SSPE: A Device-edge SNN Inference Artificial Intelligence Processor in Supporting Smart Computing

Spiking Neural Network (SNN) exhibits significant advantages in terminal devices due to its low power consumption and high computing speed. However, how to further reduce computational overhead and improve computational performance remains an urgent issue that needs to be dealt with. This article introduces an SNN processor called SSPE (Smart SNN Processing Element), which supports dynamic pruning, approximate computation, and fast data format switching. Firstly, SSPE has a Sparse Pulse Pruning Scheme (SPPS) that supports dynamic pruning skipping across different dimensions, further reducing computational complexity; Secondly, SSPE supports an Approximate Computing based on Searching System (ACSS), achieving fast computational processing by storing and matching past calculations; Thirdly, SSPE has a Switching Data Formats Method (SDFM), switching and adjusting between multiple calculation formats, ensuring further savings in computational costs. The evaluation was conducted using a 28 nm CMOS process, and the results showed that the performance of SSPE was superior to state-of-the-art processors.