At the SC26 supercomputing conference in Chicago, Fujitsu officially showcased its next-generation 2-nanometer Armv9-A processor, FUJITSU-MONAKA, alongside the commercial Fujitsu MONAKA Server lineup, presenting academic and institutional research facilities with a high-throughput, energy-efficient architecture tailored for sovereign artificial intelligence and high-performance scientific computing.
Fujitsu Unveils FUJITSU-MONAKA Hardware Portfolio at SC26
During the SC26 conference held November 15–20, 2026, at McCormick Place in Chicago, Fujitsu presented live demonstrations of its production-intent FUJITSU-MONAKA processor and rack-scale server hardware at Booth #4619. The presentation marked a critical transition from architectural disclosure to commercial hardware availability, spotlighting systems engineered to address the escalating thermal and power constraints facing academic high-performance computing (HPC) centers.
Alongside standalone server motherboards, Fujitsu exhibited its comprehensive AI-HPC Platform, configured to unify numerical simulation with generative AI acceleration under the initiative “Powering AI for Science.” Fujitsu also detailed its hybrid quantum computing roadmap, demonstrating how MONAKA host nodes interface with superconducting quantum processors to execute hybrid classical-quantum algorithms.
As national laboratories and research universities confront surging electrical infrastructure costs, MONAKA is positioned as an architectural alternative to power-dense accelerator topologies. The processor emphasizes compute density per watt, robust memory bandwidth, and an open software ecosystem designed to prevent vendor lock-in across national research cyberinfrastructures.
3D Chiplet Architecture and Technical Specifications
The FUJITSU-MONAKA processor uses an advanced 3D disaggregated System-in-Package (SiP) design that separates compute cores, cache memory, and I/O controllers onto distinct silicon dies optimized for specific process nodes. By leveraging hybrid copper bonding (HCB) and TSMC System-on-Integrated-Chips (SoIC) technology, Fujitsu achieves high interconnect density and minimal latency between stacked tiers.
At the compute layer, MONAKA incorporates 144 Armv9.3-A cores fabricated on TSMC’s 2-nanometer (N2P) process. In standard dual-socket server configurations, the platform delivers 288 physical cores per node. Rather than building a massive, monolithic 2nm die—vulnerable to steep wafer costs and defect rates—Fujitsu restricts 2nm silicon exclusively to the compute logic. The entire last-level SRAM cache and peripheral I/O subsystems are manufactured on TSMC’s mature 5-nanometer (N5) process node. The 2nm compute dies are bonded face-to-face directly atop the 5nm SRAM dies, ensuring high-cost 2nm silicon accounts for less than 30 percent of total die area while maintaining low-latency cache access.
For vector calculation, Fujitsu re-engineered its execution units. While the preceding A64FX processor used wide 512-bit Scalable Vector Extension (SVE) engines focused on double-precision scientific simulation, MONAKA implements dual 256-bit Scalable Vector Extension 2 (SVE2) pipelines per core. This configuration introduces native support for 8-bit integer matrix multiplication (INT8MM) and 8-bit floating-point (FP8) operations, optimizing the CPU for large language model inference and multi-agent AI execution while maintaining strong FP64 throughput.
| Specification | FUJITSU-MONAKA | Fujitsu A64FX (Fugaku) | Contemporary x86 / Arm Server Baselines |
|---|---|---|---|
| Fabrication Node | 2nm TSMC N2P (Compute) + 5nm TSMC N5 (SRAM/IO) | 7nm TSMC FinFET | 3nm to 5nm Monolithic or 2.5D Chiplet |
| Core Count | 144 Armv9.3-A cores per socket (288 dual-socket) | 48 cores + 4 assistant cores | 64 to 128 cores per socket |
| Vector Engines | Dual 256-bit SVE2 with FP8 / INT8MM | Dual 512-bit SVE (FP64 focused) | AVX-512 / AMX or Dual 128-bit NEON |
| Memory Subsystem | 12-channel DDR5 (MR-DIMM / MCR-DIMM up to 8800 MT/s) | 4-stack HBM2 (32 GB total capacity) | 8 to 12-channel standard DDR5 |
| Expansion & Interconnect | 96 lanes PCIe 6.0 with CXL 3.0 pooling | Tofu Interconnect D | PCIe 5.0 / CXL 2.0 |
| Thermal Design Power | 350W (Air-cooled) / 500W (Liquid-cooled) | ~150W to 170W | 300W to 400W+ |
Memory throughput and capacity represent a primary architectural shift. While A64FX integrated 32 GB of High Bandwidth Memory (HBM2) directly on-package, that fixed memory ceiling constrained modern generative AI workloads. MONAKA transitions to standard 12-channel DDR5 memory with Multiplexed Rank DIMMs (MR-DIMMs), achieving transfer speeds up to 8,800 MT/s and supporting multiple terabytes of memory per node. Interconnect features include 96 lanes of PCIe 6.0 and Compute Express Link (CXL 3.0), enabling dynamic memory pooling across disaggregated compute nodes.
From Fugaku to Sovereign AI: Architectural Evolution
The transition from A64FX to MONAKA reflects a strategic transformation in research computing. When Japan’s Fugaku supercomputer, developed by RIKEN and Fujitsu, led the TOP500 rankings in 2020, it proved that Arm architecture could match or exceed x86 supercomputers in floating-point efficiency. However, Fugaku was engineered before the widespread adoption of transformer models, where parameter counts and KV-cache footprints quickly exceed on-package HBM capacities.
