Fujitsu’s Monaka CPU will arrive in 2027 with 144 cores, 3D chiplets and a 500W appetite.
According to Tom’s Hardware, Monaka succeeds the A64FX chip used in Japan’s Fugaku supercomputer, which was the first to support Arm SVE. Fujitsu is now targeting the new part at AI and HPC workloads where CPUs still have work to do, even as GPUs hog the limelight.
The chip is based on Fujitsu’s Arm v9.3-A core design and supports Arm SVE2-256-bit for AI and HPC. Fujitsu reckons it can deliver up to twice the AI performance, more than 50 per cent lower TCO and secure in-use data protection through confidential computing.
The interesting bit is the 3D chiplet layout. Monaka puts a 2nm core die on top of a 5nm SRAM and I/O die. Fujitsu says the design cuts total 2nm silicon area by 30 per cent, which helps the cost-performance balance. Each core measures about 1.47 mm2.
The platform supports two sockets per node, giving 288 cores in total. It packs 96 PCIe 6.0 lanes with CXL 3.0, 12-channel DDR5-8000+ memory and support for liquid and air cooling.
The core die uses TSMC’s N2P process, while the SRAM and I/O die use TSMC N5. The last-level cache sits on the 5nm SRAM die beneath the core die and connects via hybrid bonding, which shortens signal paths and reduces latency.
Monaka’s enhanced SVE2 engine includes two 256-bit execution units, two 256-bit load-store units and support for FP8 and INT8MM for AI inferencing. Fujitsu has thrown in three-level TAGE branch prediction, six ALUs, ECC, parity checks and hardware instruction retry.
The high-performance SKU has 144 cores, a 2.9GHz base clock, 500W TDP, 6013 GFLOPs of DGEMM and 96.2 TOPS of INT8 grunt. The efficiency SKU keeps the cores but drops to 2.1 GHz, 350W, 4355 GFLOPs, and 69.7 TOPS.
Fujitsu is now sampling Monaka, with production shipments due in 2027. It is already eyeing Monaka-X for 2029, with SME2 for AI, a slot in FugakuNext and Nvidia NVLink Fusion support, because even CPU makers now need to bring a GPU friendship bracelet.
