Reliability-energy-performance optimisation in combinational circuits in presence of soft errors

Abstract

The reliability metric has a direct relationship to the amount of value produced by a circuit, similar to the performance metric. With advances in CMOS technology, digital circuits become increasingly more susceptible to soft errors. Therefore, it is imperative to be able to assess and improve the level of reliability of these circuits. A framework for evaluating and improving the reliability of combinational circuits is proposed, and an interplay between the metrics of reliability, energy and performance is explored. Reliability evaluation is divided into two levels of characterisation: stochastic fault model (SFM) of the component library and a design-specific critical vector model (CVM). The SFM captures the properties of components with regard to the interference which causes error. The CVM is derived from a limited number of simulation runs on the specific design at the design time and producing the reliability metric. The idea is to move the high-complexity problem of the stochastic characterisation of components to the generic part of the design process, and to do it just once for a large number of specific designs. The method is demonstrated on a range of circuits with various structures. A three-way trade-off between reliability, energy, and performance has been discovered; this trade-off facilitates optimisations of circuits and their operating conditions. A technique for improving the reliability of a circuit is proposed, based on adding a slow stage at the primary output. Slow stages have the ability to absorb narrow glitches from prior stages, thus reducing the error probability. Such stages, or filters, suppress most of the glitches generated in prior stages and prevent them from arriving at the primary output of the circuit. Two filter solutions have been developed and analysed. The results show a dramatic improvement in reliability at the expense of minor performance and energy penalties. To alleviate the problem of the time-consuming analogue simulations involved in the proposed method, a simplification technique is proposed. This technique exploits the equivalence between the properties of the gates within a path and the equivalence between paths. On the basis of these equivalences, it is possible to reduce the number of simulation runs. The effectiveness of the proposed technique is evaluated by applying it to different circuits with a representative variety of path topologies. The results show a significant decrease in the time taken to estimate reliability at the expense of a minor decrease in the accuracy of estimation. The simplification technique enables the use of the proposed method in applications with complex circuits.