EUVIMEDEUVIMED entdecken
Crossref

ELA-FPM: An Energy-Efficient LUT-Based Approximate Floating-Point Multiplier for Neural Network Acceleration

Mahla Salehi Sheikhali Kelayeh, Asma Naseri Rad, Sahand Divsalar, Omid Salehinia, Shaghayegh Vahdat

IEEE Micro · 2026-05 · 46 (3) · 109-119

DOI: 10.1109/mm.2026.3678050 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Micro, 2026-05. Verfasst von Mahla Salehi Sheikhali Kelayeh und 4 weiteren Autor:innen. Die Arbeit nennt 12 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

As artificial intelligence applications grow in practicality and complexity, efficient hardware implementation of these algorithms turns into a critical issue. Floating-point multiplication is one of the most fundamental, yet power-hungry, operations in modern artificial intelligence algorithms. This article proposes a novel floating-point approximate multiplier, called ELA-FPM, which uses a combination of static segmentation of input operands and an energy-efficient lookup table-based method to significantly alleviate the overhead of mantissa multiplication, which is the most power-consuming operation in a floating-point multiplier. The superiority of ELA-FPM manifests itself as delay, power, and area improvements of up to 68.3%, 80.3%, and 92.1%, compared to an exact floating-point multiplier and improvements of up to 65.6%, 70%, and 71%, compared to other state-of-the-art approximate multipliers. Furthermore, ELA-FPM is tested on some of the most famous convolutional neural network models, demonstrating a maximum accuracy drop of 0.67% compared to a model that uses exact multipliers.

Zitieren

APA

Kelayeh, M. S. S., Rad, A. N., Divsalar, S., Salehinia, O., & Vahdat, S. (2026). ELA-FPM: An Energy-Efficient LUT-Based Approximate Floating-Point Multiplier for Neural Network Acceleration. IEEE Micro, 46(3), 109-119. https://doi.org/10.1109/mm.2026.3678050

Vancouver

Kelayeh MSS, Rad AN, Divsalar S, Salehinia O, Vahdat S. ELA-FPM: An Energy-Efficient LUT-Based Approximate Floating-Point Multiplier for Neural Network Acceleration. IEEE Micro. 2026;46(3):109-119. doi:10.1109/mm.2026.3678050

BibTeX
@article{Kelayeh2026,
  author = {Kelayeh, Mahla Salehi Sheikhali and Rad, Asma Naseri and Divsalar, Sahand and Salehinia, Omid and Vahdat, Shaghayegh},
  title = {ELA-FPM: An Energy-Efficient LUT-Based Approximate Floating-Point Multiplier for Neural Network Acceleration},
  journal = {IEEE Micro},
  year = {2026},
  volume = {46},
  number = {3},
  pages = {109--119},
  doi = {10.1109/mm.2026.3678050}
}
Kostenlos starten →