Fugaku Next: Japan's next flagship supercomputer comes with Nvidia GPUs

Nvidia can look forward to another prestigious project with its own GPU accelerators for training AI models. The Japanese research institute Riken wants to use Nvidia GPUs in its next supercomputer, Fugaku Next.

Riken operates Japan's fastest supercomputer at the Riken Center for Computational Science (R-CCS). This is currently Fugaku, which even topped the Top500 list in 2020. This list includes the world's fastest systems, but excludes private supercomputers from hyperscalers Google, Meta, Microsoft & Co. as well as new Chinese entries. Today, Fugaku is in 7th place.

Fugaku exclusively uses ARM processors of the type A64FX, designed by the Japanese company Fujitsu. There are almost 160,000 CPUs in the system. Each processor has 48 CPU cores for general tasks and has access to 32 GB of high-bandwidth memory (HBM2). Most of the computing power of the up to 537 petaflops with double precision comes from the integrated vector units.

Nvlink also for Fujitsu CPUs

Fujitsu is designing an A64FX successor for Fugaku Next with the code name Monaka-X. At the time of the announcement in June, Riken already referred to Nvidia's Nvlink Fusion interconnect, which has since been opened. This allows the company's own CPUs and Nvidia's GPUs to be coupled together at high speed. This can facilitate the training of AI models, for example.

The three partners Riken, Fujitsu and Nvidia want to complete the design phase for Fugaku Next by February 2026. Operation from 2027 would be realistic. From then on, the supercomputer should belong to the exascale class, meaning it will be able to perform over a trillion computing operations per second. Nvidia's upcoming Blackwell successor Rubin would be a hot candidate for the GPUs.

(mma)