Japan Unveils Next-Generation Supercomputer FugakuNEXT with Advanced Fujitsu Monaka-X Processors

The RIKEN Institute in Japan has commissioned Fujitsu to develop technology for the next generation of its renowned supercomputer, FugakuNEXT. This initiative aims to enhance Japan's position in high-performance computing (HPC) as Fugaku remains a formidable contender in the global supercomputer rankings.

Fugaku, operational since 2020, has consistently ranked among the world's fastest supercomputers and holds a prominent position in the HPCG (High Performance Conjugate Gradient) benchmark, currently standing at second place. However, the Japanese authorities are pushing for advancements that will solidify their leadership in computational capabilities.

The High Performance Computing Infrastructure (HPCI) committee has tasked Fujitsu with designing the FugakuNEXT's technology, with the primary development phase set to conclude by February 27, 2026. This new supercomputer is expected to leverage the advancements made in Fujitsu's existing A64FX processors, which power the current Fugaku system.

At the core of FugakuNEXT will be the Monaka-X processors, a significant evolution of Fujitsu's Monaka CPU architecture. These processors are designed to work seamlessly with artificial intelligence (AI) accelerators, including NVIDIA's prominent GPU technology, which has recently opened its NVLink interface to other manufacturers.

If Fujitsu can deliver on the promises made regarding the Monaka architecture, FugakuNEXT is poised to reclaim a top-tier ranking among supercomputers. The predecessor, Fugaku, demonstrated exceptional efficiency, not only on benchmark tests like the HP-Linkpack but also in practical applications such as the HPCG benchmark, where it performed strongly in solving complex linear equations.

The Monaka CPU architecture, similar to the A64FX, is based on the widely adopted ARM architecture and is scheduled for market release in 2027. The Monaka-X will utilize the ARMv9-A architecture, coupled with 256-bit SVE2 vector units optimized for both AI and HPC tasks. The design innovations include rack-mounted units with a height of just 2U, capable of housing eight CPUs and a total of 1,152 cores, thanks to direct liquid cooling technology. Air-cooled versions will support up to 288 cores in two configurations.

Fujitsu is making significant changes compared to the A64FX processors, including the integration of traditional DDR5 RDIMM memory rather than the more expensive HBM2 stacked memory. This design choice allows for up to 21 GB of memory per processor core with a total memory capacity of 6 TB, comparable to offerings from AMD's Epyc and Intel's Xeon processors.

To enhance efficiency, the Monaka architecture employs a 3D stacking method that combines four compute dies, each containing 36 cores, with cache and I/O dies. This chiplet approach facilitates cost-effective manufacturing processes while using advanced 2-nanometer technology.

Furthermore, Fujitsu plans to optimize the Monaka processors for ultra-low voltage operation, potentially achieving up to twice the energy efficiency of competing CPU architectures by 2027. The goal is not merely to lower power consumption, but to translate this efficiency into superior application performance.

Notably, the integrated TofuD network interface, which was a feature of the A64FX and enhanced its performance in HPC applications beyond HP-Linkpack, will not be present in the Monaka processors. Instead, Monaka will utilize standard Infiniband and Ethernet for inter-node communication.

Details regarding the Monaka-X remain limited, but it is expected to capitalize on the Monaka architecture's strengths while integrating AI and other accelerator technologies efficiently.