5x Reliability Enhanced 40nm TaOx Approximate-ReRAM with Domain-Specific Computing for Real-time Image Recognition of IoT Edge Devices

摘要

Highly reliable Approximate-ReRAM (A-ReRAM) with Pixel-to-Pixel Data Matching (P2P-DM) and Inter-Pixel error-correcting code (IP-ECC) is proposed to recognize the image accurately by deep neural network (DNN). By specializing for the image recognition applications and modulating the image data based on pixel-to-pixel features and ReRAM error characteristics, data-retention time and endurance of ReRAM increases by 5x and 3.3x, respectively.