Impacts of Key Parameters on the Performance of NOR Flash Computing Array based Convolution

In this paper, on the basis of the NOR flash computing array (NFCA) based convolution paradigm we proposed, the impacts of key parameters including interconnect resistance, voltage drift and programming deviation on the performance of NFCA based convolution of different bit number (BN) are carefully discussed. The hardware cost is also evaluated. It is found that the critical premise for precise NFCA based convolution is BN≤2. Moreover, when BN≤4, the convolution paradigm could be applied in deep learning neural networks owing to the high tolerance of computing error, and larger BN would lead to the hardware cost reduction (~BN×) compared with binary NFCA based convolution.