A Fast Injection and Low-Overhead Compensation Method for IR-Drop in RRAM-Based CNN In-Memory Computing

In-memory computing (IMC) based on resistive random access memory (RRAM) has great potential in accelerating neural network computations, especially for convolutional neural networks (CNNs) that dominate computer vision tasks due to their intensive matrix operations. However, wire resistance in large-scale memristor arrays causes serious IR-drop issues, which significantly reduce the hardware inference accuracy of neural networks. Noise-aware training is the most widely used method to solve the IR-drop problem. Nevertheless, IR-drop noise is highly related to the spatial position and sparsity of the array, typically requiring complex modeling. Repeatedly injecting IR-drop noise during training can severely reduce the training speed. To address this, this article proposes an IR-drop noise matrix lookup table (LUT) model through sparsity analysis of the activations and weights. The proposed model facilitates rapid matching of the noise matrix according to array dimensions and sparsity patterns, thereby supporting a high-speed IR-drop inference simulation framework. Furthermore, to specifically address the IR-drop issue during IMC inference for CNN models, this article presents a retraining-free IR-drop noise compensation method based on batch normalization (BN) layers. It requires only a single BN parameter adjustment per layer to restore accuracy after the model parameters are written. The experimental results demonstrate that the proposed LUT model reduces the time per array by 54 × compared to state-of-the-art models. Compared to the uncompensated method, our approach achieves up to a 35.55% increase in accuracy, with the best inference accuracy having only a 0.16% error from the ideal case.