Fujitsu is teaming with Nvidia to build probably the world’s fastest AI supercomputer ever at 600,000 FP8 Petaflops – so Feyman GPU could well feature
- FugakuNEXT will combine Fujitsu CPUs and Nvidia GPUs in Japan’s next national supercomputer
- The system targets 600EFLOPS FP8 performance with a 100x application performance gain
- Riken, Fujitsu, and Nvidia see the project as defining a new AI-HPC standard
Japan is preparing its next national supercomputer, FugakuNEXT, through a collaboration between Fujitsu, Nvidia and Riken.
The system is planned for operation around 2030 and aims to blend simulation and artificial intelligence into one tightly integrated platform.
For the first time in a Japanese flagship project, GPUs will be used as accelerators. Nvidia will (unsurprisingly) design the GPU infrastructure, Fujitsu will handle CPUs and system integration, and Riken will be involved in the software and algorithm work.
Feynman GPU
The result is expected to be an “AI-HPC platform” designed for science, industry, and AI-driven discovery.
The performance targets for the supercomputer are certainly ambitious. FugakuNEXT is designed to deliver more than 600EFLOPS of FP8 AI performance, which would make it the most powerful AI supercomputer yet announced.
The system is also expected to achieve up to a hundredfold increase in application performance compared with Fugaku, while staying within roughly the same 40MW power budget.
Nvidia’s long-term roadmap points to the Feynman GPU architecture (named after theoretical physicist Richard Feynman) arriving near 2028, so it could well play a role in powering FugakuNEXT.
Fujitsu is developing a successor to its MONAKA CPU for the project, tentatively named MONAKA-X, with more cores, extended SIMD capabilities, and Arm’s matrix computation engine for AI inference.
Coupled with Nvidia’s accelerators, the system is expected to run large simulations alongside demanding AI workloads.
Hardware alone won’t deliver the target gains so the project will also lean on innovations such as surrogate models, mixed-precision arithmetic, and physics-informed neural networks to accelerate performance while also preserving accuracy.
Makoto Gonokami, president of Riken, said, “It is a great honor for Riken to collaborate with Fujitsu and Nvidia in advancing the development of FugakuNEXT. Since ancient times, humankind has built civilizations and advanced societies through the science of computing. Today, the emergence of AI, advanced semiconductors, and quantum computers is bringing about a discontinuous transformation in computational science.”
Ian Buck, vice president at Nvidia, added, “FugakuNEXT will deliver zettascale performance with application speeds nearly 100 times faster – within the same energy footprint as its predecessor – accelerating research, boosting industrial competitiveness, and driving progress for people in Japan and around the world.”
You might also like
FugakuNEXT will combine Fujitsu CPUs and Nvidia GPUs in Japan’s next national supercomputer The system targets 600EFLOPS FP8 performance with a 100x application performance gain Riken, Fujitsu, and Nvidia see the project as defining a new AI-HPC standard Japan is preparing its next national supercomputer, FugakuNEXT, through a collaboration between…
