Scholarlink

The main research and engineering challenges for the dependability of edge AI chips stem from the limited computing and energy resources of these devices. The chapter presents techniques for cost-efficient enhancement of reliability in deep learning accelerators. First, we study the potential of homogeneous quantization of CNNs for enabling a negative-memory overhead fault tolerance. Further, we explore a layer vulnerability-aware heterogeneous quantization and dynamic input-aware quantization to adapt precision on the fly during runtime. The released computation resources are allocated for fault tolerance through duplication and triplication.

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

Quantization-Aided Cost-Efficient Reliability of CNN Accelerators for Edge AI

  • 章节

  • 发表日期：

    2026年

查看全文
  • Maksim Jenihhin,
  • Mahdi Taheri,
  • Natalia Cherezova

摘要

The main research and engineering challenges for the dependability of edge AI chips stem from the limited computing and energy resources of these devices. The chapter presents techniques for cost-efficient enhancement of reliability in deep learning accelerators. First, we study the potential of homogeneous quantization of CNNs for enabling a negative-memory overhead fault tolerance. Further, we explore a layer vulnerability-aware heterogeneous quantization and dynamic input-aware quantization to adapt precision on the fly during runtime. The released computation resources are allocated for fault tolerance through duplication and triplication.