Accuracy Configurable FFT Processor Based On Approximate Computing

As the process of integrated circuits continues to approach physical limits,Moore's Law has gradually become ineffective,and it has become more and more difficult to improve circuit performance by reducing the size of transistors.At the same time,with the rise of mobile Internet and Internet of Things technology,the demand for IC design has gradually shifted from mainly high performance to comprehensive requirements for performance,area and power consumption.Approximate circuit simplifies the circuit by modifying the precise circuit logic,which can significantly reduce the area and power consumption of the circuit.This paper proposes an accuracy configurable FFT processor based on approximate calculation which used approximate circuit design.The accuracy configurable FFT processor can work in precise mode and 4 approximate modes to achieve dynamic adjustment of performance,power consumption and accuracy,which can meet the different needs of various scenarios.Firstly,this paper introduces sereval common FFT algorithms and hardware structures.Compares the complexity of different algorithms and the hardware overhead and resource utilization of different hardware structures.Based on this,the FFT algorithm and hardware structure used in this paper are selected.From the perspective of finite wordlength effect and calculation error,the influence on output error of 512-point radix-23 FFT algorithm is analyzed.Using MATLAB to build a512-point radix-23 FFT error simulation platform.After analyzing the simulation results,the fault-tolerant characteristics of the 512-point radix-23 FFT algorithm are obtained.Based on the512-point radix-23 FFT algorithm error simulation,an accuracy configurable FFT processor is designed.A configurable approximate butterfly unit which can truncates the carry chain is proposed at the butterfly node and a bitwidth configurable pipeline Booth multiplier is proposed at the rotation factor multiplication node.Using Design Complier to finish logic synthesize of approximate butterfly unit and bitwidth configurable Booth multiplier,combining the results of logic synthesis and error simulation,four approximation modes of the accuracy configurable FFT processor are obtained.Based on this,the RTL-level design of the accuracy configurable FFT processor is completed,and a function simulation platform is built using Model Sim and Matlab.The function simulation result shows that the design meets expectations.After that,the back-end design of the accuracy configurable FFT processor is completed by using Design Complier and IC Complier with the TSMC 180 nm CMOS technology.The result shows that under the four approximate modes,the FFT processor can achieve performance improvements of0%,4.36%,9.35%,and 14.33%,while reducing power consumption by 9.09%,9.11%,9.12,and15.61%.Compared to the accurate pipeline FFT processor,the accuracy configurable FFT processor can achieve performance improvement up to 17.44% when has the same bitwidth and can reduce the energy consumption up to 47.04% when the difference of SQNR is less than 3d B.Finally,the accuracy configurable FFT processor is applied in image processing.When using image edge extraction,the SSIM of the processed image is greater than 0.81 except the approximate mode 3 which has the largest level of approximate computing.When using image smoothing,the SSIM of all modes is greater than 0.93.The results show that the accuracy configurable FFT processor has good practical value.