A Dynamically Reconfigurable Column Streaming-based Convolution Engine for Edge AI Accelerators

Edge AI accelerators are being targeted for numerous emerging applications due to their compact size and low power consumption. However, these applications require reliability, fault-tolerance, and flexibility to overcome the defect rate caused by radiation or manufacturing defects for hard-to-reach applied environments such as space or nuclear power stations. As a result, this paper proposes a dynamically reconfigurable structure for the column streaming-based convolution engine in order to enhance the edge AI accelerator's reliability and flexibility. The high-level description code synthesized from Xilinx Vivado software shows that the design offers more mapping methods in terms of flexibility with 17.15% powerless and [0.43%,3.18%] area overhead as a trade-off compared to the commercial streaming-based CNN accelerator.