AN IMPROVED ELEMENTARY PROCESSING UNIT FOR HIGH-DENSITY CNN-BASED MIXED-SIGNAL MICROPROCESSORS FOR VISION

摘要

This paper presents the architecture of the Elementary Processing Unit ― EPU ― which has been employed to design a CNN-Based 128 x 128 Focal Plane Mixed-Signal Microprocessor for vision. The EPU contains the required building blocks to implement, on chip, vision algorithms based on the execution of linear 3x3 convolution masks, or information propagative CNN templates. Using this EPU, we have designed a prototype, called ACE16k, which contains an array of 128 X 128 EPUs and a completely digital interface, in a standard fully-digital 0.35 μm CMOS technology. The estimation results forecast 300 GOPS, 3.23 GOPS/mm~2 and 100 GOP/J.