BCIM: Efficient Implementation of Binary Neural Network Based on Computation in Memory (2211.06261v1)

Abstract: Applications of Binary Neural Networks (BNNs) are promising for embedded systems with hard constraints on computing power. Contrary to conventional neural networks with the floating-point datatype, BNNs use binarized weights and activations which additionally reduces memory requirements. Memristors, emerging non-volatile memory devices, show great potential as the target implementation platform for BNNs by integrating storage and compute units. The energy and performance improvements are mainly due to 1) accelerating matrix-matrix multiplication as the main kernel for BNNs, 2) diminishing memory bottleneck in von-Neumann architectures, 3) and bringing massive parallelization. However, the efficiency of this hardware highly depends on how the network is mapped and executed on these devices. In this paper, we propose an efficient implementation of XNOR-based BNN to maximize parallelization while using a simple sensing scheme to generate activation values. Besides, a new mapping is introduced to minimize the overhead of data communication between convolution layers mapped to different memristor crossbars. This comes with extensive analytical and simulation-based analysis to evaluate the implication of different design choices considering the accuracy of the network. The results show that our approach achieves up to $10\times$ energy-saving and $100\times$ improvement in latency compared to the state-of-the-art in-memory hardware design.