AMD supercomputers take gold and silver in latest Top500 as Chinese HPC remains shrouded in secrecy
China stays mum regarding the performance of its latest supercomputing installations.
Top500.org on Tuesday released its 65th list of the world's most powerful supercomputers, revealing the dominance of American AMD-based systems amid a lack of new entries from China, as well as the performance of supercomputers used by AI giants such as xAI and OpenAI.
The latest update to the global high-performance computing rankings places the AMD Instinct MI300A-based El Capitan at the forefront with Rmax performance of 1.7 FP64 ExaFLOPS, followed by the AMD-powered Frontier (1.353 ExaFLOPS) and Intel-based Aurora (1.012 ExaFLOPS) as the top three exascale-class systems, all of which are operated by U.S. Department of Energy laboratories.
The world's fourth-most most powerful supercomputer — Germany's Jupiter Booster, based on Nvidia's GH200 platform — is a new entrant offering Rmax performance of 0.793 ExaFLOPS. Rounding out the top five, Microsoft's Eagle achieved 0.561 ExaFLOPS.
El Capitan, located at Lawrence Livermore National Laboratory in California, achieves 1.742 ExaFLOPS on the HPL benchmark using AMD's fourth-generation EPYC processors and Instinct MI300A accelerators, all within an HPE Cray EX255a framework. It comprises over 11 million cores and uses the HPE Slingshot interconnect.
El Capitan also topped the companion High Performance Conjugate Gradients (HPCG) performance list with 17.1 PetaFLOPS and in the HPL-MxP mixed-precision benchmark with 16.7 ExaFLOPS. As for energy efficiency, it stands at 58.9 GigaFLOPS per watt.
Frontier, the second-place system in HPL, is installed at Oak Ridge National Laboratory in Tennessee and achieved 1.353 ExaFLOPS. It uses AMD's third-generation EPYC CPUs, Radeon Instinct 250X accelerators, and HPE Cray EX235a infrastructure with the same Slingshot interconnect. It operates with over 8.6 million cores.
Frontier ranks third in the HPCG benchmark (behind Japan's Fugaku) with 14.05 petaflops. Frontier also took third place in the HPL-MxP benchmark with 11.4 ExaFLOPS. Its energy efficiency works out to 54.98 GigaFLOPS per watt.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Aurora is housed at Argonne National Laboratory in Illinois and recorded 1.012 ExaFLOPS on the HPL benchmark for the bronze medal in HPL. Built on Intel Xeon CPU Max and Data Center GPU Max (aka Ponte Vecchio) components, it relies on an HPE Cray EX installation using the Slingshot interconnect. It secured the second spot in the HPL-MxP benchmark with 11.6 ExaFLOPS, but it's in fourth place on the HPCG benchmark at 5.6 PFLOPS.
JUPITER Booster is Europe's first system in this performance class. The machine is installed at the Jülich Supercomputing Centre in Germany and reaches its fourth-place spot with a preliminary 793.4 PFLOPS in HPL. JUPITER Booster uses Nvidia Grace Hopper hardware on Eviden's BullSequana XH3000 platform with direct liquid cooling and HP's Slingshot networking. It is currently being brought online.
Microsoft Azure's Eagle system holds the fifth position in HPL with 561 PFLOPS using Xeon Platinum 8480C processors and Nvidia H100 GPUs.
Although CPUs from AMD power two of the world's highest-performing processors and five out of Top 10 machines, a detailed analysis of the June 2025 Top 500 list reveals that Intel processors are used in 294 of the 500 systems. AMD follows with 173 supercomputers.
Nvidia systems appear in 13 entries, showing early traction for the Arm-compatible Grace Hopper architecture. Nine systems are based on other Arm processors, such as Fujitsu's A64FX. An additional six systems use other processor types, including IBM Power9 and China’s Sunway architecture.
The U.S. extended its numerical lead in total systems in the 65th list, while China continued its downward trend as it no longer submits results of its latest systems to Top500.org. As a result, there are 175 American systems, 47 Chinese supercomputers, and 41 German machines in the Top 500 list.
On the energy efficiency front, Germany's JEDI system leads with 72.73 GigaFLOPS per watt, followed by France’s ROMEO-2025 at 70.91 and Adastra 2 at 69.1. All three use BullSequana XH3000 infrastructure. El Capitan and Frontier ranked 26th and below on energy, reflecting a different balance between performance and efficiency.
Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.
Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.
-
Constellar The fact that Huawei has no qualms about bragging about CloudMatrix all but proves that China doesn't have a supercomputer to speak of. If they had one, they'd be bragging about it, too. There are, however, those haters who want to pretend as though China is hiding their supercomputers from the USA authorities; the general idea being that by hiding their supercomputers, there won't be investigations into how China acquired the processors to build their supercomputer. Goofy nonsense, I say. It really comes down to ability; and as always, it's the United States that has the ability, not China (although haters will keep hating. I recommend coping instead of hating)...Reply -
shings Reply
Too avoid getting sanction for its supplied chain, That's why.Constellar said:The fact that Huawei has no qualms about bragging about CloudMatrix all but proves that China doesn't have a supercomputer to speak of. If they had one, they'd be bragging about it, too. There are, however, those haters who want to pretend as though China is hiding their supercomputers from the USA authorities; the general idea being that by hiding their supercomputers, there won't be investigations into how China acquired the processors to build their supercomputer. Goofy nonsense, I say. It really comes down to ability; and as always, it's the United States that has the ability, not China (although haters will keep hating. I recommend coping instead of hating)... -
tamalero Reply
Artificially stiffing competition and then claiming the competition has no "resources" is not the kind of flex you want to pull.Constellar said:The fact that Huawei has no qualms about bragging about CloudMatrix all but proves that China doesn't have a supercomputer to speak of. If they had one, they'd be bragging about it, too. There are, however, those haters who want to pretend as though China is hiding their supercomputers from the USA authorities; the general idea being that by hiding their supercomputers, there won't be investigations into how China acquired the processors to build their supercomputer. Goofy nonsense, I say. It really comes down to ability; and as always, it's the United States that has the ability, not China (although haters will keep hating. I recommend coping instead of hating)... -
jp7189 Do new runs have to be submitted each year or is the list a historical "cumulative" of best performance ever?Reply -
Stomx But have you all noticed that the new AMD Mi300A hardware in ElCapitan is not a single bit faster than the older one Mi250x which works in several years older Frontier? ElCapitan got the first place (1.74 vs 1.35 ExaFLOPS) only because of more cores (11 vs 8.6 million )Reply -
phead128 Reply
Huawei is already sanctioned to the moon and back, so it has no qualms about bragging, because it's already sanctioned.Constellar said:The fact that Huawei has no qualms about bragging about CloudMatrix all but proves that China doesn't have a supercomputer to speak of. If they had one, they'd be bragging about it, too. There are, however, those haters who want to pretend as though China is hiding their supercomputers from the USA authorities; the general idea being that by hiding their supercomputers, there won't be investigations into how China acquired the processors to build their supercomputer. Goofy nonsense, I say. It really comes down to ability; and as always, it's the United States that has the ability, not China (although haters will keep hating. I recommend coping instead of hating)...
The academic institutes and private entities behind China's exascale supercomputers want to protect it's suppliers from getting sanctioned.
Its widely believed that China has exascale capabilities that is not ranked on the supercomputer list. -
cusbrar2 Reply
MI300A is an APU... completely different performance characteristics even though it has similar specs on paper it has unified memory.Stomx said:But have you all noticed that the new AMD Mi300A hardware in ElCapitan is not a single bit faster than the older one Mi250x which works in several years older Frontier? ElCapitan got the first place (1.74 vs 1.35 ExaFLOPS) only because of more cores (11 vs 8.6 million )
MI300A runs similar workloads to MI250X at half the power. I think they also put more effort into scalability and interconnect. -
Stomx Replycusbrar2 said:MI300A is an APU... completely different performance characteristics even though it has similar specs on paper it has unified memory.
MI300A runs similar workloads to MI250X at half the power. I think they also put more effort into scalability and interconnect.
ElCapitan/Frontier:
11,039,616 / 9,066,176 = 1.22 more cores
29,581 / 24,607 MW = 1.20 more consumed electric power
So much for their half the power -
cusbrar2 Reply
That is peak power which while true does not reflect actual usage.Stomx said:ElCapitan/Frontier:
11,039,616 / 9,066,176 = 1.22 more cores
29,581 / 24,607 MW = 1.20 more consumed electric power
So much for their half the power -
Stomx Reply
Actually even worse?cusbrar2 said:That is peak power which while true does not reflect actual usage.