RAEF: A Power Normalized System-level Reliability Analysis and Estimation Framework

System-level reliability estimation is a crucial aspect in reliable design of embedded systems. Recently reported estimation techniques use separate measurements of power consumption and reliability to demonstrate the trade-offs between them. However, we will argue in this paper that such measurements cannot determine comparative reliability of system components with different power consumptions and hence a composite measurement of reliability and power consumption is required. Underpinning this argument, we propose a SystemC based system-level reliability analysis and estimation framework, RAEF, using a novel composite metric, power normalized reliability (PNR), defined as the ratio of reliability and power consumption. We show that PNR based estimation enables insightful reliability analysis of different system components. We evaluate the effectiveness of such estimation in RAEF using a case study of MPEG-2 decoder with four processing cores considering single-event upset (SEU) based soft error model. Using this setup, we analyze and compare PNR based estimation with existing reliability evaluations at different system hierarchies. Furthermore, we demonstrate the advantages of RAEF in assessing design choices highlighting the impact of voltage scaling and architecture allocation.