Efficient Pre-Silicon Validation and Post-Silicon Tuning of Self-Healing Analog/RF Integrated Circuits

The advent of the nanoscale integrated circuit (IC) technology makes high performance analog and RF circuit increasingly susceptible to large-scale process variations. Process variations, including inter-die variations and local mismatches, significantly impact the parametric yield of analog/RF circuit, and must be properly handled at all levels of design hierarchy. Traditional approaches based on over-design are not sufficient to maintain high parametric yield, due to the large-scale process variations and aggressive design specifications at advanced technology nodes. In this context, the self-healing circuit has emerged as promising methodology to address the variability issue. In this thesis, we propose efficient pre-silicon validation and post-silicon tuning techniques, which are essential for the practical usage of self-healing methodology. One important problem in self-healing methodology is to efficiently and accurately predict the parametric yield in pre-silicon. The main challenge of this problem is caused by multiple circuit states related to tuning knobs. Given that these circuit states closely interact with process variations, they must be properly modeled in order to accurately estimate the parametric yield. Towards this goal, we develop an efficient performance modeling algorithm, referred to Correlated Bayesian Model Fusion (C-BMF) that explores the correlation between circuit states. Next, based on the performance model, the self-healing behavior and the parametric yield can be efficiently and accurately predicted. Another important problem in self-healing circuit is to efficiently perform post-silicon tuning. Towards this goal, indirect performance sensing methodology has recently attracted great attention. In the indirect performance sensing paradigm, the performance of interest (PoI) is not directly measured by on-chip sensor, but is instead accurately predicted from an indirect sensor model. Such indirect sensor model takes a set of other performances as inputs, which are referred to as the performances of measurements (PoMs). The PoMs are selected such that they are highly correlated with PoI and are easy to measure. Due to the process shift associated with manufacturing lines, the indirect sensor model must be calibrated from time to time. For the purpose of reducing the model calibration cost, we propose a Bayesian Model Fusion (BMF) algorithm that reuses the information collected in early stage of manufacturing. We further extend BMF to a Co-learning Bayesian Model Fusion (CL-BMF) algorithm that incorporates not only the early stage information, but also the current stage information that was not considered in the original modeling problem.