- Fugakunext will combine Fujitsu processors and NVIDIA GPUs in the next national supercomputer from Japan
- The system targets the performance of 600EFLOPS FP8 with a 100x application performance gain
- Riken, Fujitsu and Nvidia see the project as defining a new AI-HPC standard
Japan is preparing its next national supercomputer, Fugakunext, thanks to a collaboration between Fujitsu, Nvidia and Riken.
The system is provided for the operation around 2030 and aims to mix simulation and artificial intelligence in a closely integrated platform.
For the first time in a Japanese flagship project, GPUs will be used as accelerators. Nvidia will (not surprisingly) design the GPU infrastructure, Fujitsu will manage the integration of processors and systems, and Riken will be involved in the work of software and algorithm.
GPU Feynman
The result should be an “AI-HPC platform” designed for science, industry and discovery focused on AI.
The performance objectives of the supercalculator are certainly ambitious. Fugakunext is designed to offer more than 600 FP8 AI performance, which would make it the most powerful AI supercalculator to date.
The system should also reach an increase up to a hundred application performance compared to Fugaku, while remaining roughly the same power budget of 40 MW.
The Nvidia long -term roadmap points to the GPU architecture of Feynman (named after the theoretical physicist Richard Feynman) arriving around 2028, it could therefore play a role in the spread of Fugakunext.
Fujitsu develops a successor to its Monaka processor for the project, temporarily named Monaka-X, with more hearts, extended SIMD capacities and an ARM matrix calculation engine for IA inference.
Coupled with NVIDIA accelerators, the system should execute large simulations alongside demanding AI workloads.
The equipment alone will not provide target gains so that the project is also based on innovations such as substitution models, mixed accuracy arithmetic and neural networks informed of physics to accelerate performance while preserving precision.
Makoto Gonokami, president of Riken, said: “It is a great honor for Riken to collaborate with Fujitsu and Nvidia to advance the development of Fugakunext. Since ancient times, humanity has built civilizations and advanced societies through the science of computer science. Today, the emergence of AI, advanced scenarios and quantum computers.
Ian Buck, vice -president of Nvidia, added: “Fugakunext will offer zettascal performance with application speeds almost 100 times faster – in the same energy imprint as its predecessor – accelerate research, increase industrial competitiveness and lead progress for people in Japan and worldwide.”
