Nvidia's green500 dominance continues as France's Kairos super takes efficiency title
(2025/11/21)
- Reference: 1763725505
- News link: https://www.theregister.co.uk/2025/11/21/nvidia_green500/
- Source link:
SC25 There's a new efficiency champ at the top of the Green500 ranking of the world's most sustainable supercomputers.
CALMIP's Kairos, located at the University of Toulouse in France, achieved an efficiency of 73.28 gigaFLOPS per watt in the High Performance Linpack (HPL) benchmark.
Since 2007, the Green500 has served as something of a counterpoint to the Top500. Rather than ranking the most powerful supercomputers by sheer performance at cost, the Green500 weighed that performance against the energy required to achieve it, measured in gigaFLOPS per watt.
[1]
This year's [2]ranking represents another victory for Nvidia whose Grace Hopper GH200 superchips power the four most efficient machines on the Green500.
[3]
[4]
In fact, of the 10 most efficient supers in the ranking just two use AMD accelerators.
Rank
Top500 Rank
System
CPU / GPU
Cores
Rmax (PFlop/s)
Power (kW)
Energy Efficiency (GFlops/watt)
1
420
Kairos
Nvidia GH200
13,056
3.05
46
73.282
2
171
Romeo-2025
Nvidia GH200
47,328
9.86
160
70.912
3
225
Levante GPU Extension
Nvidia GH200
35,904
6.75
110
69.426
4
213
Isambard-AI Phase 1
Nvidia GH200
34,272
7.42
117
68.835
5
286
Otus (GPU only)
AMD Epyc 9655 / Nvidia H100
19,440
4.66
?
68.177
6
73
Capella
AMD Epyc 9334 / Nvidia H100
85,248
24.06
445
68.053
7
334
SSC-24 Energy Module
Intel Xeon Gold 6430 / Nvidia H100
11,200
3.82
69
67.251
8
96
Helios GPU
Nvidia GH200
89,760
19.14
317
66.948
9
426
AMD Ouranos
AMD MI300A
16,632
2.99
48
66.464
10
70
Portage
AMD MI300A
129,024
24.50
370
66.277
Over the past few years Nvidia's GH200 has proven to be one of the most energy efficient accelerators ever. Each chip couples a 72-core Grace GPU with 480GB of LPDDR5x memory with an 144GB H200 GPU and links them Nvidia's 900GB/s high-performance NVLink-C2C interconnect.
If any of this sounds familiar, the chip also underpins the Jülich Supercomputing Centre's (JSC) Jupiter supercomputer. That system is now the fourth most powerful supercomputer on the Top500 and Europe's first super to [5]reach an exaFLOP of FP64 performance on the HPL benchmark.
But while these systems are based on the same compute and systems platform, Kairos is much, much smaller achieving just 3.05 petaFLOPs of HPL performance.
[6]
The high-speed interconnects used to stitch each node together, which in this case use Nvidia's InfiniBand NDR200, don't scale linearly. Larger machines therefore tend to perform worse than smaller systems. So while Jupiter is roughly 333x faster than Kairos, it achieves an efficiency of 63.3 gigaFLOPS per watt, making it the 14th most efficient machine on the lineup.
While Nvidia has managed to maintain its grip on the Green500 thus far, AMD did see some movement this time around. The company's Instinct MI300A-based Portage system managed to push past the Intel-based Henri system to claim the number 10 spot on the ranking.
[7]SC25 gets heavy with mega power and cooling solutions
[8]Scientific computing is about to get a massive injection of AI
[9]Europe joins US as exascale superpower after Jupiter clinches Top500 run
[10]Nvidia-backed photonics startup Ayar Labs eyes hyperscale customers with GUC design collab
Nvidia may struggle to stay atop the Green500.
Compared to Hopper, the FP64 performance of Nvidia's newer Blackwell architecture is a bit of a mixed bag. For double precision vector operations Blackwell is 40 percent faster, while for matrix ops it's roughly a third slower.
Meanwhile, Nvidia's Blackwell Ultra GPUs, which launched earlier this year, traded nearly all of their FP64 grunt for improved performance in generative AI workloads like LLM inference.
