DE-C3: Dynamic Energy-Aware Compression for Computing-In-Memory-Based Convolutional Neural Network Acceleration

Convolutional neural networks (CNNs) are leveraged in many applications, such as image classification and natural language processing (NLP) tasks. However, the hardware implementation of CNNs not only occupies a considerable amount of memory space but also incurs high computational demands, which can lead to significant energy cost for data transfer. Therefore, an emerging computing-in-memory (CIM)- based architecture has been proposed to alleviate the energy bottleneck of CNNs. Various model compression methods have recently been studied to enhance energy efficiency of the CIM-based accelerator, including quantization and pruning. However, these works are separately discussed with the static inference scenario. In this paper, the proposed DE-C3 can dynamically and jointly design the compression strategy. Besides, it adopts trainable energy-aware thresholds for both quantization and pruning scenarios. Experimental results based on the CIFAR-10 dataset show that our DE-C3 achieves up to 3.4× the energy reduction compared to state-of-the-art works.