Analyzing the Resilience of Convolutional Neural Networks Implemented on GPUs

There have been an extensive use of Convolutional Neural Networks (CNNs) in healthcare applications. Presently, GPUs are the most prominent and dominated DNN accelerators to increase execution speed CNN algorithms improve their performance as well Latency. However, prone soft errors. These errors can impact behaviors GPU dramatically. Thus, generated fault may corrupt data values or logic operations cause errors, such Silent Data Corruption. unfortunately, propagate from physical level (microarchitecture) application (CNN model). This paper analyzes reliability AlexNet model based on two metrics: (1) critical kernel vulnerability (CKV) used identify malfunction light- each kernel, (2) layer (CLV) track light-malfunction through layers. To achieve this, we injected which was popularly applications NVIDIA’s GPU, using SASSIFI injector major evaluator tool. The experiments demonstrate average error percentage that caused models has reduced 3.7% 0.383% by hardening only vulnerable part with overhead 0.2923%. is a high improvement for