19 February 2025 in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
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.
1 January 2026 in IEEE Transactions on Instrumentation and Measurement
Ultrasonic guided waves (UGWs) are well-suited for the nondestructive inspection (NDT) of adhesive joints as they can enable the transmission and reception of signals across the joint without requiring direct access to the damaged area. Conventional damage indices (DIs), such as the peak of the signal envelope, are limited for Carbon Fiber Reinforced Polymer (CFRP) adhesive joints due to the material's anisotropy and the complexity of UGW propagation. This is especially challenging for disbond detection and sizing. This study proposes an unsupervised approach using autoencoders for automatic disbond characterization in CFRP joints. Angle beam transducers (ABTs) were used in a pitch-catch setup to collect UGW signals under pristine conditions at two excitation frequencies (320 kHz and 650 kHz). The proposed DI is based on the reconstruction error between the original and autoencoder-reconstructed signals. Three architectures, including fully connected, Transformer-based, and CNN-LSTM autoencoders, were evaluated using 5-fold cross-validation and tested on independent scans. Performance was assessed using different metrics. Results show that the CNN-LSTM model consistently achieved the highest detection accuracy, while the Transformer model exhibited the most stable performance across folds. In a controlled temperature and humidity environment, the proposed DI provides superior localization and sizing of disbonds compared to conventional manually extracted signal features.
18 June 2024 in IEEE Access
Online arithmetic, where computations are performed from the most significant digit first, has shown benefits in improving throughput and latency within high-performance computing. This computational mode is particularly beneficial for intensive data processing in sequential operations, enabling streaming computations. One such intensive operation is sparse matrix-vector multiplication (SpMV) which finds extensive application in various engineering fields. SpMVs are often carried out in compressed formats such as coordinate encoding format that inherently follows a sequential approach to execute multiply-and-accumulate (MAC) operations of sparse elements. Therefore, this paper proposes an online MAC unit, through a novel algorithm and architecture. Moreover, a SpMV unit is developed by integrating the proposed online MAC unit. The performance and efficiency of the proposed MAC design are evaluated against state-of-the-art MAC designs, synthesized using the TSMC 45 nm CMOS standard cell library with the Synopsis Design Compiler. Results demonstrate the superiority of the proposed MAC unit in terms of delay, area-delay product, and energy consumption. Besides, the SpMV employing the online MAC is compared with counterpart SpMV designs implemented in both ASIC and FPGA. The proposed SpMV design outperforms conventional designs in various performance metrics, achieving a speed-up of 1.69× and energy reductions of up to 60.4% compared to other SpMV designs for 8-bit precision. Moreover, various sparse matrix datasets with different properties are used to evaluate the SpMV using the online MAC unit, considering metrics such as power, delay, hardware resources, and energy consumption.