A CORDIC Based Configurable Activation Function for ANN Applications
Gopal Raut, Shubham Rai, Santosh Kumar Vishvakarma, Akash Kumar
Abstract
An efficient ASIC-based hardware design of activation function (AF) in neural networks faces the challenge of offering functional configurability and limited chip area. Therefore an area-efficient configurable architecture for an AF is imperative to fully harness the parallel processing capacity of an ASIC in contrast to a general-purpose processor. To address this, we propose a configurable AF based on the shift-and-add algorithm, collectively known as Co-ordinate Rotation Digital Computer(CORDIC) algorithm. The proposed versatile configurable activation function is designed using CORDIC architecture and implements both tan hyperbolic and sigmoid function. The derived model is synthesized and verified at 45nm technology. Further, in order to address leakage issues at lower technology nodes, we exploit the power-gating technique for the proposed AF based on CORDIC architecture. Our circuit design is extracted in cadence virtuoso and simulated for all physical parameters. With respect to the state-of-the-art, our design architecture shows improvement by 29% in area, 42% in power dissipation and 20% in latency. The used power gating technique saves 30% static power with minimal area overhead. The Monte-Carlo simulations for process-variations and device-mismatch are performed for both the proposed model and the state-of-the-art to evaluate expectations of functions of randomness in dynamic power variation. The dynamic power variation for our design shows that mean and σ deviation are 180.73μW and 51.7μW respectively which is 60% of the state-of-the-art.