A ReRAM-Based Adaptive Binary Neural Network Accelerator for Efficient Inference

ReRAM-based accelerators have demonstrated significant potential in neural network acceleration. Meanwhile, Binary Neural Networks (BNNs), as an extreme form of model compression, greatly optimize hardware resource utilization and power consumption. Consequently, leveraging the combined advantages of BNNs and ReRAM accelerators offers a promising solution for ultra-low power artificial intelligence (AI) hardware. In this paper, we propose a ReRAM-based Adaptive Threshold BNN accelerator (ReAT) design, which adaptively adjusts weight quantization thresholds using higher order statistical information and employs a gradient-based activation quantization strategy to minimize quantization errors. We evaluate our design using a modified NeuroSim simulator. The results show that, compared to conventional multi-bit quantized models, the ReAT-based BNN accelerators achieve a 7.14× improvement in energy efficiency with minimal accuracy loss. This study provides valuable insights into the application of ReRAM accelerators in low-power edge computing and highlights the great potential of BNNs for high-efficiency inference.