Scholarlink

<p>The optimization of power consumption, area overhead, and accuracy in digital circuits is a critical challenge, particularly in the context of artificial intelligence (AI) applications, which heavily rely on efficient hardware implementations. Among these, multipliers often contribute significantly to power and area demands. This paper presents a novel approximation scheme tailored for neural network applications, leveraging the statistical properties of neural network weights and inputs to optimize multiplier design. Unlike conventional methods, our approach selectively applies approximation to high-value bits using approximate compressors while preserving critical computations. Comprehensive simulations conducted in MATLAB evaluate the accuracy, error profile, and hardware efficiency of the proposed approximate multipliers across neural network applications and digital image processing tasks. The results show that the proposed design achieves substantial reductions in power consumption and area overhead while maintaining accuracy within acceptable bounds, demonstrating its suitability for resource-efficient AI applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient and High Accuracy Approximate Multiplier Design Approach for AI Application

  • 期刊论文

  • 发表日期：

    2025年6月26日

查看全文
  • Faraz Baraati,
  • Milad Tanavardi Nasab,
  • Abdolah Amirany,
  • Kian Jafari,
  • Reza Ghaderi

摘要

The optimization of power consumption, area overhead, and accuracy in digital circuits is a critical challenge, particularly in the context of artificial intelligence (AI) applications, which heavily rely on efficient hardware implementations. Among these, multipliers often contribute significantly to power and area demands. This paper presents a novel approximation scheme tailored for neural network applications, leveraging the statistical properties of neural network weights and inputs to optimize multiplier design. Unlike conventional methods, our approach selectively applies approximation to high-value bits using approximate compressors while preserving critical computations. Comprehensive simulations conducted in MATLAB evaluate the accuracy, error profile, and hardware efficiency of the proposed approximate multipliers across neural network applications and digital image processing tasks. The results show that the proposed design achieves substantial reductions in power consumption and area overhead while maintaining accuracy within acceptable bounds, demonstrating its suitability for resource-efficient AI applications.