Nvidia and Cineca announced the development of the Leonardo supercomputer, which promises to be ” the fastest AI supercomputer in the world ” when it arrives next year, and to make this possible, about 14,000 Nvidia A100 GPUs will be installed inside. it. which translates to about 96 million CUDA cores under the Ampere architecture.
Specifically, each node will have four Nvidia A100 GPUs, which will be accompanied by a single CPU, which this time is an Intel Sapphire Rapids @ 10nm +++ SuperFin that will give access to DDR5 memory and the PCI-Express 5.0 interface.
In regards to IA FP16 training speed performance, there is talk of the theoretical maximum performance of more than 10 Exaflops thanks to its Tensor Core cores, although of course, it also promises great power in FP64 calculations.
In addition to Leonardo, Cineca will deploy three other supercomputers powered by Nvidia A100 graphics: MeluXina, with 800 GPUs; the National Center for Supercomputing IT4Innovations with 560 GPUs; and curiously, Vega (name of AMD’s graphics architecture) with 240 Nvidia A100 GPUs, although the name was chosen by the Slovenian mathematician Jurij Vega since this team will be deployed in Slovenia while the other two will be deployed in Luxembourg and the Czech Republic respectively.