Analyzing the impact of soft errors in VGG networks implemented on GPUs. (July 2020)

• Record Type: Journal Article
• Authors: Wei, Jinghe; Ibrahim, Younis; Qian, Siyu; Wang, Haibin; Liu, Guozhu; Yu, Qingkui; Qian, Rong; Shi, Junwei
• Is Part Of: Microelectronics and reliability. Volume 110(2020)
• Publication Date: 2020-07
• Subjects: Convolutional neural networks -- VGGNet -- GPUs -- Safety-critical systems -- Soft errors -- Reliability; Electronic apparatus and appliances -- Reliability -- Periodicals; Miniature electronic equipment -- Periodicals; Appareils électroniques -- Fiabilité -- Périodiques; Équipement électronique miniaturisé -- Périodiques; Electronic apparatus and appliances -- Reliability; Miniature electronic equipment; Periodicals; 621.3815
• DOI: 10.1016/j.microrel.2020.113648
• Languages: English
• ISSNs: 0026-2714

Abstract

Abstract: Convolutional Neural Networks (CNNs) models are widely implemented on Graphic Processing Units (GPUs) due to their computational powers, particularly, in object classification and detection. In safety-critical environments, reliability issues of these models induced by soft errors are a major concern. In this paper, we analyze the reliability of VGG model, a well-known CNN architecture, through two metrics: (1) the so-called Kernels Vulnerability Factor (KVF) used to identify vulnerable kernels and (2) the Layer Vulnerability Factor (LVF) used to determine track fault propagation through layers. As our model runs on an NVIDIA's GPU, our main evaluation tool in this study is the fault injector called SASSIFI. Our results show that Im2col presents the highest vulnerability among all kernels and hardening this kernel significantly reduces silent data corruption rate by 85.67% with 35.63% time penalty. The average of errors causing misclassification may reach up to 19.7% in some injection modes and this may be extremely high and unacceptable. Fault injection mode affects both KVF and LVF because RF mode is performed at architecture level, while IOV and IOA modes are at a different level (i.e., program level). Also, a thorough comparison among VGG, AlexNet, and ResNet indicates that LVF and KVF are determined by CNN network architecture. This is because layers of these networks, particularly convolutional layers, are implemented with the Im2col kernel, the most vulnerable and invoked in these CNN networks. Highlights: Soft error performance of VGG on GPU were studied from kernel and layer perspectives. The Im2col kernel presents the most vulnerability and it is the best candidate to harden with minimum overhead. Comparisons indicates that layer and kernel vulnerabilities are determined by CNN network architecture.