Nvidia Grace and Hopper architectures reach year-end

It is stated that the Grace processor architecture built by Nvidia on ARM and the Hopper architecture, which will have a multi-chip design for servers, will take its place in a supercomputer by the end of the year.

Nvidia plays big

Based on Nvidia’s 20 years of experience, the ARM-based Grace architecture will launch early next year, but will first appear on the BullSequana XH3000 supercomputer platform built by Atos. Hopper graphics architecture, which was also developed in the 5nm process and will offer multi-chip design for servers, will also be used for the first time on this platform.

This new platform, which will be based on new generation supercomputers, will offer flexible performance ranging from 1 FP64 PetaFLOPS level to 1 FP64 ExaFLOPS level. For artificial intelligence, there are performance options ranging from 10 FP8 ExaFLOPS to 10 FP16 ExaFLOPS. The platform, which will serve for the next 6 years, includes Intel and AMD architectures as well as ARM.

With the Atos Direct Liquid Cooling (DLC) cooling system, 1000W energy consumption per node is possible, providing 50% better efficiency than previous generations.

Top 3 supercomputers in the USA are based on the Cray EX architecture designed by HP. Intel and AMD processors were used in these architectures. With Nvidia Grace and SiPearl Rhea architectures, Atos wants to gain a significant advantage.