Low-Cost and Effective Fault-Tolerance Enhancement Techniques for Emerging Memories-Based Deep Neural Networks
TimeThursday, December 9th2:16pm - 2:40pm PST
Event Type
Research Manuscript
Virtual Programs
Presented In-Person
AI/ML System Design
DescriptionEmerging non-volatile memory (NVM) technologies with better scalability, non-volatility, and good read performance are widely researched, especially for the Deep Neural Networks (DNNs) applications. However, emerging NVMs face reliability concerns due to stuck-at faults, which can severely degrade the accuracy of DNNs. This paper introduces lightweight and simple yet effective intra-block address remapping and weight encoding techniques to enhance the fault-tolerance of DNNs. Experiments using popular DNNs indicate that the proposed techniques can enhance the fault-tolerance of DNNs by up to 300x on Cifar10 dataset and up to 125x on Imagenet dataset.