Japan-based fabless semiconductor company EdgeCortix announced on September 24, 2026, its next-generation scalable AI chiplet platform, "RAIDEN." The platform is designed to support "Physical AI," which requires high real-time performance and power efficiency, such as in robotics, aerospace, and defense.
EdgeCortix Announces Next-Generation AI Chiplet Platform "RAIDEN" Achieving Up to 3.36 PFLOPS
This article is a translation. Read the Japanese original
RAIDEN employs a modular chiplet architecture, featuring a design that scales flexibly according to application needs—from the single-die "X1" (1) to the flagship "X4" model, which integrates 4 dies. The top-tier "RAIDEN X4" achieves up to 3.36 PFLOPS (3,360 TFLOPS) of AI computing performance in FP4 (4-bit floating-point format).
Regarding memory, it features a maximum capacity of 256GB and a bandwidth of 548GB/s, with inter-die communication capable of up to 1.54TB/s bandwidth. Additionally, it provides a chip-to-chip interconnect bandwidth of up to 6.4Tb/s for scaling beyond a single package.
The hardware utilizes the company's latest architecture, "DNA-X." Furthermore, by using the common "MERA" software stack, the system enables migration between different RAIDEN configurations while maintaining existing software assets.
According to EdgeCortix, customers such as Kawasaki Heavy Industries are already adopting RAIDEN-based solutions for next-generation aerospace and defense products. Samples for customers are planned to be available in early 2027, with mass production scheduled to begin in the second half of 2027.
Sources
- 日本発スタートアップのエッジAIチップがNVIDIA超え、AI処理性能は3.36PFLOPS （ITmedia AI+、2026-09-25）
- EdgeCortix