At this year’s GTC 2022, NVIDIA has further expanded on the basis of the original Grace CPU, packaged two Grace CPUs together, and launched the Grace CPU Superchip. At the same time, the latest Hopper architecture GPU and Grace CPU are combined to launch Grace Hopper. Previously, the Venado of Los Alamos National Laboratory in the United States and the Alps supercomputer of the Swiss National Supercomputing Center announced that they were built with NVIDIA Grace chips.
According to HPC Wire, Atos has signed a $160 million contract to build a supercomputer called MareNostrum 5, part of the EuroHPC JU program, that will use Nvidia’s Grace chips. It is understood that the new supercomputing system will be installed at the Barcelona Supercomputing Center (BSC) in Spain and will start operations next year.
The FP 64 peak computing performance of MareNostrum 5 is expected to be 314 Petaflops, and the computing performance involving artificial intelligence acceleration exceeds 18 Exaflops. In addition, 200PB of storage space is provided for workload usage, and 400 petabytes of active archive.
It is understood that MareNostrum 5 will use Grace CPU Superchip and NVIDIA H100. Among them, the number of NVIDIA H100 based on the Hopper architecture will reach 4500, providing 270 Petaflops of FP 64 computing performance, while the number of Grace CPU Superchips is about 5730, providing 44 Petaflops of FP 64 computing performance. The MareNostrum 5 supercomputing system will focus on applications such as medical research, chemical simulation, new drug development, climate science, and environmental engineering. Nvidia will also provide the Omniverse software package to help enable more powerful and precise industrial systems and processes.