[11]
We've since received [12]confirmation that Nvidia's next-gen Vera Rubin superchips, due out next year will support FP64. Nvidia's VP and General Manager of Hyperscale and HPC, Ian Buck, deems the data type a "requirement." But while the chip may support FP64 we don't know to what degree. As such, it may be a while before we see new systems from Nvidia that can challenge Grace Hopper on efficiency.
With that said, there's reason to believe that Blackwell's FP64 performance regression may have been an anomaly.
We have already seen several large scientific institutions announce new supercomputers based around Vera Rubin accelerators. At GTC-DC last month, the Department announced that its Mission, Vision, and Doudna supercomputers will all use Nvidia's Vera Rubin accelerators.
The first of these systems are expected to come online sometime next year, though we suspect it'll be at least 2027 before they're ready for a Top500 run. By then they'll also have to contend with AMD's fancy new [13]FP64-optimized accelerator the MI430X. ®
Get our [14]Tech Resources
[1] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/202511sycompsupercomputing&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aSCaqW77M6UudVc5rq-MYgAAAMM&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[2] https://top500.org/lists/green500/2025/11/
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/202511sycompsupercomputing&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aSCaqW77M6UudVc5rq-MYgAAAMM&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/202511sycompsupercomputing&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aSCaqW77M6UudVc5rq-MYgAAAMM&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[5] https://www.theregister.com/2025/11/17/europe_jupiter_supercomputer/
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/202511sycompsupercomputing&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aSCaqW77M6UudVc5rq-MYgAAAMM&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://www.theregister.com/2025/11/20/heavy_industry_invades_sc25/
[8] https://www.theregister.com/2025/11/18/future_of_scientific_computing/
[9] https://www.theregister.com/2025/11/17/europe_jupiter_supercomputer/
[10] https://www.theregister.com/2025/11/16/ayar_guc_collab/
[11] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/202511sycompsupercomputing&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aSCaqW77M6UudVc5rq-MYgAAAMM&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[12] https://www.theregister.com/2025/11/18/future_of_scientific_computing/
[13] https://www.theregister.com/2025/11/18/eviden_france_exascale/
[14] https://whitepapers.theregister.com/
CALMIP's Kairos, located at the University of Toulouse in France, achieved an efficiency of 73.28 gigaFLOPS per watt in the High Performance Linpack (HPL) benchmark.
Since 2007, the Green500 has served as something of a counterpoint to the Top500. Rather than ranking the most powerful supercomputers by sheer performance at cost, the Green500 weighed that performance against the energy required to achieve it, measured in gigaFLOPS per watt.
[1]
This year's [2]ranking represents another victory for Nvidia whose Grace Hopper GH200 superchips power the four most efficient machines on the Green500.
[3]
[4]
In fact, of the 10 most efficient supers in the ranking just two use AMD accelerators.
Rank
Top500 Rank
System
CPU / GPU
Cores
Rmax (PFlop/s)
Power (kW)
Energy Efficiency (GFlops/watt)
1
420
Kairos
Nvidia GH200
13,056
3.05
46
73.282
2
171
Romeo-2025
Nvidia GH200
47,328
9.86
160
70.912
3
225
Levante GPU Extension
Nvidia GH200
35,904
6.75
110
69.426
4
213
Isambard-AI Phase 1
Nvidia GH200
34,272
7.42
117
68.835
5
286
Otus (GPU only)
AMD Epyc 9655 / Nvidia H100
19,440
4.66
?
68.177
6
73
Capella
AMD Epyc 9334 / Nvidia H100
85,248
24.06
445
68.053
7
334
SSC-24 Energy Module
Intel Xeon Gold 6430 / Nvidia H100
11,200
3.82
69
67.251
8
96
Helios GPU
Nvidia GH200
89,760
19.14
317
66.948
9
426
AMD Ouranos
AMD MI300A
16,632
2.99
48
66.464
10
70
Portage
AMD MI300A
129,024
24.50
370
66.277
Over the past few years Nvidia's GH200 has proven to be one of the most energy efficient accelerators ever. Each chip couples a 72-core Grace GPU with 480GB of LPDDR5x memory with an 144GB H200 GPU and links them Nvidia's 900GB/s high-performance NVLink-C2C interconnect.
