Abstract
Reliability assessment has always been a major concern in the design of computing systems. The results of the assessment highlight and guide enhancements which trigger redesign cycles; thus early and accurate reliability assessment is of profound importance. For the purposes of early reliability analysis, abstract models of the design (which are available in early design stages) are typically used. These models, however, may not be completely accurate compared to the actual final design. Existing literature has not quantified this inaccuracy, through a comparison between Register-Transfer-Level (RTL) and microarchitecture-level reliability assessment on the same commercial microprocessor design. In this paper, we perform reliability assessment using statistical fault-injection on the RTL and Microarchitectural models of the same commercial ARM® Cortex®-A9 processor. The assessment was performed using the same benchmark workloads and equivalent configurations of the hardware structures. The results show that, compared to RTL model, the almost 200x faster microarchitectural model reports an average difference of 0.7 percentile units (10%) on the vulnerability estimation of register file and 3 percentile units (20%) on the vulnerability estimation of L1 data cache.