Detection and Diagnosis of Out-of-Specification Failures in Mixed-Signal Circuits

Date

2014-12-03

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Verifying whether a circuit meets its intended specifications, as well as diagnosing the circuits that do not, is indispensable at every stage of integrated circuit design. Otherwise, a significant portion of fabricated circuits could fail or behave correctly only under certain conditions. Shrinking process technologies and increased integration has further complicated this task. This is especially true of mixed-signal circuits, where a slight parametric shift in an analog component can change the output significantly. We are thus rapidly approaching a proverbial wall, where migrating existing circuits to advanced technology nodes and/or designing the next generation circuits may not be possible without suitable verification and debug strategies. Traditional approaches target accuracy and not scalability, limiting their use to high-dimensional systems.

Relaxing the accuracy requirement mitigates the computational cost. Simultaneously, quantifying the level of inaccuracy retains the effectiveness of these metrics. We exercise this accuracy vs. turn-around-time trade-off to deal with multiple mixed-signal problems across both the pre- and post-silicon domains. We first obtain approximate failure probability estimates along with their confidence bands using limited simulation budgets. We then generate ?failure regions? that naturally explain the parametric interactions resulting in predicted failures. These two pre-silicon contributions together enable us to estimate and reduce the failure probability, which we demonstrate on a high-dimensional phase-locked loop test-case.

We leverage this pre-silicon knowledge towards test-set selection and post-silicon debug to alleviate the limited controllability and observability in the post-silicon domain. We select a set of test-points that maximizes the probability of observing failures. We then use post-silicon measurements at these test-points to identify systematic deviations from pre-silicon belief. This is demonstrated using the phase-locked loop test-case, where we boost the number of failures to observable levels and use the obtained measurements to root-cause underlying parametric shifts.

The pre-silicon contributions can also be extended to perform equivalence checking and to help diagnose detected model-mismatches. The resultant calibrated model allows us to apply our work to the system level as well. The equivalence checking and model-mismatch diagnosis is successfully demonstrated using a high-level abstraction model for the phase-locked loop test-case.

Description

Citation