Vera Rubin chips displace Intel for Blue Lion supercomputer
The Blue Lion supercomputer in Germany will be one of the first to use the next generation Vera Rubin chips from Nvidia, displacing Intel in the leading system in Europe.
Blue Lion is planned to come online in early 2027 at the Leibniz Supercomputing Centre (LRZ), providing 30x more performance than the current SuperMUC-NG. VeraRubin is significant for Nvidia as it is displacing both Intel and AMD x86 processors in supercomputers for the first time, rather than just providing the GPU accelerators.
This highlights a key shift in scientific compute workloads to AI that is reflected in the latest TOP500 list of supercomputers. Over 380 systems in that list use Nvidia GPUs, including the Jupiter system at the Jülich Supercomputing Centre, which was the only new entry in the top ten, at number four. Eighty three of the top 100 systems on the TOP500 now use GPUs compared to only 17 which use CPUs only.
The Blue Lion system will be built by HPE Cray and use the Vera CP and Rubin GPU with fanless liquid cooling and the HP Slingshot interconnect. SuperMUC-NG has a power consumption of 2.8MW using Intel Xeon processors, but the power consumption of Blue Lion will be determined by the number of VeraRubin chips available in the €250m budget as well as the performance of the system management software from HP.
The first system to use VeraRubin (see below) will triple the power consumption, but the LRZ systems have been optimised for power efficiency, so the overall power is expected to be lower than this. The top three systems in the latest Green500 list of the world’s most energy-efficient supercomputers use the current GH200 Grace Hopper Superchips, and nine of the top 10 use Nvidia GPUs.
Vera has 88 customised ARMv9 dual thread cores with 1.8Tbits of NVLink C2C (chip2chip) interconnect. Rubin is composed of two reticle-sized chips using the Blackwell architecture on a substrate with 288Gbits of HBM4 high speed memory. The two devices have a coherent memory architecture between the CPU and GPU and will launch in the second half of 2026.
“We will have more details to share as we move to production of Vera Rubin,” said Dion Harris vice president of AI factory solutions at Nvidia, who points to the Isambard-AI supercomputer in Bristol, UK, as an example of the shift to AI using the existing GraceHopper devices.
Blue Lion in Germany will be used for the first generative climate foundation model with kilometre scale precision. “This is a huge revolution for the climate community where you can make 30 year predictions across 12 parameters at a 1km scale,” said Harris. “This ‘climate in a bottle’ uses AI to compress climate simulation data by 3000x.”
This will follow the Doudna supercomputer at the Lawrence Berkeley National Lab in the US which is scheduled to come online by the end of 2026. This will be built by Dell and will use the Nvidia’s Quantum-X800 InfiniBand networking for advanced fusion energy research and modelling as well as materials discovery.
This will provide 10x the performance for 3 to 5x the power consumption of the current Perlmutter supercomputer based on the AMD EPYC x86 processor at 2945kW. This puts the system at 1200W per VeraRubin node and an overall power consumption of 9MW.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
