Scholarlink

As convolutional neural networks (CNNs) become increasingly adopted in embedded systems, approximate computing has emerged as a promising strategy for improving hardware efficiency. While many studies focus on designing new approximate hardware components, selecting suitable approximate multipliers (AxMs) for CNN accelerators remains a practical challenge. This chapter presents a case study on the use of an error injection method guided by the error rate matrix of approximate computing-based multipliers (AxM), known as error matrix-based error injection (EMEI). This technique aids in evaluating and selecting AxMs for processing elements (PEs) within CNN hardware accelerators. Through a step-by-step application to a MobileNetV2-based CNN model trained on the CIFAR-10 and GTSRB datasets, we demonstrate how EMEI can be employed to balance accuracy and efficiency by integrating AxMs with varying precision levels. The chapter also discusses how this method impacts hardware resource usage and inference performance, offering practical insights into designing efficient CNN accelerators for embedded applications.

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

Performance Optimization of CNN Hardware

  • 章节

  • 发表日期：

    2026年

查看全文
  • Peiyao Sun,
  • Haosen Yu,
  • Tom J. Kazmierski,
  • Basel Halak

摘要

As convolutional neural networks (CNNs) become increasingly adopted in embedded systems, approximate computing has emerged as a promising strategy for improving hardware efficiency. While many studies focus on designing new approximate hardware components, selecting suitable approximate multipliers (AxMs) for CNN accelerators remains a practical challenge. This chapter presents a case study on the use of an error injection method guided by the error rate matrix of approximate computing-based multipliers (AxM), known as error matrix-based error injection (EMEI). This technique aids in evaluating and selecting AxMs for processing elements (PEs) within CNN hardware accelerators. Through a step-by-step application to a MobileNetV2-based CNN model trained on the CIFAR-10 and GTSRB datasets, we demonstrate how EMEI can be employed to balance accuracy and efficiency by integrating AxMs with varying precision levels. The chapter also discusses how this method impacts hardware resource usage and inference performance, offering practical insights into designing efficient CNN accelerators for embedded applications.