Fujitsu used Hot Chips 2026 to reveal new details of FUJITSU-MONAKA, its 144-core Arm server processor set to arrive commercially in 2027. The CPU combines 2nm compute dies with 5nm SRAM and I/O dies, places all of its last-level cache on separate silicon, and will use two 256-bit SVE2 vector units per core. The company is preparing 350W and 500W versions for data centers, artificial intelligence, and high-performance computing (HPC).
Fujitsu Monaka in 30 seconds
- Monaka will have 144 Arm cores and will support two-socket servers with up to 288 cores.
- Fujitsu reserves the 2nm process for the compute dies and uses 5nm for SRAM and I/O.
- All of the last-level cache sits on separate dies located beneath the CPU dies.
- Each core packs two 256-bit SVE2 units, down from the 512-bit SVE used in A64FX.
- There will be air-cooled 350W models and liquid-cooled 500W models, with production slated for 2027.
The August 24 presentation offers a clearer picture of a CPU Fujitsu has been previewing for years. Monaka inherits some of the experience built up with the K and Fugaku supercomputers, but its commercial target is considerably broader: conventional servers, AI infrastructure, data analytics, telecommunications, HPC, and edge systems.
One of the most interesting decisions isn’t the core count, but what sits underneath it.
Fujitsu designed Monaka to use the expensive 2nm manufacturing process only where it delivers the greatest benefit. Less than 30% of the total silicon area uses 2nm, while other functions remain on a 5nm node.
2nm is reserved for the cores
Monaka uses a three-dimensional chiplet architecture.
The design essentially splits into three types of silicon: compute dies, SRAM dies for cache, and a die handling I/O functions.
The cores use a 2nm process, while SRAM and I/O are manufactured on 5nm. Fujitsu stacks the compute dies on top of the SRAM dies using a high-density 3D connection.
The decision has an economic rationale as well as a technical one.
Moving every component of a processor to a more advanced manufacturing node doesn’t necessarily deliver the same benefits across the board. Compute logic can gain a lot from smaller transistors, but SRAM, I/O interfaces, and certain analog circuits don’t scale the same way.
Manufacturing all of it on 2nm would raise costs and consume more of the most advanced wafer capacity.
Fujitsu instead chose to split the functions apart.
| Component | Process | Main function |
|---|---|---|
| Core Die | 2nm | Arm cores and execution |
| SRAM Die | 5nm | Last-level cache |
| I/O Die | 5nm | Memory, PCIe, and interconnect |
| Interposer | Silicon | Communication between the different dies |
The strategy echoes an increasingly visible trend in data center processors: using each manufacturing node only for the parts of the chip that actually need it.
AMD has spent years separating compute and I/O in its EPYC processors. Intel is also adopting multi-tile architectures. Monaka adds another twist by physically placing the cache underneath the compute logic.
The company argues that this layout cuts costs and speeds up the arrival of a 2nm-based processor. That advantage will need to be verified once the product reaches the market.
All the last-level cache sits on another die
The cache layout is probably Monaka’s most striking technical feature.
Rather than integrating the last-level cache alongside the cores, Fujitsu places the entire LLC (Last Level Cache) on separate 5nm SRAM dies.
The compute dies sit directly on top.
This sets Monaka apart from solutions like AMD’s 3D V-Cache. In that case, extra SRAM is added on top of a chip that already has its own cache. Fujitsu instead chose to move the entire last-level cache to a separate layer.
The design also keeps the 2nm silicon concentrated on the logic that benefits most from the advanced process.
There’s also a thermal reason behind the layer ordering.
The compute dies generate the most heat, so they sit at the top of the stack, closer to the cooling system. The SRAM stays underneath.
Fujitsu hasn’t yet published detailed figures on latency between cores located in different parts of the package, one of the questions that will need to be examined once commercial hardware ships.
The architecture, then, looks promising on paper, but independent benchmarks are still missing.
From Fugaku’s 512 bits to two 256-bit SVE2 units
Monaka also introduces a significant change from the A64FX used in Fugaku.
A64FX was one of the most important processors in Fujitsu’s recent history and the first to implement Arm Scalable Vector Extension (SVE). It used 512-bit vector units and HBM2 memory, a combination aimed squarely at HPC.
Monaka adopts SVE2, but narrows each unit to 256 bits.
Each core packs two 256-bit SVE2 units alongside two load/store units of the same width. The processor also adds FP8 and INT8 matrix support aimed at AI inference.
Going from 512 to 256 bits might initially look like a step back.
Fujitsu’s explanation is that Monaka is designed for a broader data center market, and that shrinking the units helps contain each core’s area and improve the cost-performance ratio.
It’s also worth remembering there are two vector units per core.
The decision reflects the difference between designing a CPU specifically for a supercomputer and building a processor meant to compete in general-purpose servers, HPC, and AI simultaneously.
No more HBM: Monaka will use twelve DDR5 channels
There’s another significant break from A64FX.
Monaka won’t use HBM memory integrated in the package. Fujitsu has opted for twelve DDR5 channels, with announced support for DDR5-8000.
The change reduces the available bandwidth compared with an HBM-based architecture, but it simplifies building servers with large amounts of conventional memory and lets Fujitsu target a much bigger market.
Fujitsu estimates around 500 GB/s in STREAM Triad for both versions of the processor, though these are manufacturer estimates that will need to be verified against commercial systems.
The platform also includes PCI Express 6.0 and CXL 3.0, two technologies that matter for the next generation of data centers.
