Over the past decade, leading supercomputers worldwide have rapidly adopted computational accelerators such as GPUs in pursuit of the highest absolute performance and performance per watt. In addition to achieving high performance in floating-point operations (FLOPS) and memory bandwidth (bytes/sec), recent years have seen a dramatic increase in the demand for AI performance (AI-FLOPS), which relies on low-precision matrix operations.
To meet these requirements, the latest GPUs are designed to deliver not only high FLOPS performance for scientific computing but also significantly improved low-precision matrix computation capabilities, ranging from 16-bit to 4-bit arithmetic, through dedicated architectural design and silicon area allocation. High Bandwidth Memory (HBM) technology, which is inherently suited to GPU architectures, continues to advance in support of these performance gains.
Against this backdrop, GPU-based architectures have rapidly become the standard for top-tier systems. In the November 2025 TOP500 ranking, all of the top ten systems except Fugaku are equipped with the latest GPUs. In Japan, however, flagship systems such as the K Computer and Fugaku have traditionally adopted general purpose CPU-based architectures. In aiming to further improve performance per watt and enhance AI capabilities, the post-Fugaku development plan has taken a GPU-centric architecture as its foundation. Under the code name “FugakuNEXT,” RIKEN is advancing development of the next generation ultra-high performance system that integrates GPUs and CPUs.
In computational science where high performance computing forms the backbone application software development is just as critical as hardware innovation. However, domestic application development has largely been optimized for CPU-based architectures, following the lineage of the K Computer and Fugaku. MEXT support programs have likewise not placed GPU-based systems at the center of their activities.
Meanwhile, several national university supercomputing centers have already begun transitioning to GPU systems, and GPU deployment is progressing in advance of the next flagship system. Application developers have started migrating their codes to GPUs in various ways, and early results are emerging.
To strengthen domestic support for GPU-porting efforts, Advanced HPC-AI Research and Development Support Center (HAIRDESC) was established within Research Organization for Information Science and Technology (RIST). Under a 4.5 year plan starting in October 2025, HAIRDESC will support GPU application development through nation-wide framework. This effort involves not only RIST but also the three core institutions, University of Tsukuba, the University of Tokyo and Institute of Science Tokyo, responding to “Next-Generation HPC-AI Development Support Center Formation” program by MEXT. Moreover, six additional national university supercomputing centers, RIKEN Center for Computational Science (R-CCS), and two major GPU vendors, AMD and NVIDIA also join as the partner institutions and companies.
Until now, various centers have independently conducted tutorials and hackathons to promote GPU application development, and each has produced its own educational materials. However, these materials are typically tailored to individual center environments with their own GPU systems and are distributed across different locations.
HAIRDESC will consolidate and organize these resources, providing unified guidance so that application developers at beginner, intermediate, and advanced levels, can efficiently learn GPU development methodologies and prepare for the increasingly GPU-oriented supercomputing landscape, including the upcoming FugakuNEXT.
Moreover, GPUs play an essential role not only in high performance computing but also in advancing AI-driven methodologies such as data science, surrogate modeling, and AI-assisted code generation, collectively referred to as “AI for Science.” To fully utilize GPU capabilities, HAIRDESC will also support research in areas such as mixed-precision computing and low-precision arithmetic techniques that help balance the computational efficiency and numerical accuracy.
The GPU transformation of computational science applications will also facilitate greater international collaboration. By adopting computational platforms common to leading overseas research efforts, particularly those centered in EU and USA, the domestic ecosystem will be better positioned to promote the global use of Japanese applications and to adopt advanced GPU applications from abroad. HAIRDESC will actively support such international dissemination of GPU-enabled applications.
We sincerely ask for your understanding and support for HAIRDESC and the objectives of the Next-Generation HPC-AI Development Support Center Formation initiative. Your collaboration will contribute to the advancement of GPU application development, computational science, AI research, and international cooperation in Japan.
January 2026
Advanced HPC-AI Research and Development Support Center (HAIRDESC)
Director, Taisuke BOKU