Leveraging ReRAM Crossbar for Octave Convolution in Deep Neural Networks

This paper proposes a co-design method that integrates the central processing unit (CPU) and resistive random-access memory (ReRAM) crossbar to enhance energy efficiency of Octave convolution. In this method, low-frequency operations within the Octave framework are executed by the energy-efficient ReRAM crossbar, while high-frequency computations, which are crucial for accuracy, are processed by the CPU and results are combined within the CPU to finalize the classification task. Compared to vanilla convolution, for ResNet-50 trained on the CIFAR-10 dataset, this approach reduces the number of CPU operations by 37% while improving accuracy by 4.7%. Compared to traditional Octave convolution, the proposed approach reduces the number of CPU operations by 17% while maintaining accuracy within 2%. Approximately 18.5% of the total operations is executed by the energy-efficient ReRAM in the proposed approach.