An end-to-end RNS CNN Accelerator

This work presents a Residue Numbering System (RNS)-based Convolutional Neural Network (CNN) accelerator. The proposed system is fully RNS-based, requiring no intermediate conversions to a binary representation. RNS operation overhead is minimized by designing the architecture in such a way that the usage of the non-trivial RNS operations is amortized over a large number of MAC operations. This allows to exploit their periodic usage and further reduce power consumption through clock-gating. Implementation results on a 22-nm process, show that RNS can not only increase the maximum achievable frequency of the arithmetic circuits, but also results in 58% more energy-efficient processing, compared to the traditional binary counterparts. Compared to the state-of-the-art, RNS-based CNN accelerator, the proposed architecture is shown to be 8.5× more energy efficient, with an average energy efficiency of 1.91 TOPS/W.