Block-Based Compression and Corresponding Hardware Circuits for Sparse Activations.

Abstract

In a CNN (convolutional neural network) accelerator, to reduce memory traffic and power consumption, there is a need to exploit the sparsity of activation values. Therefore, some research efforts have been paid to skip ineffectual computations (i.e., multiplications by zero). Different from previous works, in this paper, we point out the similarity of activation values: (1) in the same layer of a CNN model, most feature maps are either highly dense or highly sparse; (2) in the same layer of a CNN model, feature maps in different channels are often similar. Based on the two observations, we propose a block-based compression approach, which utilizes both the sparsity and the similarity of activation values to further reduce the data volume. Moreover, we also design an encoder, a decoder and an indexing module to support the proposed approach. The encoder is used to translate output activations into the proposed block-based compression format, while both the decoder and the indexing module are used to align nonzero values for effectual computations. Compared with previous works, benchmark data consistently show that the proposed approach can greatly reduce both memory traffic and power consumption.