Japanese supercomputer: FugakuNEXT with new Fujitsu Monaka-X processors

Five years after its launch, the Japanese supercomputer Fugaku is still in the top ten of the 500 fastest supercomputers in the world and ranks second in the HPCG rankings. But that is no longer enough for those responsible in Japan.

The High Performance Computing Infrastructure (HPCI) committee has now commissioned Fujitsu to design the technology for the next generation of national supercomputers in the form of the FugakuNEXT. The primary development phase is scheduled to last until February 27, 2026.

Fujitsu already builds the 48-core A64FX processors used in the current Top7 supercomputer Fugaku.

FugakuNEXT will be based on the Monaka-X, a further development of Fujitsu's Monaka CPU, which will also work closely with AI accelerators such as GPUs. Specifically, Fujitsu refers to Nvidia's currently dominant AI accelerators and their NVLink interface, which was recently opened to other manufacturers.

If Fujitsu fulfills the promises of the Monaka architecture, FugakuNEXT could once again become a hot candidate for a top position in the supercomputer list. When it was released in 2020, its predecessor was already very efficient. It demonstrated its performance not only on paper and when solving gigantic linear equation systems in the HP linkpack, but also in applications such as the HPCG benchmark (High Performance Conjugate Gradient, currently in second place just behind El Capitan) or in mixed-precision calculations (HPL-MxP, currently in sixth place).

Monaka multi-core in 2 nm technology

Like the A64FX processors, the Monaka CPU is based on the widespread ARM architecture and is set to be launched on the market in 2027 –. It is therefore not only being developed for supercomputers. Specifically, ARMv9-A is used in combination with 256-bit SVE2 vector units, whose SIMD arithmetic units are intended to be equally suitable for AI and HPC applications.

In contrast to the introduction of Monaka, rack inserts are now planned, which combine eight CPUs with a total of 1152 cores in a 2U height thanks to direct liquid cooling. Air-cooled versions come with a maximum of 288 cores spread over two racks.

Fujitsu is changing a few things compared to the A64FX CPUs. For example, Monaka will connect conventional DDR5 RDIMM memory bars and not expensive and not too large HBM2 stack memory. With 12 memory channels and 24 RDIMMs, up to 21 GByte per processor core are possible with a total of 6 TByte memory expansion, which is also offered by current Epyc and Xeon processors from AMD and Intel.

3D stacking combines four compute dies with 36 cores each, cache dies and an I/O die. The chiplet technology makes it possible to combine adapted, cost-effective production processes in the I/O die with 2-nanometer technology with gate-all-around transistors.

Fujitsu also aims to be up to twice as energy efficient as the (CPU) competition in 2027 with special optimizations for operation at particularly low voltages for the Monaka. The benefits of the technology are estimated to be the equivalent of a new level in process technology. This puts the 2 nm processors on a par with sub-2 nm designs. Apparently, however, the aim is not to achieve lower power consumption per se, but to translate this advantage into higher application performance.

In addition, the TofuD network connection integrated into the A64FX packages is history, which was used particularly in HPC applications outside of the HP linkpack. Monaka, like most other processors, uses Infiniband and Ethernet to connect the individual nodes to each other.

Little is known about Monaka-X, on the other hand, except that it is based on Monaka technology and, as described, is intended to efficiently integrate AI and other accelerator technology.

(csp)