Tokyo-based fabless AI chip designer EdgeCortix has unveiled RAIDEN, a modular AI chiplet platform purpose-built for Physical AI workloads that tops out at 3.36 PFLOPS of FP4 compute in its flagship X4 configuration and is being pitched squarely at robotics, aerospace and defense customers scaling past what off-the-shelf edge accelerators can deliver.
The RAIDEN spec sheet
Announced on 24 September 2026 via HPCwire, RAIDEN X4 delivers up to 3,360 TFLOPS of FP4 AI compute, 256 GB of memory, 548 GB/s of memory bandwidth, 1.54 TB/s of die-to-die bandwidth and up to 6.4 Tb/s of chip-to-chip scale-out connectivity. Two smaller configurations — X2 (dual-die) and X1 (single-die) — share the same DNA-X accelerator architecture and MERA software stack, so a customer can pick a compute point without redoing its software integration.
Not just another matmul unit
EdgeCortix's pitch is that Physical AI needs more than raw matmul throughput. RAIDEN's runtime-reconfigurable matrix and vector engines are designed to handle perception, reasoning and control workloads alongside classic neural-network inference — the specific mix that shows up when a robot is running visuomotor policies, a drone is fusing radar and camera streams, or an aerospace platform is running mission autonomy on-vehicle.
Roadmap and early customers
Customer sampling of RAIDEN begins early 2027 and volume production is scheduled for the second half of 2027. EdgeCortix says its Early Access Program is open now, and CEO Dr. Sakyasingha Dasgupta framed the launch bluntly: "Physical AI will not be won by simply building a faster accelerator. Most importantly, customers are already designing around RAIDEN."
Where it lands competitively
RAIDEN slots into a Physical AI silicon race that has been dominated by NVIDIA's Jetson Thor and Blackwell edge SKUs and, more recently, by Qualcomm's Dragonwing IQ8 line — the chip underneath the newly announced Arduino VENTUNO Q robotics platform — and the AI chiplet approach Qualcomm is pursuing following its PickNik Robotics acquisition. EdgeCortix's differentiators are the chiplet packaging economics, the FP4 peak number and a software layer that already ships across product generations — a bet that Physical AI customers value SDK continuity as much as top-line TFLOPS.
Reporting based on coverage from HPCwire and EdgeCortix's official announcement.