FLoPAD-GRU: A Flexible, Low Power, Accelerated DSP for Gated Recurrent Unit Neural Network

Abstract

Recurrent neural networks (RNNs) are efficient for classification of sequential data such as speech and audio due to their high precision on tasks. However, power efficiency, the required memory capacity and bandwidth requirements make them less suitable for battery powered devices. In this work, we introduce FLoPAD-GRU: a system on a chip (SoC) for efficient processing of gated recurrent unit (GRU) networks, that consists of a digital signal processor (DSP), supplemented with an optimized hardware accelerator, which reduces memory accesses and cost. The system is programmable and scalable, which allows for execution of different network sizes. Synthesized in 28 nm CMOS technology, real-time classification is achieved at 4 MHz, with an energy dissipation of 4.1 pJ/classification, an improvement of 15 × compared to a pure DSP realization. The memory requirements are reduced by 75 %, which results in a silicon area of 0.7 mm2for the entire SoC.