Japan Launches FugakuNEXT A Zetta-Scale Supercomputer Built by Fujitsu and Nvidia
Japan is raising its bar on high-performance computing with the introduction of FugakuNEXT-a hybrid AI-HPC supercomputer with an ambitious goal to achieve zetta-scale performance. The venture, announced in Tokyo, is a major collaboration between RIKEN, Fujitsu, and for the first time Nvidia, combining Japanese CPU technology with American GPU architecture.
Building on the Legacy of Supercomputing
FugakuNEXT is the latest in a line of world-leading Japanese supercomputers, namely, the Earth Simulator (2002), the K computer (2011), and its direct predecessor, Fugaku (2020). The Fugaku, which captured the top global ranks for two years, was used in research responses to the COVID-19 pandemic. The new system is scheduled to be brought into service around 2030 in RIKEN's Kobe campus with a development budget that surpasses 110 billion Yen (roughly $740 million USD).
Making the Paradigm Shift Towards a Hybrid AI-HPC Architecture
Unlike the previous Japanese flagship systems, the architecture specified for FugakuNEXT is based on a GPU-centric design. The architecture will consist of:
- CPU: A new MONAKA-X CPU developed by Fujitsu.
- GPU: High-performance GPUs supplied by Nvidia.
- Interconnect: A high-bandwidth NVLink Fusion fabric co-designed by Nvidia to bridge the CPU and GPU.
This hybrid approach resonates with the current unification trend across traditional HPC simulation workloads with next-generation AI applications onto a single platform.
Performance Targets Planning a 100x Jump
RIKEN set out particularly ambitious performance targets on FugakuNEXT, to achieve a total one-hundred-fold advance in the performance of real-world applications over today's machines. This will require both hardware and software upgrades.
- Hardware Performance: 5x anticipated improvement compared with the original Fugaku.
- Software & Algorithmic Improvements: 20x improvement enabled by optimized mixed-precision computing, among other optimizations.
- Peak Performance: Early targets are in the region of 600 exaFLOPS in FP8 sparse precision.
- Power Efficiency: To operate in the same 40-megawatt power envelope as Fugaku.
More Than Hardware An "AI for Science" Platform
FugakuNEXT will become the ultimate combined platform to hasten the pace of scientific discovery around it. On that platform, aspects of the research process will become automated in order to accelerate progress on the climate modeling, drug development, and resilience to disasters. Key to this will be the complete integration of Nvidia's software stack, including CUDA-X libraries, TensorRT inference, and NeMo for LLM training, placing AI at the core of science.
Strategic National Investment
To Japan, FugakuNEXT is now part of the deep strategy to strengthen domestic semiconductor technologies and claim a leading position in the global AI-HPC landscape. Satoshi Matsuoka, RIKEN's Center for Computational Science director, called it "a major strategic move" in the partnership with Nvidia to build capability within Japan. The project, however, aims not only to succeed Fugaku but also to redefine the very notion of national supercomputers in science and society for the zetta-scale era.
