Design And Application Of Approximate Binary Multiplier Based On Evolutionary Algorithm

With the rapid development of technologies such as big data,cloud computing and machine learning,chips are facing significant challenges in energy consumption.Approximate computing,as a new computing paradigm,aims to reduce the energy consumption by relaxing the requirement of accuracy and reducing the complexity of circuits.As a basic computing unit,the multiplier has great influence on the frequency and power consumption of the system because of its high complexity and power consumption.Therefore,the approximate multiplier has high research value.In this thesis,an approximate method for binary multipliers based on the evolutionary algorithm(EA)is proposed,and various approximate multipliers are proposed based on this method.Proposed approximate multipliers are applied to the Fast Fourier Transform(FFT)and image sharpening to evaluate the effectiveness and validity of the proposed approximations in fault-tolerant systems.In proposed method,energy efficient approximate multipliers are designed based on EA.By taking multipliers and compressors as individuals and genes respectively,the process of natural evolution is simulated to explore multipliers with better precision and hardware performance.In addition,based on the characteristics of partial product distribution,an error model is proposed,which reduces the number of test vectors by 50%.The proposed approximation method is applied to unsigned and signed multipliers with operation bit widths of 8×8 and 12×12.Then,proposed multipliers are simulated and synthesized,and the results show that the proposed design has the best trade-off in multiple metrics.Finally,proposed approximate multipliers are applied to FFT and image sharpening,and the hardware performance of the system is significantly improved,with errors within an acceptable range.In conclusion,proposed approximation method achieves significant optimization in both the accuracy and hardware performance of approximate multipliers,and proposed approximate multipliers show great potentials in fault-tolerant systems.