DANNA: A Dimension-Aware Neural Network Accelerator for Unstructured Sparsity

With the development of technology in deep neural network (DNN) pruning, Neural Acceleration Processor can be designed to get more efficiency advantage by exploiting all sparsity in neural network. However, sparse data has irregular coordination, which result that the exist sparse designs have serious storage waste and high latency. In this work, we present a Dimension-Aware Neural Network Acceleration (DANNA) to optimize those problems. Specifically, DANNA employs novel Dimension-first dataflow and custom microarchitecture, which substantially reduce both memory amounts and memory access collision (hash collision). Furthermore, DANNA leverages a fixed channel-share-LUT based on channel reuse characteristics in convolution to replace the traditional tedious coordinator comparison, which saves consumption of logic.