The ten most energy-efficient supercomputers in the world, ranked by GFlops per watt on the November 2025 Green500 list, led by France's KAIROS system.
Quick answer: 1. KAIROS, 2. ROMEO-2025, 3. Levante GPU Extension, 4. Isambard-AI Phase 1, 5. Otus (GPU Only), 6. Capella, 7. SSC-24 Energy Module, 8. Helios GPU, 9. AMD Ouranos, 10. Portage.
The Green500 list ranks the world's most powerful supercomputers not by raw speed, but by how much computational performance they deliver per watt of electricity consumed. This energy-efficiency metric — measured in Gigaflops per watt (GFlops/W) — has become a critical benchmark as the cost and environmental impact of powering these machines continue to grow. This list captures the top ten from the November 2025 Green500 edition. The top three positions are held by BullSequana XH3000 systems using NVIDIA Grace Hopper Superchips, led by the KAIROS system at the CALMIP center in Toulouse, France. The ranking spans academic research centers, national labs, and corporate sites across France, Germany, the United Kingdom, Poland, South Korea, and the United States.
A BullSequana XH3000 system at the CALMIP / University of Toulouse – CNRS in France, KAIROS became the new No. 1 on the Green500 with an efficiency of 73.28 GFlops/W. It uses NVIDIA Grace Hopper Superchips and achieved 3.05 PFlop/s on the HPL benchmark while consuming only 46 kW of power.
The ROMEO-2025 system at the ROMEO HPC Center in Champagne-Ardenne, France, shares the same BullSequana XH3000 architecture as KAIROS but is substantially larger, resulting in slightly lower efficiency at 70.91 GFlops/W. It achieved 9.86 PFlop/s on HPL.
The Levante GPU extension at DKRZ (Deutsches Klimarechenzentrum) in Germany is the third BullSequana XH3000 system in the top three, achieving 69.43 GFlops/W. It delivered 6.75 PFlop/s on HPL and is used primarily for climate research.
The first HPE Cray EX254n system on the list, Isambard-AI Phase 1 at the University of Bristol uses NVIDIA Grace Hopper Superchips with Slingshot-11 interconnect, achieving 68.84 GFlops/W and 7.42 PFlop/s on HPL.
The Otus system at Universitaet Paderborn in Germany is a Lenovo ThinkSystem SD665-N V3 built on AMD EPYC processors and NVIDIA H100 accelerators, achieving 68.18 GFlops/W. It delivered 4.66 PFlop/s on HPL.
Capella at TU Dresden is a MEGWARE-built Lenovo system with AMD EPYC processors and NVIDIA H100 accelerators, achieving 68.05 GFlops/W. It is one of the larger systems in the top ten by core count at 85,248 cores.
The SSC-24 Energy Module at Samsung Electronics in South Korea is an HPE Cray XD670 system with Intel Xeon Gold processors and NVIDIA H100 accelerators, achieving 67.25 GFlops/W.
Helios GPU at Cyfronet in Poland is an HPE Cray EX254n system with NVIDIA Grace Hopper Superchips, achieving 66.95 GFlops/W. It delivered 19.14 PFlop/s on HPL — one of the highest raw performances in the top ten.
The AMD Ouranos system at Atos in France is a BullSequana XH3000 using AMD 4th Gen EPYC processors and AMD Instinct MI300A accelerators, achieving 66.46 GFlops/W. It is the only AMD-accelerated system in the top ten.
Rounding out the top ten, Portage at Hewlett Packard Enterprise in the United States is an HPE Cray EX255a with AMD EPYC processors and AMD Instinct MI300A accelerators, achieving 66.28 GFlops/W. It is the largest system in the top ten by core count at 129,024 cores.
| # | Name | Location | Rmax (PFlop/s) | Power (kW) | Cores | Architecture |
|---|---|---|---|---|---|---|
| 1 | KAIROS | CALMIP / University of Toulouse – CNRS, France | 3.05 | 46 | 13056 | BullSequana XH3000, NVIDIA GH200 |
| 2 | ROMEO-2025 | ROMEO HPC Center – Champagne-Ardenne, France | 9.86 | 160 | 47328 | BullSequana XH3000, NVIDIA GH200 |
| 3 | Levante GPU Extension | DKRZ – Deutsches Klimarechenzentrum, Germany | 6.75 | 110 | 35904 | BullSequana XH3000, NVIDIA GH200 |
| 4 | Isambard-AI Phase 1 | University of Bristol, United Kingdom | 7.42 | 117 | 34272 | HPE Cray EX254n, NVIDIA GH200 |
| 5 | Otus (GPU Only) | Universitaet Paderborn – PC2, Germany | 4.66 | 68.2 | 19440 | Lenovo SD665-N V3, NVIDIA H100 |
| 6 | Capella | TU Dresden – ZIH, Germany | 24.06 | 445 | 85248 | Lenovo SD665-N V3, NVIDIA H100 |
| 7 | SSC-24 Energy Module | Samsung Electronics, South Korea | 3.82 | 69 | 11200 | HPE Cray XD670, NVIDIA H100 |
| 8 | Helios GPU | Cyfronet, Poland | 19.14 | 317 | 89760 | HPE Cray EX254n, NVIDIA GH200 |
| 9 | AMD Ouranos | Atos, France | 2.99 | 48 | 16632 | BullSequana XH3000, AMD MI300A |
| 10 | Portage | Hewlett Packard Enterprise, United States | 24.5 | 370 | 129024 | HPE Cray EX255a, AMD MI300A |
Systems are ranked by energy efficiency as measured in Gigaflops per watt (GFlops/W), using the maximal achieved LINPACK (Rmax) performance divided by system power consumption, as recorded on the Green500 list published in November 2025. The Green500 is a re-ranking of the TOP500 systems by performance-per-watt. All figures — Rmax in PFlop/s, power in kW, and efficiency in GFlops/W — are drawn directly from the Green500 record. Rank 1 is the most energy-efficient.
The Green500 is a re-ranking of the TOP500 supercomputers by energy efficiency — GFlops per watt — rather than raw LINPACK performance. It was created to highlight sustainable computing and to encourage system designers to prioritize power efficiency alongside raw speed.
GFlops/W stands for Gigaflops per watt — billions of floating-point operations per second per watt of electrical power consumed. A higher value means the system delivers more computation for each watt of electricity it draws.
The top three systems — KAIROS, ROMEO-2025, and Levante GPU Extension — all use the same BullSequana XH3000 architecture with NVIDIA Grace Hopper Superchips, built by the French company Bull. The architecture is particularly efficient, and the smaller systems (like KAIROS) benefit from reduced power overhead relative to compute performance.