Scholarlink

As the complexity of deep learning models (DLMs) increases, particularly Large Language Models (LLMs), high-performance accelerators are required to improve computational efficiency. SRAM-based Field Programmable Gate Arrays (FPGAs) are the most common choice due to their adaptability, parallelism, and high performance per watt. However, their vulnerability to radiation-induced faults raises reliability issues, particularly in safety–critical applications such as aerospace, healthcare, and autonomous systems. Reliable operation in such environments necessitates the resilience of FPGA-accelerated DLMs to these faults. This chapter explores various fault injection (FI) techniques for evaluating the reliability of FPGA-based DLMs, as well as offering insights into advanced fault-tolerance (FT) techniques to improve the reliability and efficiency of deep learning applications based on FPGA.

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

Fault Injection and Tolerance Techniques for Deep Learning Models Deployed on SRAM-Based FPGAs

  • 章节

  • 发表日期：

    2026年

查看全文
  • Mehmoona Gul,
  • Laiq Hasan,
  • Ali Zahir,
  • Anees Ullah

摘要

As the complexity of deep learning models (DLMs) increases, particularly Large Language Models (LLMs), high-performance accelerators are required to improve computational efficiency. SRAM-based Field Programmable Gate Arrays (FPGAs) are the most common choice due to their adaptability, parallelism, and high performance per watt. However, their vulnerability to radiation-induced faults raises reliability issues, particularly in safety–critical applications such as aerospace, healthcare, and autonomous systems. Reliable operation in such environments necessitates the resilience of FPGA-accelerated DLMs to these faults. This chapter explores various fault injection (FI) techniques for evaluating the reliability of FPGA-based DLMs, as well as offering insights into advanced fault-tolerance (FT) techniques to improve the reliability and efficiency of deep learning applications based on FPGA.