Physical AI has a system problem. Machines that perceive, reason and act in the real world cannot scale by adding AI accelerators alone. Compute, memory capacity, memory bandwidth, connectivity, and software must scale together and continue scaling as the AI models and application workloads evolve. Today, EdgeCortix®, a leading Japanese fabless semiconductor company specializing in energy-efficient AI processing, unveiled RAIDEN™, its next-generation scalable AI chiplet platform purpose-built to address that challenge. Through this chiplet platform, RAIDEN carries EdgeCortix’s energy-efficient design into a new performance class, scaling to meet the demands of Physical AI deployed outside centralized data centers.
RAIDEN introduces a unified, energy-efficient architecture that scales from a single compute die to a four-die flagship within one hardware and software platform. At maximum scale, RAIDEN is architected to deliver up to 3.36 PFLOPS (3,360 TFLOPS) of FP4 AI compute, up to 256 GB of memory, 548 GB/s of memory bandwidth and up to 1.54 TB/s of aggregate die-to-die bandwidth, together with configurable power settings and up to 6.4 Tb/s of chip-to-chip scale-out connectivity for scaling beyond a single package.
RAIDEN is entering the market with customer design wins already secured, spanning demanding Physical AI applications from next-generation aerospace and defense systems, robotic platforms, to high-performance edge AI servers.
One Architecture. From One Die to Four.
RAIDEN is designed for Physical AI at the thick edge, an emerging class of high-performance AI systems deployed outside centralized data centers and closer to machines, sensors and operational environments. Rather than creating separate architectures for different performance classes, RAIDEN scales through a modular chiplet architecture: single-die (X1), two-die (X2) and the four-die (X4) flagship.
All three configurations share EdgeCortix’s latest DNA-X® accelerator architecture and MERA™ software stack. Compute, memory capacity, memory bandwidth and inter-die bandwidth scale with the platform, allowing customers to address different performance, power and deployment requirements without moving to an entirely different AI architecture or software environment. This common architecture is designed to allow customers to move across RAIDEN configurations while preserving their software investment and application environment.
Detailed specifications and deployment positioning for the X1 and X2 configurations will be introduced in subsequent phases of the RAIDEN launch program.
Four Dies. One AI System to Run the Entire Workload.
At the top of the platform, the RAIDEN X4 flagship is architected to operate four compute dies as a tightly integrated system rather than as a collection of isolated accelerators.
The RAIDEN X4 flagship configuration enables:
- Up to 3.36 PFLOPS of FP4 AI compute
- Up to 256 GB of memory
- 548 GB/s of memory bandwidth
- Up to 1.54 TB/s of die-to-die (DNA-X-to-DNA-X) bandwidth
- Up to 6.4 Tb/s of chip-to-chip scale-out connectivity
- Configurable power settings for different performance and deployment requirements
- Integrated multi-core host processing per-die and high-speed system connectivity for self-contained designs
As multimodal, generative, and agentic AI models become part of a diverse range of applications across the physical AI landscape, systems must simultaneously accommodate larger models, greater context, multiple concurrent workloads and increasingly demanding real-time constraints, all within practical system power envelopes.
RAIDEN’s scalable chiplet architecture is designed to address this continuum through a common platform, scaling heterogeneous compute, memory and connectivity from a single die to four dies while maintaining a common software environment. For applications that require additional system-level scale, RAIDEN’s high-speed chip-to-chip connectivity provides up to 6.4 Tb/s of scale-out bandwidth, extending the architecture beyond an individual multi-die package.
The architecture helps reduce the system-level compromises that increasingly constrain advanced Physical AI workloads: swapping models and data between internal and external memory, aggressive quantization, AI and non-AI workload partitioning, and the need to divide a single Physical AI pipeline across multiple discrete computing systems.
Purpose-Built for AI That Perceives, Reasons and Acts.
