relevant to their own work. They also say politicians tend to read a rank as a general measure valid for all applications. The authors defend the single benchmark as the simplest way to drive competition between manufacturers, countries and sites, and argue that running LINPACK also serves as a first reliability test for new systems.[1]
HPCG was introduced because HPL's computation and data access patterns represent some important applications but not all.[17] HPL-MxP addresses the opposite gap, the low-precision throughput that HPL rules exclude.[18][3] Coverage is a further limit: entries depend on submission, Chinese centers withheld results for their fastest systems for several years, and the organizers themselves point to AI clusters as the place where much new hardware now goes.[26][27][6]
Continue reading
- High-performance computingHigh-performance computing (HPC) is the use of aggregated computing power, usually many networked computers working together, to tackle large problems in science, engineering and business that a single desktop…Mentioned twice above
- AI computeAI compute is the processing capacity used to train and run artificial intelligence models, together with the chips, memory, networks, data centers, electricity and capital that supply it.Mentioned twice above
- FLOPSFLOPS (floating-point operations per second, also written FLOP/s) is a measure of computing speed: the number of arithmetic operations on floating-point numbers that a processor, accelerator or supercomputer…Listed in See also
- Data CenterA data center is a purpose-built facility, or a dedicated part of a facility, that houses and interconnects information technology and telecommunications equipment together with the power…Listed in See also
- MLPerfMLPerf is the AI industry's standard suite of benchmarks for measuring how fast machine learning systems train and run models, developed and maintained by MLCommonsListed in See also
- AI chip export controlsAI chip export controls are a set of United States regulations that restrict the export, reexport, and in-country transfer of advanced artificial intelligence accelerators (graphics processing units, tensor…Listed in See also
More to read
References
- ^1 ^2 ^3 ^4TOP500. "About". TOP500.org, accessed 2026-11-18. top500.org/project
- ^1 ^2 ^3 ^4TOP500. "Introduction and Objectives". TOP500.org, accessed 2026-11-18. top500.org/...introduction
- ^1 ^2 ^3 ^4 ^5TOP500. "The Linpack Benchmark". TOP500.org, accessed 2026-11-18. top500.org/...linpack
- ^TOP500. "TOP500 Authors". TOP500.org, accessed 2026-11-18. top500.org/...authors
- ^1 ^2 ^3 ^4TOP500. "Call for Participation in the TOP500 / Green500 Lists". TOP500.org, accessed 2026-11-18. top500.org/...call-for-participation
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8TOP500. "TOP500 Leadership Holds Steady as New System Turnover Falls to a Historic Low". TOP500.org, 2026-11-16. top500.org/...w-system-turnover-falls-historic-low
- ^1 ^2TOP500. "June 1993". TOP500.org. top500.org/...06
- ^1 ^2TOP500. "June 2008". TOP500.org. top500.org/...06
- ^1 ^2TOP500. "ORNL's Frontier First to Break the Exaflop Ceiling". TOP500.org, 2022-05-30. top500.org/...r-first-to-break-the-exaflop-ceiling
- ^1 ^2TOP500. "November 2025". TOP500.org. top500.org/...11
- ^1 ^2 ^3 ^4 ^5 ^6 ^7TOP500. "LineShine Debuts at No. 1 as the TOP500 Enters a New Global Exascale Era". TOP500.org, 2026-06-23. top500.org/...op500-enters-new-global-exascale-era
- ^TOP500. "June 2016". TOP500.org. top500.org/...06
- ^1 ^2 ^3TOP500. "June 2020". TOP500.org. top500.org/...06
- ^A. Petitet, R. C. Whaley, J. Dongarra, A. Cleary. "HPL - A Portable Implementation of the High-Performance Linpack Benchmark for Distributed-Memory Computers", version 2.3. Netlib, 2018-12-02. netlib.org/...hpl
- ^1 ^2 ^3 ^4TOP500. "TOP500 Description". TOP500.org, accessed 2026-11-18. top500.org/...top500_description
- ^1 ^2 ^3 ^4 ^5TOP500. "Highlights - November 2026". TOP500.org. top500.org/...highs
- ^1 ^2HPCG Benchmark. "HPCG Benchmark". hpcg-benchmark.org, accessed 2026-11-18. hpcg-benchmark.org
- ^1 ^2 ^3 ^4 ^5HPL-MxP. "HPL-MxP Mixed-Precision Benchmark". hpl-mxp.org, accessed 2026-11-18. hpl-mxp.org
- ^1 ^2 ^3HPL-MxP. "Results". hpl-mxp.org, accessed 2026-11-18. hpl-mxp.org/results.md
- ^TOP500. "TOP #1 Systems". TOP500.org, accessed 2026-11-18. top500.org/...top-systems
- ^TOP500. "June 2018". TOP500.org. top500.org/...06
- ^1 ^2 ^3TOP500. "November 2023". TOP500.org. top500.org/...11
- ^TOP500. "November 2024". TOP500.org. top500.org/...11
- ^1 ^2 ^3 ^4 ^5TOP500. "November 2026". TOP500.org. top500.org/...11
- ^1 ^2Bureau of Industry and Security, Department of Commerce. "Addition of Entities to the Entity List". Federal Register, 2021-04-09 (document 2021-07400). federalregister.gov/...entities-to-the-entity-list
- ^1 ^2Nicole Hemsoth Prickett. "Why Did China Keep Its Exascale Supercomputers Quiet?". The Next Platform, 2021-11-15. nextplatform.com/...1645477
- ^1 ^2 ^3Luke James. "China tops the list of fastest supercomputers with a CPU-only behemoth, ending US champion El Capitan's reign". Tom's Hardware, 2026-06-24. tomshardware.com/...puter-ending-el-capitans-reign
- ^Associated Press. "Chinese supercomputer displaces US machines as world's fastest for first time since 2017". Spectrum News NY1, 2026-06-23. ny1.com/...orlds-fastest-for-first-time-since-2017
- ^Amin Vahdat. "Ironwood: The first Google TPU for the age of inference". Google, 2025-04-09. blog.google/...ironwood-tpu-age-of-inference
- ^Tobias Mann. "Google's 7th-gen Ironwood TPUs promise 42 AI exaFLOPS pods". The Register, 2025-04-10. theregister.com/...1058980
- ^Timothy Prickett Morgan. "Supercomputer Installs Take A Dive In The Second Half Of 2026". The Next Platform, 2026-11-16. nextplatform.com/...5310221
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
1 revision · v1 · 3,440 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check before publication on November 18, 2026: 104 claims checked against 31 cited and independent sources; 2 material and 8 minor corrections applied before publication.
Cite this page: AI Wiki. "TOP500." aiwiki.ai, updated 18 Nov 2026, fact-checked 18 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/top500