Innovative “ARTS” Approximate Multiplier: An Effective Step Towards Efficient Hardware for Artificial Intelligence

According to the College of Engineering's Public Relations Office, the new approximate multiplier research project was conducted by the college's researchers as a direct response to the growing challenge of efficient hardware implementation of machine learning and artificial intelligence applications in daily life.

Dr. Shaghayegh Vahdat, a faculty member at the School of Electrical and Computer Engineering and the supervisor of this project, explains the necessity of this structure, stating: "Considering the ever-increasing prevalence of artificial intelligence in all aspects of human life, from mobile phones to medical diagnostic systems, the need for fast, energy-efficient, and compact computational units is more substantial than ever, and multipliers are the beating heart of these applications. The ARTS multiplier was developed to overcome these serious hardware limitations."

“The core idea of ARTS is partitioning the input operands into static segments and then approximately summing the partial products using simplified reduction trees. This approach makes high-volume computations intelligent, simple, and efficient,” added the faculty member from the School of Electrical and Computer Engineering.

Emphasizing the significant improvements in key hardware metrics demonstrated by the results of the project, Dr. Vahdat stated: "The results of this project show a 68.6% reduction in power consumption, a 16.5% reduction in latency, and a 60% reduction in chip area compared to a conventional exact 8-bit multiplier. Compared to other advanced approximate multipliers proposed in the literature, improvements of up to 59.8% in power, 37.2% in latency, and 52.8% in area have been observed."

 “Unlike many approximation schemes that focus only on hardware metrics, ARTS shows excellent hardware-efficiency in real-world applications while maintaining acceptable accuracy.”

She continued, "Used in image multiplication and Sobel edge detection, which are common image processing applications, ARTS achieves up to 91.4% and 28.3% higher image quality based on peak signal-to-noise ratio (PSNR), and up to 52.4% and 20.5% better structural similarity index measure (SSIM) compared to the state-of-the-art (SoTA) approximate multipliers. Also, in deep neural networks (DNN) application,  ARTS  demonstrates over 84.8% higher accuracy compared to other approximate schemes with similar hardware characteristics."

Referring to the future of this technology, Dr. Vahdat emphasized: “ARTS is just the beginning. This architecture has great potential for employment in a wide range of low-power and high-speed systems, from Internet of Things (IoT) edge processors to artificial intelligence accelerators, and can be an effective step towards the localization and optimization of advanced hardware design in the country.”

This research project was completed thanks to the efforts and collaborations of students Mahla Salehi and Sahand Divsalar. This achievement is a successful example of integrating modern approximate computing theories in digital hardware engineering, which can promise a bright path in the smart chip design industry.

I hope this milestone would inspire other students at the School of Electrical and Computer Engineering to achieve even greater accomplishments for the university by working in the field of approximate computing and designing innovative hardware systems."

https://www.sciencedirect.com/science/article/pii/S0167739X25003929