Enhanced convolutional neural network accelerators with memory optimization for routing applications

Abstract

Currently, Convolutional Neural Networks (CNN) accelerators find application in various digital domains, each highlighting memory utilization as a significant concern leading to system degradation. In response, our present work focuses on optimizing the memory usage of CNN through a strategic approach. The resulting system is coined as the Memory Optimized Zebra CNN (MOZC). In the initial stage, the CNN accelerator is constructed with optimized features, specifically addressing the network routing function. In this context, our approach draws inspiration from zebras, aiming to identify the shortest path between network nodes. The Field-Programmable-Gate-Arrays (FPGA) are employed for evaluating MOZC performance, considering parameters such as lookup table (LUT), Flip-Flop (FF), memory utilization, power consumption, Digital-Signal Processing (DSP), and Giga-Operations-Per-Second per watt (GOPS/W). Additionally, key parameters like data delivery and Throughput assess routing and data transmission robustness. Video data is utilized to determine routing efficiency, and the achieved highest GOPS/W is recorded as 30.43, marking a substantial improvement over conventional CNN accelerators.