Fujitsu has launched its 144-core FUJITSU-MONAKA server processor, bringing 2nm 3D-stacked silicon to AI inference and HPC systems.
Global sales start in November 2026, both as a standalone Arm processor and inside Fujitsu’s new MONAKA Server range. The chip reaches 3.8GHz and supports memory speeds up to 8,800MT/s.
According to Fujitsu, MONAKA has 144 cores made on a 2nm process, with cache and I/O sections built using 5nm silicon. Fujitsu uses 3D stacking and ultra-low-voltage technology to improve power efficiency without wasting leading-edge silicon on every part of the package.
The processor supports SVE2 vector processing and dedicated matrix instructions for AI inference. Fujitsu claims MONAKA can deliver twice the AI inference throughput of rival CPUs, although it has not identified the comparison processors or published the underlying benchmark results.
The chip includes confidential computing through Arm CCA, using hardware encryption to protect applications and data while they are running in memory.
Fujitsu’s initial MONAKA Server range includes 1U and 2U rack systems. The 2U model carries two 144-core processors, 24 RDIMM slots, PCIe Gen6 expansion and GPU support while using conventional air cooling.
The 1U system can carry one or two processors. Air-cooled models run at a 2.1GHz base frequency, while liquid-cooled versions increase that to 2.9GHz.
Fujitsu says its air-cooled 1U system can operate at ambient temperatures up to 40C, while a water-cooled configuration supports coolant temperatures up to 45C. The company claims its cooling technology can cut cooling power use by up to 80 per cent, although it has not published the full test conditions.
The 1U system supports CDI and CXL technology for pooling memory and accelerators beyond the limits of a single server. That should help customers share expensive memory and accelerator resources across larger AI workloads.
Fujitsu offers a separate 2U multi-node MONAKA machine for academia and HPC. It has four nodes with two processors each, giving the chassis eight CPUs and 1,152 cores, backed by 96 RDIMM slots.
MONAKA follows Fujitsu’s A64FX processor used in Japan’s Fugaku supercomputer, but targets a broader mix of AI, cloud, enterprise and HPC workloads. Fujitsu is working with Nvidia to connect MONAKA CPUs to Nvidia GPUs through NVLink Fusion.
Standalone processors will go on sale globally from November 2026, with 1U and 2U MONAKA servers launching in Japan and Europe during the same month. Broader shipments are scheduled to begin progressively from April 2027.
