Abstract: Memristor crossbar array (MCA) is a natural medium for achieving in-memory computing which supports high energy-efficient neural computing by breaking the “memory wall” and “power wall” of the current von Neumann system. However, the manifestation of the real MCA is affected by many non-ideal factors reducing the inference accuracy of the neural network calculation. In this paper, a hard compensation method of “multiple differential memristor adjustment lines” is proposed to compensate the affections of IR-Drop to the inference precision of MAC based neural network. Through the simulation experiments of VGG16 network and LeNet-5 network under MNIST data set, the proposed hardware compensation method can effectively reduce the precision loss caused by IR-Drop.