MONAKA responds directly to the sovereign AI movement, wherein universities, research agencies, and regional computing centers seek autonomous compute infrastructure free from proprietary GPU licensing and supply bottlenecks. By enabling high-performance AI inference directly on general-purpose CPUs, MONAKA eliminates the necessity of deploying dedicated accelerator sleds for every inference workload. When estimating institutional compute demands, calculating a reliable compute (FLOPs) estimate allows system architects to model whether SVE2-accelerated CPU nodes can absorb routine embedding, fine-tuning, and inference pipelines at lower operational expense.
Energy Efficiency and Operational Carbon Metrics
Power availability has emerged as the limiting factor for campus data center expansions. Academic facilities constructed decades ago frequently encounter thermal ceilings between 10 kW and 25 kW per rack. Installing high-density GPU clusters often requires 40 kW to 80 kW per rack, necessitating costly chilled-water retrofits.
Fujitsu engineered MONAKA with ultra-low-voltage operating points to suit existing institutional facilities. The 350-watt air-cooled SKU operates at a 2.1 GHz base frequency, permitting drop-in installation into conventional enterprise server racks without facility reconfiguration. For higher-density supercomputing installations, the 500-watt direct liquid-cooled (DLC) SKU operates at a 2.9 GHz base frequency with peak boost capability reaching 3.8 GHz. In direct liquid-cooled facilities, heat recapture systems can route warm water into campus heating networks, significantly reducing the aggregate compute carbon footprint of university computing centers.
Industry and Research Community Assessment
Technical analyses from independent hardware analysts, including evaluations published by Tom’s Hardware and ServeTheHome, emphasized Fujitsu’s packaging execution. Hardware reporters observed that by demonstrating functional 2nm TSMC compute chiplets bonded to 5nm base dies, Fujitsu has validated disaggregated 3D silicon as a viable manufacturing model for enterprise servers, mitigating the extreme wafer costs associated with leading-edge semiconductor lithography. Further analysis from Unite.AI emphasized that the accompanying Fujitsu MONAKA Server lineup provides sovereign infrastructure buyers with a turn-key platform engineered specifically for autonomous, auditable enterprise AI deployment.
Server manufacturers at SC26, including system integrator partners exhibiting liquid-cooled rack platforms, confirmed that the MONAKA platform broadens server choice beyond traditional x86 monopolies. Research computing administrators noted that while discrete GPUs remain dominant for heavy foundation model pre-training, CPU-based execution offers significant advantages in job scheduling elasticity. In mixed-workload university environments, nodes running MONAKA can seamlessly pivot from MPI-based finite element analysis to transformer inference without stranding specialized silicon. Furthermore, deterministic floating-point operations across standardized Armv9 vector pipelines reinforce computational reproducibility across multi-institutional scientific collaborations.
Implications for University Procurement and Research Infrastructure
For university Chief Information Officers, research computing directors, and sponsored research administrators, the arrival of MONAKA introduces concrete considerations across multi-year technology roadmaps:
- Federal Infrastructure Grant Alignment: Programs funded through the NSF CISE (Directorate for Computer and Information Science and Engineering) increasingly mandate life-cycle energy accountability. TCO models that factor in 5-year electrical costs favor architectures featuring high memory capacity and energy-efficient vector execution over high-wattage accelerator nodes.
- Colocated Institutional Data Repositories: MONAKA’s 12-channel DDR5 memory and CXL 3.0 interconnect permit rapid, direct indexing of large-scale datasets hosted within a campus data repository, expediting analytics without requiring high-latency data transfers over external accelerator PCIe buses.
- Open Cyberinfrastructure Standards: Relying on standard Armv9-A architecture ensures compatibility with mainstream Linux distributions, open-source compiler infrastructures, and containerized research pipelines, minimizing the overhead of proprietary accelerator toolchains.
Deployment Roadmap and Commercial Availability
Fujitsu confirmed that sample silicon and early validation systems have entered testing with strategic partners, including national laboratories and supercomputing research consortia. Commercial server deployments are scheduled to ramp through late 2026, with widespread enterprise and university deliveries slated for 2027.
Software enablement is progressing across the open-source scientific computing stack. Upstream Linux kernels, GCC, LLVM, OpenMP runtimes, and optimized mathematical libraries have integrated tuning profiles for Armv9-A and SVE2 matrix extensions. Research computing teams evaluating upcoming hardware refreshes are advised to profile existing simulation and AI workloads against SVE2 instruction targets to evaluate potential throughput gains before procurement specifications are finalized.
Sources
- Fujitsu Global. (2026-11-15). “Fujitsu at SC26: Showcasing FUJITSU-MONAKA Next-Generation 2nm Arm Processor and Sovereign AI Server Portfolio.” https://global.fujitsu/
- SC26 Conference Committee. (2026-11-15). “SC26: The International Conference for High Performance Computing, Networking, Storage, and Analysis.” https://sc26.supercomputing.org/
- Tom’s Hardware. (2026-11-15). “Fujitsu Monaka Architecture Analysis: 144-Core 2nm Compute with 5nm 3D-Stacked SRAM.” https://www.tomshardware.com/
- Unite.AI. (2026-09-14). “Fujitsu Unveils 2nm FUJITSU-MONAKA CPU and Sovereign AI Server.” https://www.unite.ai/fujitsu-unveils-2nm-fujitsu-monaka-cpu-and-sovereign-ai-server/
- ServeTheHome. (2026-11-15). “Inside the Fujitsu MONAKA Processor: Armv9 SVE2 and Disaggregated 3D Chiplets.” https://www.servethehome.com/







