Energy- and Quality-Efficient Approximate Multipliers for Neural Network and Image Processing Applications

Approximate computing is a new trend that trades off computational accuracy for lower energy dissipation and design complexity in various applications, where high precision is not a critical need. In this paper, energy- and quality- efficient approximate multipliers based on new approximate compressors are proposed. We use NAND gates for generating the complemented partial products, which reduces the number of transistors. Furthermore, new approximate compressors with different accuracy and performance characteristics are designed. Accordingly, three hybrid approximate multipliers offering different trade-offs between accuracy and hardware efficiency are proposed. The proposed designs are simulated using HSPICE with the 7nm FinFET model as a modern technology. Furthermore, the efficacies of the approximate multipliers in the neural network and image processing applications are evaluated using MATLAB. According to the results, the proposed designs provide far better compromises between the quality and energy metrics in comparison with the previous designs and can be considered as efficient alternatives for the exact multipliers in neural network and image processing applications.