StoryPartnership
RIKEN picks NVIDIA GPUs and Fujitsu CPUs for FugakuNEXT, targeting 600+ EFLOPS for AI around 2030
RIKEN launched an international initiative with Fujitsu and NVIDIA to build FugakuNEXT, its first flagship to use GPUs as accelerators. The target is more than 600 exaFLOPS of AI performance in FP8 and up to a hundredfold faster applications than Fugaku, within roughly the same 40 MW of power. It is planned to start operating around 2030 next to RIKEN's Kobe site.
- Fujitsu leads the system and MONAKA-X CPU; NVIDIA designs the GPUs; RIKEN leads software and applications.
- Hardware alone should give 5x or more over Fugaku; software and AI methods another 10-20x.
- Power budget held at about 40 MW, as for Fugaku.
- Japan-US collaboration with the US Department of Energy runs under a MEXT-DOE arrangement on HPC and AI.