An Accuracy-and-Efficiency-Configurable Blade-Type Approximate Multiplier With Genetic Algorithm-Based Automatic Training Framework
IEEE Transactions on Circuits and Systems I: Regular Papers · 2026-08 · 73 (8) · 5406-5419
DOI: 10.1109/tcsi.2026.3678032 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems I: Regular Papers, 2026-08. Verfasst von Yuxuan Du und 7 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Macau Science and Technology Development Fund.
Abstract (Originalsprache)
In artificial intelligence and image processing chips, the power consumption of arithmetic units has become a critical concern, motivating approximate techniques for higher efficiency. However, existing researches still face challenges in systematically and automatically exploring efficiency boundary under accuracy constraints. To address this, we propose a configurable blade-type approximate multiplier with a genetic algorithm-based automated training framework. First, a blade-type architecture is proposed for coarse accuracy and efficiency control, where Masks act like a “blade” to slice the partial-product array into multiple regions for localized truncation and compensation. Second, a data-correlation and error-distribution based approximate compressor design method is used to achieve configurable compensation within each region for fine control. Then, a genetic algorithm-based automated framework is designed to train dedicated compressors across all Masks in a multiplier, achieving inter-region error compensation and allowing automatic convergence toward the optimal efficiency -accuracy trade-off under different constraints and applications. Simulated in 28nm CMOS, 8-bit and 16-bit blade-type multipliers achieve up to 31.0% and 37.7% power savings, and 39.8% and 51.6% PDP (Power-delay-product) gains, with only $13.98\times 10^{-4}$ and $14.08\times 10^{-6}$ NMED (Normalized mean error distance), respectively, compared to Design-Compiler baselines. When applied to image processing, the designs deliver up to 65.24 dB PSNR (Peak Signal-to-Noise Ratio) and 60.67% power gain for smoothing, and 57.94 dB PSNR with 62.98% power reduction for sharpening.
Förderung
- National Natural Science Foundation of China
- Macau Science and Technology Development Fund
Zitieren
APA
Du, Y., Zhou, K., Yan, Z., Qian, Z., Chen, Z., Yin, J., Lu, Y., & Shan, W. (2026). An Accuracy-and-Efficiency-Configurable Blade-Type Approximate Multiplier With Genetic Algorithm-Based Automatic Training Framework. IEEE Transactions on Circuits and Systems I: Regular Papers, 73(8), 5406-5419. https://doi.org/10.1109/tcsi.2026.3678032
Vancouver
Du Y, Zhou K, Yan Z, Qian Z, Chen Z, Yin J, et al. An Accuracy-and-Efficiency-Configurable Blade-Type Approximate Multiplier With Genetic Algorithm-Based Automatic Training Framework. IEEE Transactions on Circuits and Systems I: Regular Papers. 2026;73(8):5406-5419. doi:10.1109/tcsi.2026.3678032
BibTeX
@article{Du2026,
author = {Du, Yuxuan and Zhou, Kaize and Yan, Zhihao and Qian, Zhangrui and Chen, Zhuo and Yin, Jun and Lu, Yan and Shan, Weiwei},
title = {An Accuracy-and-Efficiency-Configurable Blade-Type Approximate Multiplier With Genetic Algorithm-Based Automatic Training Framework},
journal = {IEEE Transactions on Circuits and Systems I: Regular Papers},
year = {2026},
volume = {73},
number = {8},
pages = {5406--5419},
doi = {10.1109/tcsi.2026.3678032}
}
EUVIMED