CXL will make it possible to build systems where CPUs, accelerators, and memory expansion devices can share resources more flexibly.
A Monaka server will support up to two sockets, reaching 288 cores per node. Fujitsu is planning Ethernet and InfiniBand connectivity for commercial systems.
Two Monaka variants: 350W air-cooled, 500W liquid-cooled
Fujitsu is preparing at least two main configurations.
The first will have a TDP of 350 watts, a 2.1 GHz base clock, and will support air cooling.
The second will push power up to 500 watts and the base clock up to 2.9 GHz. This version is designed for liquid cooling.
Having both models is especially interesting for data center operators.
Not every facility is set up for direct liquid cooling. Keeping a 144-core CPU within a 350W envelope lets operators deploy Monaka in conventional infrastructure without a full rack overhaul.
The 500W variant will extract more performance wherever there’s enough thermal headroom.
| Feature | Monaka 350W | Monaka 500W |
| Cores | 144 | 144 |
| Base clock | 2.1 GHz | 2.9 GHz |
| TDP | 350W | 500W |
| Cooling | Air | Liquid |
| SVE2 | 2 × 256 bits per core | 2 × 256 bits per core |
| Memory | 12 DDR5 channels | 12 DDR5 channels |
| Production | 2027 | 2027 |
Evaluation samples are reportedly already available to select partners, while volume production remains on track for 2027.
Fujitsu wants to compete in AI too, without turning Monaka into a GPU
Artificial intelligence comes up constantly in Monaka’s presentation, but it’s worth clarifying what Fujitsu is actually aiming for.
Monaka remains a server CPU, not an accelerator directly comparable to an NVIDIA Blackwell or AMD Instinct GPU.
Its strategy is to improve performance for AI workloads that can run efficiently on a CPU.
The FP8 and INT8 matrix extensions are especially relevant for inference, where using lower numerical precision allows more operations and reduces memory requirements.
Fujitsu is also preparing optimized libraries for AI, data analytics, and HPC.
The company claims Monaka could reach up to twice the application performance and roughly twice the performance per watt compared with competing processors expected in 2027, alongside a lower total cost of ownership. These are Fujitsu’s internal estimates, and there are no commercial systems yet to verify them.
Energy efficiency is one of the project’s priorities.
Monaka uses especially low-voltage operating techniques and per-core frequency and voltage control. Fujitsu argues that this approach significantly cuts power consumption while keeping a high core count.
The project is also backed by Japan’s New Energy and Industrial Technology Development Organization (NEDO), under programs aimed at developing more efficient data centers.
A Japanese processor manufactured by TSMC
Fujitsu presents Monaka as a processor developed in Japan, but that doesn’t mean it will be manufactured there.
The advanced dies use TSMC processes.
The information presented at Hot Chips identifies TSMC N2P for the compute dies and TSMC N5 for SRAM and I/O.
The distinction matters at a time when Japan is putting large amounts of public money toward rebuilding domestic capacity for advanced semiconductor manufacturing.
Monaka is Japanese in design and architecture, but its production depends on TSMC’s manufacturing capacity.
Fujitsu is working in parallel on later generations and on technologies that could eventually use manufacturing processes available in Japan.
Monaka, however, has a more immediate mission: reaching the server market in 2027.
Arm keeps gaining ground in the data center
Fujitsu won’t be entering an empty market, either.
AWS has developed several generations of Graviton. Microsoft has Cobalt. Ampere has been selling Arm server processors for years, and NVIDIA is expanding its own CPU strategy around Grace and Vera.
Monaka will have to compete against all of them, and especially against the newer generations of AMD EPYC and Intel Xeon, including Intel’s own Diamond Rapids.
Core count alone won’t be enough anymore.
What will matter to operators is performance per watt, performance per rack, server cost, memory capacity, available software, and how easy it is to migrate applications.
Fujitsu is trying to lower one of the traditional barriers facing alternative architectures through compatibility with Linux distributions such as Red Hat Enterprise Linux, SUSE Linux Enterprise Server, and Ubuntu, along with tools like Slurm, Lustre, OpenMPI, and LLVM.
The company is also working with Arm, Broadcom, AMD, Supermicro, and Linaro around the project and its associated infrastructure.
We’ll need to wait until 2027 to see how all this architecture translates into real-world performance.
But Monaka already points to a trend that will likely matter as much as the 144 cores themselves: advanced processors are no longer being built as a single piece of silicon where everything uses the most expensive node.
Fujitsu reserves 2nm for compute, keeps SRAM and I/O on 5nm, and connects them through a 3D architecture.
In a generation where manufacturing advanced transistors keeps getting more expensive, deciding which parts of a CPU actually need those transistors can end up mattering as much as making them smaller.
Frequently asked questions
How many cores will Fujitsu Monaka have?
FUJITSU-MONAKA will have 144 Arm cores per processor and will support two-socket configurations, with up to 288 cores per server.
What manufacturing process will Monaka use?
The compute dies will use a 2nm process, identified in the technical documentation as TSMC N2P, while the SRAM and I/O dies will use TSMC’s 5nm N5.
When will Fujitsu Monaka reach the market?
Fujitsu is still targeting commercial availability for 2027. Specifications remain subject to change until the final launch.
Is Monaka designed for artificial intelligence?
Yes, although it remains a CPU rather than a dedicated GPU. Fujitsu includes SVE2, FP8 and INT8 matrix extensions, and optimized libraries for AI workloads, alongside HPC and data analytics.
