Neural Network's Reliability to Permanent Faults: Analyzing the Impact of Performance Optimizations in GPUs

This work evaluates, for the first time, the impact of several performance optimizations on the DNN's reliability when considering permanent faults affecting the underlying GPU. The reliability evaluation relies on software-based fault campaigns deployed on the NVIDIA RTX 3060Ti GPU with the Ampere architecture. The results show that setting up optimization configurations, with reduced usage of the number of registers/threads and computations on the Stream Processors only, can improve the reliability, against permanent faults in the register file, by up to 20% for most of the evaluated DNNs.