NVIDIA Grabs 81% of TOP500, 2x AI Training of Everyone Else Combined

NVIDIA now powers 81% of the TOP500 supercomputers - that's 405 of the world's 500 fastest machines, up 17 from six months ago. Nine out of every 10 systems new to the list run on NVIDIA gear. Those numbers land a week after ISC High Performance in Hamburg, where the latest rankings dropped.

2x AI Training, 3x Inference: The Gap Widens

Aggregate throughput from NVIDIA systems across the TOP500 now exceeds double the AI training and nearly triple the AI inference of every other platform combined. That's not a fluke of a few outliers - it's a deliberate architectural preference for machines built from the ground up for AI, simulation, and science together. GPU adoption hit a record 238 systems, and NVIDIA networking (mostly Quantum InfiniBand plus some Ethernet) connects 376 machines.

Grace CPU Gets Real: 2.5 Million Cores Shipped

The NVIDIA Grace CPU appeared in 26 systems on the current list, up eight from the previous ranking. Total shipped Grace CPUs hit nearly 2.5 million cores. Grace-based machines occupy both the TOP500 top ten and the Green500 top spot: JUPITER (No. 5), Alps (No. 10), and KAIROS (No. 1 Green500 at 73.3 gigaflops per watt). The Grace Hopper Superchip pairs CPU and GPU with shared memory, cutting overhead for memory-hungry AI workloads. The next-gen Vera CPU, announced earlier this year, builds on that foundation.

Efficiency Crown: KAIROS Leads Green500, Top Eight All NVIDIA

KAIROS, an NVIDIA Grace Hopper system at the University of Toulouse in France, tops the Green500 at 73.3 gigaflops per watt. NVIDIA GPUs occupy the entire top eight of the efficiency list, and nine of the top ten use NVIDIA technologies. Four Grace Hopper systems took the top four spots across France, Germany, and the UK. That's not a coincidence - accelerated computing is the most energy-efficient path to performance.

Global Buildout: 35 European AI HPC Systems in the Pipe

A record 35 NVIDIA-powered AI HPC supercomputers are under development across Europe, equipping over 3 million researchers. JUPITER at Jülich in Germany - Europe's first exascale system - is already mapping the human brain at cellular scale, simulating Earth's climate, and advancing 6G AI models. New Blackwell-based systems (B200 and GB200) entered the list from Asia, Europe, and the US, with the first GB200 machines appearing in Japan. National AI factories are popping up in South Africa, Saudi Arabia, Singapore, and Vietnam.

The story from the TOP500 is simple: the world's AI and HPC buildout runs on NVIDIA silicon. The next list in six months will show whether that share keeps growing - and whether the Vera CPU and Blackwell GPU accelerate the trend further.

Source: NVIDIA Powers Over 400 of the World's 500 Fastest Supercomputers
Domain: blogs.nvidia.com