Sapphire Rapids, Nvidia Hopper, Nvidia Grace, Emerald Rapids and Rialto Bridge – all of this is to be installed in the new MareNostrum5. In addition, there is also technology from IBM and Nvidia's InfiniBand technology, which produce a real experimental system, not with maximum performance but with interesting variety.
With a computing power of up to 314 PetaFLOPS, MareNostrum5 is designed to support medical research in Europe in drug research, vaccine development, virus spread simulation, AI applications and big data processing. The system will also provide computing power for HPC-specific complex applications such as climate research, engineering, materials and geosciences that need to be managed outside of the cloud, Atos explains in the press release.
From the racks of previous versions of MareNostrum to the site that will host #MareNostrum5
💻Our machines have allowed us to take research further
💜Looking forward to hosting #MN5 to continue working for science and society
🤝@EuroHPC_JU @CienciaGob @recercat @la_UPC pic.twitter.com/oEzfQ2Nv4e
— BSC-CNS (@BSC_CNS) June 22, 2022
It gets interesting when it comes to the technical data – but above all it is completely confusing at first. As Timothy Prickett Morgan from The Next Platform lists as a site specializing in this topic, half and incorrect information has been circulating for over a week, but it is now finally finally sorted. It's the additional experimental platforms and evaluation solutions that have caused a mess.
The main components of the system and performance are Intel Sapphire Rapids and Nvidia Hopper H100. The basis for this is the new Atos BullSequana XH3000, two CPUs and four GPUs are installed per node. Lenovo contributes further CPU-only nodes, based on the ThinkSystem SD650 V3 “Neptune” they also rely on Intel Sapphire Rapids.
Then there are two more stages: Nvidia Grace as the upcoming ARM CPU is also used in the supercomputer, but this CPU cluster only offers around 2 PetaFLOPS of performance, i.e. not even one percent of the entire system. It is therefore probably intended for evaluation. The same probably also applies to the nodes with Intel Emerald Rapids and Intel Rialto Bridge – both next-gen solutions for the Xeon processors and GPU accelerator cards. With 4 PetaFLOPS, however, the systems are also primarily test models for later supercomputers.
A lot, a lot of hardware, but still nothing European, can be the summary here at the end. However, the system will “only” cost around 150 million euros and should also be ready in the coming year, the European project is still not likely to be ready. In Europe, Europe's first exascale supercomputer is to be in Jülich by the end of next year, but the hardware is still officially unknown. Costs here: 500 million euros.