Block-Based Compression for Reducing Indexing Cost of DNN Accelerators

Sparse compression is often used in DNN (deep neural network) accelerators to reduce the data traffic of convolutional computation. As the quantization technique greatly reduces the data bit-width, the indexing cost may account for a large proportion of data traffic after sparse compression. In this paper, we point out the similarity of feature maps in different channels. Based on this observation, we propose a block-based compression approach to reduce the indexing cost. Benchmark data show that the proposed approach can reduce 21.1% indexing cost.