An Fpga-Based Dynamic Reconfigurable Cnn Edge Computing Device for Object Classification
Yu Wang,Yibing Liao,Yuxuan Zhao,Chengyu Zhang,Bende Xiao,Jiamei Yang,Hui Wang,Fei Xu,Yifan Gao,Jianbin Zheng
DOI: https://doi.org/10.2139/ssrn.4399163
2023-01-01
Abstract:A reconfigurable CNN edge computing device, based on a field-programmable gate array (FPGA), is proposed to reduce memory burden and required runtime for resource-constrained devices. This system utilizes the pipeline structure of FPGAs to speed up network computation and employs off-chip memory to store network models, thereby avoiding the need for tiling techniques with high intermediate cache requirements. A parallel data scheduling model is also developed to reduce additional memory access costs caused by interactions with off-chip memory. The speed, power consumption, and average intersection over union (IoU) were measured separately to evaluate performance. Results showed the proposed system achieved a detection speed of 16 frames per second (FPS), with a power consumption rate of 0.79 W and an average IoU of 41.2%. This represents a speed increase of 178% and a power consumption reduction of 60%, with a reduction in classification accuracy of only 2.52%, compared with the FPGA-based FitNN implementation.