The Japanese Riken institute is working with Fujitsu and Nvidia on FugakuNext. This supercomputer is designed to deliver zetta-scale performance for AI and HPC applications.
The Japanese research institute Riken is collaborating with Fujitsu and Nvidia on a successor to the current Fugaku supercomputer. The new system, codenamed FugakuNext, is set to become the world’s first supercomputer delivering zettascale performance. The developers are aiming for more than 600 exaflops (FP8, sparse) for AI applications. Commissioning is planned for around 2030.
Take with a Grain of Salt for Zetta
A grain of salt is useful here. Computing power for HPC purposes cannot simply be equated with the computing power measured by the Linpack benchmark, which is used to compile the official Top 500. Moreover, zetta-scale here doesn’t mean the system will exceed zettaflops according to its own measurement method: 600 exaflops is 0.6 zettaflops.
The new supercomputer is expected to be up to five times faster than Fugaku at the hardware level. Fugaku currently achieves 442 petaflops. FugakuNext will thus, in principle, easily break the exascale barrier.
Thanks to optimizations in software and algorithms, including techniques such as mixed-precision computing and Physics-Informed Neural Networks, the project aims for a total performance gain of one hundred times for specific (AI) workloads. In any case, Nvidia and the Japanese are planning to build an exceptionally powerful supercomputer.
Combination
FugakuNext combines computing power for traditional simulations with AI capabilities. This happens within an integrated AI-HPC platform. For the first time in the Japanese supercomputer series, the system will include GPUs as accelerators, with Nvidia responsible for the GPU design.
Fujitsu is responsible for the general system design, the compute nodes, and the CPUs. The new CPU, Fujitsu-Monaka-X, is a further development of the existing Monaka processor. This chip is in turn a derivative of the A64FX that powers the first powerful Fugaku supercomputer. The new CPU will get additional SIMD capabilities and matrix calculation functions for better performance in AI inference and HPC applications.
Besides scientific research, FugakuNext aims to contribute to Japan’s technological independence in semiconductors and advanced computing systems. The project also wants to contribute to the automation of scientific research through “AI for Science”. In this respect, Japan has similar ambitions to Europe.