CONFIGURABLE CNN SOC CO-PROCESSOR ARCHITECTURE

• Main Author: Adiel Wijaya, Joshua
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/46110
• Institution: Institut Teknologi Bandung

Big data analytics are one of the pillars of Industry 4.0, and artificial intelligence (AI) is one of the newest technology on big data processing in industry. This research is aimed to provide solution in AI technology application that could be suitably applied for processing various types of data. Convolutional neural network (CNN) is one of the most utilized AI algorithm in the industry. One of the problems with the CNN algorithm is that it is computationally extensive, requiring specialized equipment for processing. The other problem with the CNN algorithm that it requires to be configured for different data types. This research proposed a configurable CNN architecture design for use in a SoC co-processor. The co-processor is configured and generated by the proposed design tools utilizing folding architecture and multiple processing elements working in parallel. The proposed system utilized a configurable system designer that can automatically generate the Verilog source file that defines a CNN processor that can process various image and kernel sizes. The system designer also able to generate the program code to be run on the SoC platform. C platform and shows that the difference between the processing result and the simulation results are insignificant ( ). The system can reach the processing speed of 72.727 MHz.