MSDF-Based MAC for Energy-Efficient Neural Networks

This article presents an energy-efficient serial multiply-accumulate (MAC) unit based on the most significant digit first (MSDF) approach, specifically aimed at neural networks operating in resource-constrained and low-energy environments. The proposed MAC unit has been integrated into two neural network architectures: a multilayer perceptron (MLP) and a denoising autoencoder. For the MLP, we employ a pre-trained model for handwritten digit classification using the Modified National Institute of Standard and Technology (MNIST) dataset. Hardware synthesis using 45 nm CMOS technology shows that the MSDF-based MLP achieves a favorable trade-off between hardware metrics, such as reduced circuit area and energy per neuron, and reaches a good classification accuracy of 97.4%. In the case of the autoencoder, the proposed MAC unit is utilized in a pre-trained denoising autoencoder for the same dataset. The autoencoder employs MSDF-based mixed precision across layers to reduce computational resources and loads while preserving high-quality image reconstruction. A comprehensive evaluation of image correctness, signal fidelity, and visual quality metrics was conducted to assess the impact of varying precision levels across different layers. The results indicate that the proposed approach achieves an effective denoising performance with balanced hardware requirements.