Aug. 22, 2025 — RIKEN is collaborating with Fujitsu Limited and NVIDIA on an initiative to develop “FugakuNEXT,” which RIKEN called Japan’s next-generation flagship supercomputer.
RIKEN said the system will be the first flagship supercomputer to adopt GPUs as accelerators, with NVIDIA leading the design for the GPU infrastructure. Fujitsu, which has already been working with RIKEN, will lead the basic system design, compute nodes, and CPUs. As a foundation for high-performance computing infrastructure, the three organizations will work together to build an “AI-HPC platform, according to RIKEN.
The FUJITSU-MONAKA-X CPU being designed by Fujitsu, will be an enhanced version of the FUJITSU-MONAKA. With ultra-many cores and extended SIMD functionality providing high scalability, it will deliver performance for HPC applications such as numerical simulations. In addition, by integrating the matrix computation engine (Arm SME) into the CPU, it enables low-latency AI inference processing. When combined with the GPU with high-bandwidth, tightly coupled integration, the system will deliver exceptional performance in AI training and GPU-optimized HPC applications. These advancements will enable the system to address a wide spectrum of computing needs, from general-purpose workloads to numerical simulations and AI processing.
The computation accelerator (GPU) designed by NVIDIA will enable performance improvements for HPC applications requiring advanced parallel processing and for AI applications including generative AI whose adoption is rapidly expanding in industry as well as for next-generation applications that integrate AI and HPC, RIKEN said.
To further improve application performance, software-side innovations will encompass the use of mixed-precision computing, the use of AI-oriented hardware, such as low-precision computing units for high-precision calculations (e.g., the Ozaki scheme), and algorithmic optimizations including the adoption of surrogate models and PINNs (Physics-Informed Neural Networks) to replace complex computations with AI.
Ultimately, by combining the hardware performance improvements achieved through the joint development by RIKEN, Fujitsu, and NVIDIA with the algorithmic and software innovations driven by their collaboration, the project aims to deliver, within the same approximately 40MW power constraint as during the development of Fugaku, up to a hundredfold overall increase in application performance – comparable to the scale of improvement targeted in the transition from the “K” computer to Fugaku, according to RIKEN. In addition, by advancing AI for science – including AI-driven hypothesis generation and validation, automated code generation, and the automation of physical experiments – the project will seek to accelerate the entire cycle of scientific discovery.
“Aiming for operation around 2030, FugakuNEXT is envisioned to deliver not only the simulation performance pursued by previous supercomputers, but also world-leading capabilities in both simulation and AI, with the two working in close integration as a unified AI-HPC platform,” RIKEN said. “To this end, FugakuNEXT will extend the CPU technologies developed for Fugaku, while incorporating GPUs as accelerators, aiming to maximize execution performance for both simulation and AI applications. At the same time, to continuously generate scientific achievements through the use of next-generation computing infrastructure, it is becoming increasingly important to pursue not only advances in hardware technology, but also comprehensive innovations in software and algorithms.”
RIKEN said the three organizations will combine RIKEN’s strengths in software and algorithm technologies, Fujitsu’s expertise in CPU and system integration technologies, and NVIDIA’s GPU technologies and ecosystem.
FugakuNEXT will pursue up to a hundredfold increase in application performance over Fugaku, according to RIKEN. Furthermore, through joint Japan-U.S. efforts in developing FugakuNEXT, the project seeks to establish an “AI-HPC platform” as a new global standard, and to accelerate “AI for Science” – the automation and advancement of research processes such as hypothesis generation and validation. The development of this world-class “AI-HPC platform” will enhance globally competitive domestic technologies and strengthen Japan’s strategic indispensability in the information industry and semiconductor technologies, thereby strengthening its industrial competitiveness.
Fugaku delivered approximately a forty-fold increase in hardware performance over its predecessor, the “K computer,” through the adoption of cutting-edge semiconductor technologies and high-performance memory available at the time. Furthermore, through RIKEN’s leadership in enhancing and optimizing applications, performance was accelerated by approximately three-fold, enabling total gains in application performance that exceeded the target of a hundred-fold improvement for certain applications. However, in recent years, the slowdown in performance and power efficiency gains from semiconductor process technology has made it increasingly difficult to achieve significant improvements in hardware performance. Therefore, in addition to advances in hardware, new approaches driven by innovations in software and algorithms are indispensable.
FugakuNEXT will contribute to strengthening Japan’s semiconductor and computing infrastructure, while ensuring national sovereignty in advanced AI technologies and computing platforms. At the same time, it aims to establish a global ecosystem, in both hardware and software, through deployment in the international market. On the hardware side, the system will feature a successor to the general-purpose CPU “FUJITSU-MONAKA”, currently under development by Fujitsu, further evolved into a new CPU tentatively named “FUJITSU- MONAKA-X”, combined with a GPU designed by NVIDIA that emphasizes parallel processing performance and bandwidth as the accelerator. The basic design will explore the adoption of cutting-edge connection technologies between the CPU and GPU, while also considering the incorporation of advanced memory technologies. As a result, FugakuNEXT will target more than a fivefold improvement in hardware performance over Fugaku.
Designed on the basis of advanced technologies from both Japan and the U.S, FugakuNEXT will feature AI-oriented hardware performance exceeding 600 exaFLOPS (EFLOPS) in FP8 precision (sparse), and is expected to become the world’s first “Zetta-scale” system as a supercomputer for HPC. The software, AI models, and applications developed in this project will be made openly available via cloud environments – such as through the “virtual Fugaku” – even prior to the launch of FugakuNEXT system, thereby fostering an ecosystem for high-performance computing infrastructure in the era of AI-simulation convergence.
In the FugakuNEXT project, alongside hardware development, support for the development of advanced applications is also recognized as a critical component of the initiative. RIKEN will promote GPU porting and optimization support as well as the construction of the FugakuNEXT testbed, leveraging available CPU/GPU systems such as the “AI for Science Development Supercomputer”. In particular, for system design and application co-design incorporating CI/CD/CB technologies, RIKEN will work in collaboration with the U.S. Department of Energy (DOE) under the “Project Arrangement for High-Performance Computing and Artificial Intelligence” between Japan’s Ministry of Education, Culture, Sports, Science and Technology (MEXT) and the DOE. This collaboration will promote the development of a continuous performance evaluation environment through the use of the automated benchmarking environment “Benchpark”, and will involve close engagement with the broader application community.
On the software side, RIKEN will lead efforts in close collaboration with universities and research institutions, working together with Fujitsu and NVIDIA to develop advanced environments, such as enabling effective use of AI processing capabilities, preparing libraries that facilitate the acceleration of complex computations through AI, and supporting AI-driven code development and optimization for FugakuNEXT. It will also be important to ensure that commercial applications and a wide variety of open-source software (OSS) run smoothly, and thereby, the software stack developed by RIKEN for Fugaku and the virtual Fugaku will be further advanced and made available for use in FugakuNEXT.
