EdgeCortix Unveiled Scalable RAIDEN AI Chiplet Platform
The modular architecture aims to address compute and memory bottlenecks in physical AI through early 2027 sampling.
Updated on Sept. 23, 2026 in Semiconductors
Live Poll
Do you believe new AI chip technologies will lead to better performance for physical machines?
EdgeCortix has announced the RAIDEN platform, a scalable AI chiplet system designed to aggregate compute and memory across one to four dies. The technology, which is currently in the early access stage, utilizes a proprietary accelerator architecture and software stack to manage high-bandwidth AI workloads.
Why it matters
Current AI accelerators often struggle to scale memory, compute, and connectivity simultaneously, creating system-level constraints for physical AI applications. This modular approach seeks to mitigate those performance compromises, potentially streamlining how aerospace and server developers partition workloads.
The flagship RAIDEN X4 configuration delivers 3.36 PFLOPS of FP4 compute with 256 GB of memory and 548 GB/s bandwidth. It utilizes 6.4 Tb/s chip-to-chip connectivity to maintain throughput across its modular dies.
The players
EdgeCortix
A fabless semiconductor company focused on energy-efficient AI inference hardware and software stack development.
Kawasaki Heavy Industries
An industrial manufacturer with a diverse portfolio in aerospace, defense, and power systems.
Unigen Corporation
A developer of custom hardware, server platforms, and embedded computing solutions.
The details
The platform scales using a modular chiplet architecture—a design technique where individual functional silicon dies are connected within a single package—that can be configured with one, two, or four dies. A consistent MERA software stack—an abstraction layer used to map AI models onto hardware accelerators—maintains environmental consistency regardless of the specific RAIDEN configuration chosen. This allows for unified workload management despite varying hardware footprints.
Timeline
September 23, 2026: EdgeCortix unveiled the RAIDEN platform.
Early 2027: Customer sampling is expected to begin.
Second half of 2027: Volume production is planned.
The Tech Race
This development follows a trajectory supported by the New Energy and Industrial Technology Development Organization subsidies, which aim to bolster domestic high-performance computing capabilities. The platform competes in the evolving market for scalable, physical-AI-ready inference hardware.
Enterprise customers in aerospace and server development can apply for early access now to begin testing the platform. Wider industry impact remains dependent on the planned volume production rollout in the second half of 2027.
The takeaway
The platform highlights a growing industry pivot toward modular chiplet architectures to bypass the limitations of monolithic AI chips. Potential users should monitor the customer sampling phase in early 2027 for third-party validation of the claimed 3.36 PFLOPS performance metrics.
Further reading
For more on the current state of modular silicon, see our reporting on Semiconductors.
More information
Interested parties can register for the RAIDEN early access program registration on the official site.
Live Poll
Do you believe new AI chip technologies will lead to better performance for physical machines?
