Forschungszentrum Jülich celebrates Europe's first Exaflops Supercomputer
The 66th edition of the Top500 list of supercomputers now features Jülich's "JUPITER Booster" with exactly 1 trillion FP64 floating-point operations per second.
Supercomputer Jupiter Booster at FZ Jülich
(Image: Sascha Kreklau/Forschungszentrum Jülich)
Goal achieved: The Jupiter Booster supercomputer at Forschungszentrum Jülich has now reached its planned computing power of 1 trillion FP64 floating-point operations per second. This makes it and remains the fastest European computer.
It also maintained its 4th rank on the 66th edition of the Top500 list of supercomputers. It was published at the International Conference for High Performance Computing, Networking, Storage and Analysis (SC'25) in St. Louis.
Jupiter Booster debuted on the 65th Top500 list in June with 0.793 Exaflops/s (EFlops). Since then, experts in Jülich have put further modules into operation and optimized the system. With 63.3 billion flops per watt of power consumed (63.3 GFlops/W), energy efficiency also increased by 4.5 percent.
Some other of the top ten supercomputers on the latest Top500 list have also been optimized. However, the top 14 positions on the ranking list are occupied by the same systems as in June. Only further down are there some minor shifts and also several new entries, but in the performance range below 140 Petaflops/s (0.14 EFlops).
In addition, there is a new leader on the Green500 list of the most efficient supercomputers, KAIROS at the University of Toulouse. Like Jupiter Booster, it is a Bull Sequana XH3000 machine from the French manufacturer Eviden (formerly Atos), but with a fraction of the performance (3.046 PFlops, Top500 rank 422). However, it operates with 73.28 GFlops/W, making it almost 16 percent more efficient.
Videos by heise
1 Exaflops more Top500 performance
The 66th Top500 list features 45 new supercomputers, including 12 from the USA, 7 from Japan, 6 from Canada, and three in Germany. The most powerful has a performance of 135.4 PFlops, the weakest 2.91 PFlops. Together, they increase the aggregated computing power of the current Top500 list by almost exactly 1 EFlops. This means the performance increase of the 66th Top500 list is relatively weak.
Half of the top ten supercomputers on the 66th Top500 list are located in Europe: two in Italy and one each in Germany, Switzerland, and Finland. However, the USA has the absolute most supercomputer computing power. China, on the other hand, has not participated in the Top500 competition for several years.
Even in light of the enormous growth of AI data centers, the significance of the Top500 list is gradually fading. However, AI behemoths are optimized for different algorithms than most Top500 systems, specifically for AI. They mostly compute with more compact number formats optimized for AI, for example, with TF32 or bFloat16 values during training, and with single-digit bit integer formats during inference. The AI computing performance figures are therefore not comparable to those from the Top500 list.
The Top500 list exclusively considers the results of the High Performance LINPACK benchmark when processing "double precision" floating-point numbers (Dual Precision/DP, FP64). 128 Top500 systems additionally ran the High Performance Conjugate Gradients (HPCG) benchmark for other scientific tasks. Many of the newer Top500 systems are also designed for the use of AI software.
| 66. Top500 List of Supercomputers (November 2025): Top 10 | ||||||
| Rank | Name | Country | CPU Type | Accelerator | RMax* | Efficiency |
| 1 | El Capitan | USA | MI300A | AMD MI300A | 1809 PFlops | 60.9 GFlops/W |
| 2 | Frontier | USA | Epyc | AMD MI250X | 1353 PFlops | 55.0 GFlops/W |
| 3 | Aurora | USA | Xeon | Xeon GPU Max | 1012 PFlops | 26.2 GFlops/W |
| 4 | Jupiter Booster | Germany | GH200 | Nvidia GH200 | 1000 PFlops | 63.3 GFlops/W |
| 5 | Eagle (MS Azure) | USA | Xeon | Nvidia H100 | 561 PFlops | n.a. |
| 6 | Eni HPC6 | Italy | Epyc | AMD Instinct MI | 478 PFlops | 56.5 GFlops/W |
| 7 | Fugaku | Japan | A64FX | – | 442 PFlops | 14.8 GFlops/W |
| 8 | Alps | Switzerland | GH200 | Nvidia GH200 | 435 PFlops | 61.1 GFlops/W |
| 9 | LUMI | Finland | Epyc | Instinct MI 250X | 380 PFlops | 53.4 GFlops/W |
| 10 | Leonardo | Italy | Xeon | Nvidia A100 | 241 PFlops | 32.2 GFlops/W |
| *RMax is the floating-point performance (FP64) measured by Linpack | ||||||
(ciw)