Fujitsu has developed the Fujitsu-Monaka, an ARM-based processor designed for high performance computing (HPC), artificial intelligence (AI), and data centers, featuring innovative 3D many-core architecture and ultra low voltage technology for energy efficiency. The processor supports a wide array of applications with a comprehensive open-source software ecosystem and aims to enhance operational efficiency while contributing to carbon neutrality. Notably, it incorporates confidential computing features to protect user data and is designed for user-friendliness in various computational tasks.

1. Next Arm-based
Processor
FUJITSU-MONAKA
and Its Software
Ecosystem
Fujitsu Limited
© Fujitsu 2024

2. Next Gen. Arm-based Processor “FUJITSU-MONAKA”
Application Performance Performance per Watt
3D many-core architecture Confidential Computing
Fujitsu microarchitecture
High-performance Easy to Use
High Reliability & Security
Energy Efficient
FUJITSU-MONAKA
0%
50%
100%
150%
200%
250%
Competitor X in 2027 MONAKA in 2027
Normalized
Performance(%)
Competitor X in 2027
2x
FUJITSU-MONAKA in 2027
250
200
150
100
50
0
0
50
100
150
200
250
Another ~ 2027 MONAKA in 2027
Normalized
Performance(%)
Competitor X in 2027
2x
FUJITSU-MONAKA in 2027
FUJITSU-MONAKA is the processor developed for HPC, AI, and Data Center
2 © 2024 Fujitsu

3. 3D Microarchitecture
PCIe/
Interconnect
Si Interposer
IO Die
SRAM Die (LLC)
Core Die
SRAM Die (LLC)
Core Die
SRAM Die (LLC)
Core Die
SRAM Die (LLC)
Core Die
DDR5
PCIe/
Interconnect
DDR5
Top View
SRAMDie
Core Die
Si Interposer
SRAM Die
Core Die
Package
Side view
High-performance
FUJITSU-MONAKA
specifications
⚫ Armv9-A Architecture (SVE2*1, CCA*2)
⚫ 144 cores x 2 sockets (288 cores / node)
⚫ Ultra low voltage
⚫ 3D chiplet
- Core die 2nm
- SRAM die/IO die 5nm
⚫ DDR5 (12 channels)
⚫ PCI Express 6.0 (CXL3.0)
⚫ Air cooling
3 © 2024 Fujitsu
*1 SVE2(Scalable Vector Extension version two) is a superset of the
SIMD instruction set SVE and NEON.
*2 CCA(Confidential Compute Architecture) is Arm architecture for
confidential computing.

4. Energy Efficient
⚫Lowering voltage reduces power
consumption.
⚫Our proprietary CAD contributes to
realize stable operation at an ultra low
voltage.
P ∝ C V2 f
C : Capacity
V : Voltage
f : Clock Frequency
*1 Fujitsu estimation
7nm
Semiconductor Technology Node
Power
Consumption
5nm
3nm
2nm
Post 2nm
Ultra low voltage
operation on
FUJITSU-MONAKA
Trend of Semiconductor Power*1
FUJITSU-MONAKA realizes carbon neutrality by ultra low voltage technology.
4 © 2024 Fujitsu
Ultra Low Voltage Technology

5. ⚫ Confidential Computing
⚫ Protecting end-user data in memory by encrypting every VM
⚫ This is expected to be an essential technology in cloud, edge and HPC
5 © 2024 Fujitsu
Hardware Vendor
(Fujitsu)
Cloud Provider
End Users
Existing Processors
Cloud Infrastructure
(Hypervisor)
User
VM
User
VM
User
VM
FUJITSU-MONAKA
(Hardware/Firmware)
Cloud Infrastructure
(Hypervisor)
User
VM
User
VM
User
VM
VM memory encryption protects the user
data from the cloud provider.
VMs are not protected from
the cloud provider.
Existing Cloud
FUJITSU-MONAKA-based
Cloud
No access possible
from the cloud
provider
Processor is the
origin of Root of
Trust
Security Enhancements
High Reliability & Security

6. Easy to Use
AI DC
HPC
MLOps
File System / Storage
Hardware：Arm/FUJITSU-MONAKA
Container
Infra
Virtual
Infra
OpenStack
Ceph Harbor Knative
Istio
Serving
Framework
Lustre
KServe
MLServer MLflow
MinIO
Jupyter
Notebook
GitLab
Service Development
CCA
LLM Milvus
LangServe
Cilium
Kubernetes
Toolchain FFTW3
ArmPL
OpenBLAS
ArmCL UXL
LLVM GCC OpenMP oneTBB Arm Optimized Routines
pypiserver Quetz
Keycloak
PostgreSQL
Ceph CSI
CoCo Veraison
RA-TLS
OS Linux (Major distribution) QEMU/KVM libvirt PAPI/libpfm
containerd/CRI-O/Kata Container
REST API
FastAPI
Triton
PyTorch
TensorFlow ONNX
LangChain llama.cpp
scikit-learn
Flask
6 © 2024 Fujitsu
Users can use OSS-based software stack without modification.
Batch job based
HPC cluster system
Inference and AI model
development platform
Data center workload and
secure computing
Job/system management
Slurm Ansible
Parallelization Analysis tools
Linaro Forge
perf
Warewulf
MPICH
OpenMPI
xCCL
PMIx
UCC/UCX Libfabric
Monitoring
AlertManager
Grafana/Loki
Prometheus
FUJITSU-MONAKA Target Domain
* Software list may be changed.

7. 7 © 2024 Fujitsu
OpenBLAS
Meta
MIT LLVM
CCC*1
30% speed up
(Thread control
improvement)
15% speed up
(Arm SIMD support)
31x speed up ML performance on Arm
33% speed up SWPL*2 and quality improvement
Random Forest Training
Stock
Fujitsu enhances software ecosystem with OSS communities
Fujitsu
Collaboration with OSS Community
Easy to Use
*1 Confidential Computing Consortium
*2 Software Pipeline

8. 8 © 2024 Fujitsu
⚫ Explore use cases with customer to expand Arm ecosystem
⚫ Expand the FUJITSU-MONAKA application areas
Use Case 1
⚫ FUJITSU-MONAKA improves operational efficiency
of the generative AI tasks (e.g. LLM+RAG).
⚫ FUJITSU-MONAKA accelerates the physics simulation
with AI surrogate model.
Accelerating CAE design with PINNs
Further acceleration with FUJITSU-MONAKA
LLM + RAG*1 Use Case 2 AI Surrogate Model
Simulations
Complex process
Slow execution
Problem-specific
PINNs*2
Simple process
High speed inference
High versatility
VS
Lang
Chain
LLM
CPU
Vector DB
CPU GPU
CPU
Use Case Exploration with Customer
Easy to Use
*1 Retrieval-Augmented Generation
*2 Physics Informed Neural Networks

9. ⚫ Fujitsu develops the processor called FUJITSU-MONAKA with 3 architectural
innovative key technologies.
⚫ FUJITSU-MONAKA supports open-source based software stack for 3 domains.
⚫ FUJITSU-MONAKA contributes carbon neutrality by ultra low voltage technology.
⚫ Fujitsu enhances software ecosystem with OSS communities and co-creates use
case with customers.
* This presentation is based on results obtained from a project subsidized by the New Energy and Industrial Technology Development Organization (NEDO).
Conclusion
9 © 2024 Fujitsu
Our Own Microarchitecture 3D many-core Confidential Computing
HPC AI Data Center

10. Thank you