Fujitsu Monaka CPU Arrives in 2027: What Shoppers and Data Centers Should Know

A New Player in Server CPUs

Fujitsu has unveiled detailed specifications for its Monaka processor at Hot Chips 2026, marking a significant entry into the competitive Arm-based server CPU market. The 144-core chip introduces an unconventional three-tier design that separates computation from memory and power management in ways that could reshape how data centers think about performance and efficiency. With two power-focused models launching in 2027, Monaka represents Japan’s push toward domestically-engineered AI infrastructure.

The timing matters: as cloud providers and enterprises evaluate CPU options for the next generation of systems, understanding what Monaka offers helps shoppers make informed decisions about their infrastructure investments. This isn’t just another processor announcement. It’s a deliberate engineering choice that trades core count for cache efficiency and power optimization.

Three Dies Instead of One

Most modern processors stack everything on a single piece of silicon. Fujitsu chose differently. The Monaka splits into three distinct layers: a 2nm core die where all the computation happens, a 5nm cache die that holds the entire last-level cache, and a 5nm input-output die. These three components bond together using advanced hybrid bonding techniques, with the hottest core die positioned on the cooling side for thermal management.

This stacking strategy serves a practical purpose. By keeping the 2nm process node to just 30 percent of total die area, Fujitsu accelerates its ability to manufacture the chip while pushing less-dense components onto mature 5nm technology where yields are higher and costs are lower. The approach mirrors design decisions in other next-generation Arm server CPUs with multi-die configurations, though Fujitsu’s complete separation of cache to its own die is distinctive.

Voltage regulators also live on the 5nm cache die rather than the compute core. This matters because analog circuits don’t shrink efficiently at 2nm. By placing regulators directly beneath the floating-point execution units, Fujitsu enables fine-grained voltage and frequency scaling on a per-core basis, helping unlock the efficiency claims the company has made.

Narrower Vector Instructions, Practical Focus

Fujitsu’s previous Fugaku supercomputer CPU featured 512-bit vector processing. Monaka cuts that to dual 256-bit SVE2 units per core. This isn’t a step backward, but a deliberate choice for data center workloads where smaller vectors reduce core area and improve cost-per-unit performance. Each 256-bit vector unit pairs with a matching load-store unit, keeping the pipeline balanced.

The redesign adds FP8 and INT8 matrix support specifically for inference tasks, a nod to real-world AI deployment where lower-precision math dominates. This focus on inference over training reflects market demand: most deployed AI workloads spend far more compute cycles on inference than on model training.

Power Consumption and Real-World Performance

Fujitsu offers two SKUs designed for different cooling scenarios. The 350-watt model runs at a 2.1 GHz base clock and requires only air cooling. The 500-watt variant pushes to 2.9 GHz and needs liquid cooling. Both are achieved partly through aggressive undervolting, running cores at roughly 30 percent below standard voltage levels to achieve half the typical power draw. This ultra-low-voltage approach, combined with custom SRAM and proprietary design tools, yields efficiency gains equivalent to a full process node advance.

Performance estimates show the 350W model delivering 4,355 GFLOPS in double-precision math and 69.7 trillion integer operations per second in 8-bit mode. The 500W SKU reaches 6,013 GFLOPS and 96.2 TOPS respectively. Both achieve approximately 500 GB/s of memory bandwidth using 12-channel DDR5 at 8000 MT/s. Fujitsu claims up to double the AI performance and over 50 percent lower total cost of ownership compared to unspecified competitors, though independent testing will be necessary before accepting these figures.

Where Monaka Fits in the Market

By 2027, when Monaka launches, it will occupy a middle position among Arm server options. AWS Graviton5 offers 192 cores on a single 3nm die. Ampere targets 512 cores in its Aurora design. Microsoft’s Cobalt 200 delivers 132 cores with per-core voltage scaling. Custom and specialized CPU designs continue to proliferate, making the processor market increasingly fragmented.

Monaka’s competitive advantage doesn’t rely on core count. Instead, the complete cache-on-die stack and dual-channel DDR5 bandwidth offer compelling advantages for memory-intensive workloads. The 256-bit vector width matches some competitors while exceeding the 128-bit implementations common to hyperscaler designs.

Government Backing and Supply Chain Reality

Japan’s New Energy and Industrial Technology Development Organization subsidizes Monaka as part of a green data center initiative targeting 40 percent energy savings by 2030. Yet the chips will be manufactured at TSMC in Taiwan, not Japan. This gap between domestic design and overseas fabrication reflects current realities in semiconductor supply chains, though Japan continues investing heavily in building domestic 2nm capacity through its Rapidus initiative.

Evaluation samples are available now, with full production volume expected in 2027. Fujitsu has not yet disclosed pricing, and performance metrics remain preliminary estimates pending independent benchmarking upon release.