Japan fires up the supercomputing furnace | Types of cpu | Hardware restaurant | Cpu vs gpu vs ram | Turtles AI

FugakuNEXT is Japan’s new supercomputer, born from a collaboration between RIKEN, Fujitsu, and NVIDIA. A next-generation hybrid AI-HPC, it aims to be the first "zetta-scale" system with a MONAKA-X CPU, NVIDIA GPUs, NVLink Fusion, and 40 MW of efficiency.

Key Points:

• Japan’s first flagship supercomputer with integrated NVIDIA GPUs
• Goal: 5x the hardware power and 100x the application performance of Fugaku
• Expected AI performance: Up to 600 exaFLOPS (sparse FP8), within current power limits
• Planned operation: by 2030 at the RIKEN campus in Kobe (Port Island)

On a Tokyo morning, or perhaps evening, with only global time zones setting the pace, a concrete dream emerged: FugakuNEXT, a computing giant that isn’t content to chase Earth, GPUs, and the cloud, but rather envisions the future. It’s not just a hardware update or a facelift: RIKEN, Fujitsu, and NVIDIA have chosen to jointly usher in the dawn of an "AI-HPC" era, with capital letters. The official launch on August 22nd felt like a cross between an institutional ritual and a conference on the wonders of science: on one side, Makoto Gonokami and Satoshi Matsuoka, the historic experts at RIKEN; on the other, Vivek Mahajan, a Fujitsu pedigree holder, and Ian Buck, an NVIDIA representative, were ready to weave a hybrid architecture that we will observe on Earth, in tornadoes, in pharmaceutical patents, and in the aesthetic inventions produced by the artisanal intelligence of simulations.

This machine, under construction amidst server secrets and performance graphs, promises to surpass its predecessor, the Fugaku: 5 times more raw hardware power, but above all, 100 times more agile in real-world projects, thanks to software refinements and physics algorithms with so-called "mixed precision."

Behind the scenes, beats a Made-in-Japan heart, cooled by logic and innovation: the fires of the MONAKA-X CPU (a product of the Fujitsu roadmap, on 2nm technology, with 3D packaging and low voltages), combined, via NVLink Fusion silicon, with an NVIDIA GPU—a first for a Japanese flagship machine. The stated goal? A peak AI performance of over 600 exaFLOPS in FP8 sparse precision requires only 40 MW of electrical power, the same powering the current Fugaku.

It’s interesting how the rules of the game have changed: where monolithic CPUs once reigned supreme, now the symphonic CPU-GPU duet takes center stage; Where speed was once the sole focus, today it is colored by sustainability, scalability, and automation. Thinking of workloads that generate hypotheses, write code, and simulate experiments autonomously is almost science fiction, if it weren’t already a film already partly written: the machine will handle surrogate models, physics-informed neural networks, aesthetic optimization in production, and atmospheric simulation for earthquakes and hurricanes.

The investments? Over 110 billion yen (approximately $740 million), roughly the budget of a technological blockbuster. The setting? The RIKEN campus in Kobe, on Port Island, where, however, the curtain will rise only around 2030.

Meanwhile, the global HPC landscape remains crowded: Fugaku has slipped to seventh place in the TOP500 ranking, overtaken by American and European giants, while FugakuNEXT promises not only to regain the top spot, but to redefine its boundaries. According to experts, the system will need to churn out software innovations (continuous benchmarking, optimization, hybrid algorithms, even QC integration) that reflect the ambition of making "AI for science" a universal language.

In short, Japan isn’t just chasing computing power: it has developed a national superbrain, co-developed with foreign partners, driven by logic, advanced silicon, energy efficiency, and advanced AI, and ready, perhaps, to inspire other countries. Nothing more to add, but the stage is now set.