Products

Edge AI chip series

3rd generation [chichibu] (2024~)

Chichibu is designed for full-scale mass production. It features an ARM system CPU along with proprietary engines such as DNN, RISC-V, GPU, GeMM, image processing, digital signal processing, FFT, and more. Additionally, it incorporates H.264, built-in memory, a secure engine, and an ISP, enabling diverse information processing at the edge.

The performance of this chichibu has been recognized, and our company is currently on business development.

Overview

  • Small, high-speed, and low power consumption, enabling fanless operation.
  • A processor that can be used in embedded products requiring waterproof and dustproof capabilities.
  • This heterogeneous architecture not only features our proprietary neural processing engine, but also includes hardware engines needed for image and audio processing.

Key Features

  • In addition to deep learning inference, it also supports a variety of computer vision image processing and audio signal processing tasks.
  • Its hardware engine group, driven by software-controlled parameters, ensures high efficiency and flexibility.
  • With AI performance of 6.5 TOPS at 2 W (3.25 TOPS/W), it offers highly efficient power performance.
  • Equipped with an ARM Cortex A5 quad-core and various I/O IP blocks, it enables a one-chip solution.
  • It can also be used as a co-processor alongside existing processors. Moreover, a DRAM-less system using only on-chip memory (8 MB cache) is possible, making the solution compact and low-cost.

Examples of Possible Functions

[Self-Position Estimation]

  • Visual/LiDAR SLAM, distance measurement

 

[AI Inference]

  • YOLO, OpenPOSE, fingerprint authentication, palmprint authentication, face detection, etc.

 

[Image Processing (OpenCV library)]

  • Correction: Lens distortion correction, perspective correction, arbitrary-coefficient filtering, bilateral filtering, non-linear filtering, distance filtering
  • Recognition: SAD, SSD
  • Transformation: Morphological operations, thinning, blending, Hough transform
  • Feature Extraction: FAST, SIFT, Harris corner detection, labeling, perimeter/area calculation, FFT, etc.

 

[Audio Processing]

  • Frequency conversion, FIR filtering, modulation, rate conversion, mixing, etc.

2nd generation [beppu] (2021~)

Beppu was developed as an evaluation platform for designing mass-production chips.
The camera module equipped with beppu has successfully operated YOLOv3 (2–3 fps) and OpenPose on battery power.

1st generation [arima] (2019~)

Arima A was developed under a project*¹ by NEDO*². This project has achieved more than 10 times higher power efficiency in AI recognition and image processing compared to conventional general-purpose GPUs.

*1 Joint Project between Socionext, Toyota Industries Corporation, and ArchiTek
*2 NEDO : New Energy and Industrial Technology  Development Organization