Japanese fabless semiconductor company EdgeCortix has announced RAIDEN, a chiplet-based AI accelerator platform it says is designed to scale compute, memory and connectivity together for so-called “physical AI” systems deployed outside centralised data centres.
The company argues that real-world AI applications—machines that perceive, reason and act—face system-level constraints that cannot be addressed by adding accelerators alone, citing the need to scale memory capacity, memory bandwidth, interconnects and software alongside compute as models and workloads evolve.
EdgeCortix said the RAIDEN architecture is intended to scale from a single compute die (X1) to a four-die configuration (X4) within a common hardware and software platform. At maximum scale, the company lists target specifications of up to 3.36 PFLOPS of FP4 AI compute, up to 256GB of memory, 548GB/s of memory bandwidth, up to 1.54TB/s of aggregate die-to-die bandwidth, and up to 6.4Tb/s of chip-to-chip scale-out connectivity.
The platform is positioned for “thick edge” deployments—high-performance systems closer to sensors and operational environments—where power envelopes and real-time requirements can constrain model size and the ability to run multiple workloads concurrently.
EdgeCortix said all configurations will use its DNA-X accelerator architecture and MERA software stack. The company said a common software environment is intended to allow customers to move between RAIDEN configurations without reworking their application environment, and that work done on its current SAKURA-II products would carry forward to RAIDEN.
EdgeCortix said detailed specifications for the X1 and X2 configurations will be released in later phases of the company’s launch program.
As part of the announcement, EdgeCortix said it has secured “design wins” for RAIDEN. It named Kawasaki Heavy Industries as selecting RAIDEN-based solutions for multiple next-generation aerospace and defence products, and Unigen Corporation as developing RAIDEN-based server platforms as part of an existing collaboration that currently ships SAKURA-II-based modules and systems.
EdgeCortix founder and CEO Dr. Sakyasingha Dasgupta said the company built RAIDEN as a platform where compute, memory, bandwidth, connectivity and software scale together, and said customers were already designing around the product.
On timing, EdgeCortix said customer sampling is expected to begin in early 2027, with volume production planned for the second half of 2027. The company said an early access program is available for customers and partners seeking engagement during development.
