Methodology

TOP500 for the benchmark, operator pages for the facts only operators publish, one link per number.

Which systems are included

105 systems: 100 from the June 2026 TOP500 list (ranks 1 to 100, plus a few lower-ranked systems that an operator documents in detail) and 5 historic landmarks that left the list, such as Summit and Sierra. Each system is one benchmarked configuration, so a site with two partitions has two rows. Anonymised and unnamed list entries are left out.

Where the numbers come from

Rank, Rmax, Rpeak, cores, power, processor, accelerator, interconnect and operating system come from the TOP500 list and system records. What each system is used for, how researchers apply for time, cooling and power design, the facility, upgrade or successor plans, node counts, schedulers and production dates come from the operator’s own pages, its funder or the system vendor. 35 systems, most of the top 40 among them, carry this operator context, and it leads their system pages. The table cites 144 TOP500 pages and 98 other pages; each number links to the page it came from.

Rmax, Rpeak and what is computed

Rmax is measured HPL (LINPACK) performance in double precision and Rpeak is the theoretical peak, both in PFLOP/s (older TFLOP/s and GFLOP/s figures are converted). HPL efficiency is Rmax divided by Rpeak. Efficiency in GFLOP/s per watt is Rmax divided by the power the site reported for its benchmark run, and is empty when no power is reported. Region, vendor group, accelerator vendor and fabric are groupings applied to the sourced values.

How we verify

Each row shows the date we last checked it. A field is filled only when a source states it; empty cells mean the figure is not published. Historic rows keep the rank of the edition where they last appeared, and the Edition column says which. Featured placement is paid and labelled; it never changes a row’s data or its position in the sorted table.

Corrections

Operators and readers can send corrections to hello@hpclist.com with a public source. We update the row and its verified date.

Related lists

Renting capacity instead of studying systems? hpcproviders.com lists HPC clouds and public supercomputer programmes with prices. Tracking AI training clusters? gpuclusterlist.com covers the largest GPU clusters.

Credit and reuse

Benchmark figures are from the TOP500 list, © TOP500.org, operated by Prometeus GmbH; TOP500 is their registered trademark. TOP500 publishes the list freely but under no open licence, so we cite it on every value and do not relicense it or offer it for bulk download. Operator context, summaries, groupings and the computed efficiency columns are written and maintained by HPC List. To cite this table, link hpclist.com and credit TOP500.org for the benchmark figures.

Field definitions

41 fields

Rank: Position in the TOP500 edition named in the Edition column. Historic systems keep the rank of the edition where they last appeared.
Rmax PFLOP/s: Measured double-precision LINPACK (HPL) performance of the listed configuration. Older TFLOP/s and GFLOP/s figures are converted to PFLOP/s.
Rpeak PFLOP/s: Theoretical peak double-precision performance of the listed configuration. This is not AI or mixed-precision throughput.
HPL efficiency %: Rmax divided by Rpeak: the share of theoretical peak the system delivered on LINPACK. Computed from the two sourced figures.
Power kW: Power reported for the benchmark run. Measurement boundaries vary by site; this is not facility power, TDP or cooling capacity. Empty when the system does not report it.
Efficiency GFLOP/s per W: Rmax per watt of reported benchmark power, computed from the two sourced figures (GFLOP/s per W). Empty when power is not reported.
Site: Host or operator named by TOP500 for the installation.
Country: Country or region reported by TOP500 for the site.
City: Site city as reported by TOP500 or the operator. It can be the operator address rather than the machine room.
Region: World region of the installation country (grouping applied by this site).
Vendor: Manufacturer or system supplier exactly as submitted to TOP500.
Vendor group: Supplier grouped for filtering. Bull is the Eviden/Atos line; Other covers single-system suppliers.
System model: Hardware platform in the cited TOP500 edition.
Processor: CPU description in the cited TOP500 edition; clock and core count refer to the processor model.
Accelerator: Accelerator model listed for the system. Empty when TOP500 lists none.
Accelerator vendor: Maker of the accelerator. None listed means TOP500 names no accelerator for the system.
Interconnect: Network fabric exactly as listed in the cited TOP500 edition.
Fabric: Interconnect family derived from the listed interconnect (grouping applied by this site).
Cores cores: Total cores submitted to TOP500, including accelerator cores where TOP500 counts them. Not a CPU-only count and not a CUDA-core count.
Installed: Installation year reported in the TOP500 system record. Not necessarily the year of production acceptance.
Edition: TOP500 edition (YYYY-MM) that supplies the ranking and benchmark data. A historic edition does not assert current operating status.
OS: Operating system from the TOP500 system record.
TOP500 record: Link to the TOP500 database record for this exact submission.
Used for: What the operator says the system is used for: science domains, missions or workloads. Taken from the operator, its funder or the system vendor, not from TOP500.
How to get time: How researchers or customers get time on the system, naming the allocation programme where the operator states one.
Apply: Operator page where researchers apply for time or read the access rules.
Power design: Facility power, energy supply or heat reuse as stated by the operator. Not the benchmark power in the Power column.
Facility: Building, data centre or campus that houses the system, as described by the operator.
Upgrade or successor: Announced upgrade, expansion, successor or retirement plan, with dates where the operator gives them.
Nodes nodes: Compute node count as described by the operator. It can exceed the nodes used in the benchmark run.
Accelerators per node per node: Accelerator packages or superchips per node, as the operator describes them.
CPU memory per node GB: Operator-reported CPU-accessible memory per node.
Node bandwidth Gb/s: Aggregate network bandwidth per compute node, as stated by the operator.
Module bandwidth Gb/s: Network bandwidth per module or GPU, as stated by the operator.
Cooling: Cooling design stated by the operator or supplier.
Scheduler: Job scheduler software reported by the operator.
Production start: Month the operator says the system entered production.
Access model: How the operator describes who can use the system.
Status: Operating status, filled only when the operator states it.
Retired: Operator-reported decommissioning month.
Configuration note: Operator-supported note on partitions or overlapping configurations. Do not add benchmark figures across configurations without checking overlap.