COMPUTE ATLAS Supercomputer supply-chain graph
1153 systems 482 sites Sourced data

Records

The extremes, as a timeline

Not who is fastest right now, who was, and for how long. Everything below comes from fields already on each system's own row; nothing here is a new fact. The first table is the one place this reads closest to a ranking, so it is worth being precise about what it is: only systems in this dataset that we have recorded as holding TOP500 rank 1, on the edition where that is true, the same stored pointer every system page already links out from. It is not a copy of any edition's table, and it has real gaps where no system in our dataset held #1 that edition.

From our stored TOP500 pointers

The fastest machine in the world, by edition

EditionSystemRmax
June 1993 CM-5/1024 59.7 GFlop/s
November 1993 Numerical Wind Tunnel 124.0 GFlop/s
June 1994 Intel Paragon XP/S 140 143.4 GFlop/s
November 1994 Numerical Wind Tunnel 124.0 GFlop/s
June 1995 Numerical Wind Tunnel 124.0 GFlop/s
November 1995 Numerical Wind Tunnel 124.0 GFlop/s
June 1996 Hitachi SR2201/1024 220.4 GFlop/s
November 1996 CP-PACS 368.2 GFlop/s
June 1997 ASCI Red 1.07 TFlop/s
November 2000 ASCI White 4.94 TFlop/s
June 2002 Earth Simulator 35.86 TFlop/s
November 2004 BlueGene/L 478.20 TFlop/s
June 2008 Roadrunner 1.03 PFlop/s
November 2009 Jaguar 1.76 PFlop/s
November 2010 Tianhe-1A 2.57 PFlop/s
November 2011 K computer 10.51 PFlop/s
June 2012 Sequoia 17.17 PFlop/s
November 2012 Titan 17.59 PFlop/s
June 2013 Tianhe-2A 61.44 PFlop/s
June 2016 Sunway TaihuLight 93.01 PFlop/s
June 2018 Summit 148.6 PFlop/s
June 2020 Fugaku 442.0 PFlop/s
June 2022 Frontier 1.35 EFlop/s
November 2024 El Capitan 1.81 EFlop/s
June 2026 LineShine 2.20 EFlop/s

Sourced individually on each system's own page; this table only reorders what is already there. An edition missing from this list is not necessarily one where a system in our dataset failed to reach #1, it may be one we have not yet recorded a pointer for.

From our own Rmax field, not a stored pointer

The fastest system running, year by year

Among systems in this dataset with a measured (not vendor-peak) result, live in a given year, the fastest. This fills years where no system in our dataset held the #1 pointer above but our own data still knows who was fastest among what we track.

YearsSystemRmax
1993-1995 Intel Paragon XP/S 140 143.4 GFlop/s
1996 CP-PACS 368.2 GFlop/s
1997 ASCI Red 1.07 TFlop/s
1998-1999 ASCI Blue-Pacific 2.14 TFlop/s
2000-2001 ASCI White 4.94 TFlop/s
2002-2003 Earth Simulator 35.86 TFlop/s
2004-2006 BlueGene/L 478.20 TFlop/s
2007 HECToR 660.24 TFlop/s
2008-2010 Pleiades 5.95 PFlop/s
2011 K computer 10.51 PFlop/s
2012 Titan 17.59 PFlop/s
2013-2015 Tianhe-2A 61.44 PFlop/s
2016-2017 Sunway TaihuLight 93.01 PFlop/s
2018-2019 Summit 148.6 PFlop/s
2020-2021 Fugaku 442.0 PFlop/s
2022-2023 Frontier 1.35 EFlop/s
2024-2025 El Capitan 1.81 EFlop/s
2026 LineShine 2.20 EFlop/s

Rmax divided by power, in MW

The most power-efficient system running, year by year

Among systems with a measured result and a stated power figure, live in a given year, the highest PFlop/s per megawatt. See power and cooling for the installed-base view this record is drawn from.

YearsSystemPFlop/s per MW
2004-2006 BlueGene/L 0.3
2007 JUGENE 0.4
2008-2010 Pleiades 1.4
2011 Endeavor 3.0
2012 SANAM 3.0
2013 TSUBAME-KFC 4.5
2014-2015 L-CSC 5.5
2016 Sakura (PEZY) 16.7
2017-2018 Suiren2 16.8
2019 A64FX Prototype 16.9
2020 MN-3 40.9
2021 Frontier TDS 62.2
2022-2023 Henri 65.4
2024 JEDI 66.9
2025-2026 Otus 68.3

Stated power draw

The largest power draw running, year by year

Among systems with a stated power figure, live in a given year, the highest.

YearsSystemPower
1993-2003 Cray-1 115 kW
2004 BlueGene/L 1.43 MW
2005 ASC Purple 1.67 MW
2006-2007 NASA Discover 2.00 MW
2008 Pleiades 4.41 MW
2009-2010 Jaguar 6.95 MW
2011-2012 K computer 12.66 MW
2013-2019 Tianhe-2A 18.48 MW
2020-2022 Fugaku 29.90 MW
2023-2025 Aurora 38.70 MW
2026 LineShine 42.22 MW

Power is as reported by the operator or the ranking; the boundary it measures (compute only, or the whole facility) varies by source, so a run here is a record within this dataset, not a verified physical maximum.