The RIKEN Institute of Physical and Chemical Research (RIKEN) recently announced the launch of the "Fugaku Next" supercomputer development project. In addition to collaborating again with Fujitsu, it also announced that NVIDIA will be a key partner. The new system is expected to be operational in Kobe's central district in 2030. The goal is to create a next-generation supercomputer that combines traditional high-performance computing (HPC) and artificial intelligence (AI). Overall application performance is expected to increase 100 times over the current Fugaku, achieving approximately 50 EFLOPS, or Zetta-level computing performance.
From Arm-based HPC to AI HPC
When it opened in 2020, Fugaku became one of the world's fastest supercomputers, powered by the Fujitsu A64FX Arm-based processor, and established Arm-based CPUs as a key player in the HPC market. However, with the recent rise of large-scale AI language models and generative AI applications, the role of HPC is shifting.
RIKEN stated that Fugaku Next will not only maintain traditional simulation and scientific computing capabilities but will also be closely integrated with AI computing, encompassing diverse application scenarios such as vehicle aerodynamics simulation and design, drug development, and quantum simulation. The goal is to achieve a 5-10x performance improvement on traditional HPC while simultaneously challenging Zetta-class computing performance in AI computing. Through this dual-track approach, the company aims to achieve a 100-fold increase in overall application performance.
Fujitsu MONAKA-X with NVIDIA GPU
Fugaku Next will utilize Fujitsu's next-generation Arm instruction set CPU, the "FUJITSU-MONAKA-X," and will utilize NVLink technology for high-speed interconnection with NVIDIA GPUs. This marks the first time a leading Japanese HPC supercomputer has adopted GPU acceleration architecture, signaling the country's entry into the AI HPC era.
Crucially, RIKEN and Fujitsu aren't simply introducing hardware; rather, they aim to leverage the NVIDIA CUDA-X software stack to enable Fugaku Next to directly utilize a variety of science and AI development tools. These include NVIDIA cuQuantum (quantum simulation), RAPIDS (data science), TensorRT (AI inference), and NeMo (large-scale language model training and application). This means Fugaku Next is not only an upgrade to traditional supercomputers, but also a new national-level platform capable of supporting AI model training and industrial applications.
According to the performance plan previously planned by the RIKEN Institute of Physical and Chemical Research, the theoretical FP64 vector computing performance target of Fugaku Next is to reach more than 48 PFLOPS for the CPU part and more than 3.0 EFLOPS for the acceleration part.
Furthermore, in terms of theoretical FP16/BF16 matrix computing performance, the CPU will reach over 1.5 EFLOPS, while the accelerator will reach over 150 EFLOPS. As for theoretical FP8 matrix computing performance, the CPU will reach over 3.0 EFLOPS, while the accelerator is targeting over 300 EFLOPS.
In addition, in terms of main memory capacity, the CPU and acceleration unit are each planned to be more than 10 PiB, and the main memory bandwidth is expected to reach more than 7 PB/s for the CPU part and more than 800 PB/s for the acceleration unit.
In terms of storage system performance and capacity configuration, the first tier will serve as an on-premises file system. Currently, CHFS and other file system solutions are under consideration. The design goals are to complete total memory dumping time within 1 minute, IOPS within 1 second, and capacity at least twice the total memory size.
The second level is positioned as shared storage, with Lustre or DAOS being considered as the file system. The design goal is to keep the memory dump time within 5 minutes, the IOPS to be one-tenth of the first level, and the capacity to be more than 10 times the total memory size.
The research team stated that they will ensure smooth data transfer between Fugaku and Fugaku NEXT, while also addressing unstable I/O access performance caused by insufficient storage node memory. They will also promote hardware and software design that achieves stable performance. Furthermore, they will promote the development of sustainable file system software.
In terms of system networking, Fugaku Next will be comprised of two interconnection methods: a scale-up network for accelerated computing and a scale-out network for node interconnection. The scale-up network can not only interconnect multiple GPUs within a node, but also interconnect GPUs on multiple nodes through switches to form pod computing clusters. The scale-out network is considered to adopt open specifications that support IP or RDMA, and may use an indirect network architecture consisting of multiple layers of switches.
Fugaku NEXT's total power consumption (including computing nodes and storage systems) will be kept below 40MW. Compared to the existing Fugaku system, which consumes approximately 30MW, Fugaku NEXT significantly increases computing performance but also significantly increases power consumption.
The specific design of Fugaku Next will begin in 2026, is expected to be completed in 2028, and will be put into manufacturing, installation and adjustment in 2029, with the goal of being put into operation in 2030.
Standardization and global supply strategy
Unlike the current Fugaku system, which is tailor-made for Japan, RIKEN emphasizes that Fugaku Next will adopt a "co-manufactured with Japan" model. Through standardized design, the same architecture can be extended to other international HPC systems, further enhancing Japan's global influence.
Fujitsu also revealed that MONAKA-X will not only be available for use in Fugaku Next, but will also be available to companies such as Supermicro in the future, demonstrating the commercial potential of this architecture CPU.
Process and GPU specifications are still to be confirmed
Fujitsu has yet to reveal the foundry for the MONAKA-X, revealing only that it will utilize a 1.4nm process. The A64FX is currently manufactured by TSMC, leading to speculation that TSMC remains a potential partner. However, Fujitsu is also involved in the Rapidus chip project in Japan, leaving other potential candidates open.
As for the GPU specifications provided by NVIDIA, they are still undetermined. NVIDIA Vice President of HPC Ian Buck revealed that Fugaku Next may not adopt the current Blackwell or the upcoming Rubin display architecture GPU, but will continue to introduce the most suitable solution as technology evolves. Therefore, considering that there are still five years before the official launch in 2030, it may be possible to adopt the previously announced 2028 release.Feynman (Feynman), or a later display architecture GPU.