Physical AI requires more than neural-network inference. Central to RAIDEN is the latest DNA-X® architecture, engineered to execute the full workload while scaling the system as a unified platform. It maintains the runtime-reconfigurable advantages of earlier generations while introducing micro-code programmable matrix and vector engines. This innovation extends versatility beyond standard neural-network inference, providing the necessary support for perception, reasoning, application processing and control workloads to operate together within a common scalable platform.
A single MERA software stack provides a common development and deployment environment across the RAIDEN family, allowing models, tools and software investments to carry across different platform configurations. The same stack spans EdgeCortix product generations, so work done on SAKURA-II today carries forward to RAIDEN rather than being rebuilt.
“Physical AI will not be won by simply building a faster accelerator. It requires a platform where compute, memory, bandwidth, connectivity and software scale together – and where the hardware deployed today can run the models customers adopt years from now,” said Dr. Sakyasingha Dasgupta, Founder and CEO of EdgeCortix. “That is what we built RAIDEN to do. One architecture and one software platform scaling from a single die inside an intelligent machine to a four-die flagship for the most demanding Physical AI workloads. Most importantly, customers are already designing around RAIDEN.”
RAIDEN extends EdgeCortix’s vision of Powering AI That Acts™ beyond traditional edge inference into a broader class of Physical AI systems, spanning robotics and autonomous systems, intelligent manufacturing, telecommunications, aerospace and defense, and edge AI infrastructure.
Customers Are Already Designing Around RAIDEN.
RAIDEN is already moving from architecture into multiple customer programs.
Kawasaki Heavy Industries, Ltd., has selected RAIDEN-based solutions for multiple next-generation AI-enabled aerospace and defense products, where computing platforms must operate under demanding system constraints while supporting complex AI workloads that continue evolving throughout long product lifecycles.
Unigen Corporation, a U.S. headquartered enterprise hardware provider, is developing RAIDEN-based server platforms, expanding an existing collaboration with EdgeCortix that today ships SAKURA-II-powered modules and server systems.
Together, these customer engagements demonstrate the breadth of the architecture RAIDEN was created to address: from AI deployed directly within intelligent and autonomous machines to high-performance multi-die systems for on-premise and edge AI infrastructure. Further RAIDEN design wins, application categories, and ecosystem partners will be announced as the program progresses.
Availability and RAIDEN Early Access Program.
Customer sampling is expected to begin in early 2027, with volume production planned for the second half of 2027. The RAIDEN Early Access Program is now open, giving qualified customers and ecosystem partners an opportunity to engage with EdgeCortix during the development program and receive priority consideration for customer samples. The program has already generated strong customer interest, building on the design wins secured to date.
Register your interest in the RAIDEN Early Access Program: https://www.edgecortix.com/en/raiden-early-access-program
Learn more about RAIDEN: https://www.edgecortix.com/en/products/raiden
RAIDEN, EdgeCortix next-generation chiplet solution, is based on results obtained from the project, ‘Research and Development Project of the Enhanced Infrastructures for Post-5G Information and Communication Systems’ (JPNP20017), subsidized by the New Energy and Industrial Technology Development Organization (NEDO).
About EdgeCortix Inc.
EdgeCortix is a fabless semiconductor company delivering energy-efficient AI compute across edge and Physical AI applications, from embedded devices to on-premise systems. Headquartered in Kanagawa, Japan, with operations in the United States and India, the company develops advanced silicon and software platforms built on its patented hardware–software co-design approach. Its Dynamic Neural Accelerator® (DNA) architecture and MERA™ software stack support inference, generative and agentic AI workloads on one runtime-reconfigurable platform that scales from the SAKURA-II single-chip accelerators to the RAIDEN multi-die chiplet platform. EdgeCortix solutions serve robotics and autonomous systems, aerospace and defense, intelligent manufacturing, telecommunications and edge AI infrastructure, markets that demand real-time performance, energy efficiency and adaptability.
For more information, please visit https://www.edgecortix.com/en/
View source version on businesswire.com: https://www.businesswire.com/news/home/20260923260504/en/
Media gallery
