Data-driven Processing Element for Sparse Convolutional Neural Networks

Improving neural network throughput and energy efficiency may be achieved by leveraging their intrinsic or induced data sparsity. This article introduces a dedicated processing element that dynamically reduces the number of computations during inference by exploiting the data relevance without sacrificing accuracy. It exploits the asynchronous paradigm, enabling fine-grained control on the data propagation at run-time. The resulting design consumes less power, even with a low sparsity. The energy saved is directly proportional to the sparsity and can reach up to a factor of five, while the area overhead is negligible. These outcomes are highly encouraging for the application of this method to well-known convolutional neural networks.