Scholarlink

This chapter presents a fault injection framework for logic-in-memory systems. Two simulators are introduced: X-Fault for device-level accuracy and FLIM for high-speed operation. X-Fault maps BNN workloads to crossbars and models controllers, logic families, and device faults. Covered faults include SAF, RDF, DRDF, IRF, SWF, and coupling. FLIM abstracts faults to XNOR operations with precomputed masks and a fault injector inside Larq and TensorFlow. The chapter details application mapping, a crossbar simulator, a memristor model, a mask-based fault generator, and layer-specific injection in Conv2D and dense layers. Two resilience metrics are defined. Quality of Logic measures faulty outputs across gates under one fault model. Impact of Fault measures faulty outputs for one gate across fault models. Case studies compare MAGIC and IMPLY gates and assess BNN resilience across fault types, layers, and models. Results show major speedups for FLIM with modest loss of fidelity. Limitations and future extensions are discussed, including injection during training.

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 in Logic-in-Memory Architectures

  • 章节

  • 发表日期：

    2026年

查看全文
  • Felix Staudigl,
  • Rainer Leupers

摘要

This chapter presents a fault injection framework for logic-in-memory systems. Two simulators are introduced: X-Fault for device-level accuracy and FLIM for high-speed operation. X-Fault maps BNN workloads to crossbars and models controllers, logic families, and device faults. Covered faults include SAF, RDF, DRDF, IRF, SWF, and coupling. FLIM abstracts faults to XNOR operations with precomputed masks and a fault injector inside Larq and TensorFlow. The chapter details application mapping, a crossbar simulator, a memristor model, a mask-based fault generator, and layer-specific injection in Conv2D and dense layers. Two resilience metrics are defined. Quality of Logic measures faulty outputs across gates under one fault model. Impact of Fault measures faulty outputs for one gate across fault models. Case studies compare MAGIC and IMPLY gates and assess BNN resilience across fault types, layers, and models. Results show major speedups for FLIM with modest loss of fidelity. Limitations and future extensions are discussed, including injection during training.