Nonlinear Function Computation Based on Cordic with Stochastic Bitstreams

Stochastic computing (SC) is a computational paradigm that employs stochastic bitstreams to represent data. Although SC has demonstrated success in numerous tasks, the computation of functions with classical stochastic bitstreams remains severely affected by low accuracy and high latency. Consequently, computing complex tasks, such as the computation of nonlinear activation functions of neural networks, is a significant challenge. To address this issue, we propose the CORDIC-SC algorithm for low-latency and highprecision nonlinear function computation and design a logic circuit unit for bitstream format input. The experimental result shows that the proposed unipolar method is 7.7 × more energy efficient and 10.2 × more area efficient than the existing SC method while maintaining the same degree of accuracy. In addition, the proposed bipolar method is capable of reducing the error by 99.995% while maintaining the same latency.