If any of this sounds familiar, the chip also underpins the Jülich Supercomputing Centre's (JSC) Jupiter supercomputer. That system is now the fourth most powerful supercomputer on the Top500 and Europe's first super to [5]reach an exaFLOP of FP64 performance on the HPL benchmark.
But while these systems are based on the same compute and systems platform, Kairos is much, much smaller achieving just 3.05 petaFLOPs of HPL performance.
[6]
The high-speed interconnects used to stitch each node together, which in this case use Nvidia's InfiniBand NDR200, don't scale linearly. Larger machines therefore tend to perform worse than smaller systems. So while Jupiter is roughly 333x faster than Kairos, it achieves an efficiency of 63.3 gigaFLOPS per watt, making it the 14th most efficient machine on the lineup.
While Nvidia has managed to maintain its grip on the Green500 thus far, AMD did see some movement this time around. The company's Instinct MI300A-based Portage system managed to push past the Intel-based Henri system to claim the number 10 spot on the ranking.
[7]SC25 gets heavy with mega power and cooling solutions
[8]Scientific computing is about to get a massive injection of AI
[9]Europe joins US as exascale superpower after Jupiter clinches Top500 run
[10]Nvidia-backed photonics startup Ayar Labs eyes hyperscale customers with GUC design collab
Nvidia may struggle to stay atop the Green500.
Compared to Hopper, the FP64 performance of Nvidia's newer Blackwell architecture is a bit of a mixed bag. For double precision vector operations Blackwell is 40 percent faster, while for matrix ops it's roughly a third slower.
Meanwhile, Nvidia's Blackwell Ultra GPUs, which launched earlier this year, traded nearly all of their FP64 grunt for improved performance in generative AI workloads like LLM inference.
[11]
We've since received [12]confirmation that Nvidia's next-gen Vera Rubin superchips, due out next year will support FP64. Nvidia's VP and General Manager of Hyperscale and HPC, Ian Buck, deems the data type a "requirement." But while the chip may support FP64 we don't know to what degree. As such, it may be a while before we see new systems from Nvidia that can challenge Grace Hopper on efficiency.
With that said, there's reason to believe that Blackwell's FP64 performance regression may have been an anomaly.
We have already seen several large scientific institutions announce new supercomputers based around Vera Rubin accelerators. At GTC-DC last month, the Department announced that its Mission, Vision, and Doudna supercomputers will all use Nvidia's Vera Rubin accelerators.
The first of these systems are expected to come online sometime next year, though we suspect it'll be at least 2027 before they're ready for a Top500 run. By then they'll also have to contend with AMD's fancy new [13]FP64-optimized accelerator the MI430X. ®
Get our [14]Tech Resources
[1] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/202511sycompsupercomputing&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aSCaqW77M6UudVc5rq-MYgAAAMM&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[2] https://top500.org/lists/green500/2025/11/
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/202511sycompsupercomputing&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aSCaqW77M6UudVc5rq-MYgAAAMM&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/202511sycompsupercomputing&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aSCaqW77M6UudVc5rq-MYgAAAMM&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[5] https://www.theregister.com/2025/11/17/europe_jupiter_supercomputer/
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/202511sycompsupercomputing&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aSCaqW77M6UudVc5rq-MYgAAAMM&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://www.theregister.com/2025/11/20/heavy_industry_invades_sc25/
[8] https://www.theregister.com/2025/11/18/future_of_scientific_computing/
[9] https://www.theregister.com/2025/11/17/europe_jupiter_supercomputer/
[10] https://www.theregister.com/2025/11/16/ayar_guc_collab/
[11] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_specialfeatures/202511sycompsupercomputing&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aSCaqW77M6UudVc5rq-MYgAAAMM&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[12] https://www.theregister.com/2025/11/18/future_of_scientific_computing/
[13] https://www.theregister.com/2025/11/18/eviden_france_exascale/
[14] https://whitepapers.theregister.com/
Pointless...
"There's no point in FP1"