Abstract

Nonlinear Analog and Mixed Signal (AMS) circuits are very complex and expensive to design and verify. Deeper technology scaling has made these designs susceptible to noise and process variations which presents a growing concern due to the degradation in the circuit performances and risks of design failures. In fact, due to process parameters, AMS circuits like phase locked loops may present chaotic behavior that can be confused with noisy behavior. To design and verify circuits, current industrial designs rely heavily on simulation based verification and knowledge based optimization techniques. However, such techniques lack mathematical rigor necessary to catch up with the growing design constraints besides being computationally intractable. Given all aforementioned barriers, new techniques are needed to ensure that circuits are robust and optimized despite process variations and possible chaotic behavior. In this thesis, we develop a methodology for optimization and verification of AMS circuits advancing three frontiers in the variability-aware design flow. The first frontier is a robust circuit sizing methodology wherein a multi-level circuit optimization approach
is proposed. The optimization is conducted in two phases. First, a global sizing phase powered by a regional sensitivity analysis to quickly scout the feasible design space that reduces the optimization search. Second, nominal sizing step based on space mapping of two AMS circuits models at different levels of abstraction is developed for the sake of breaking the re-design loop without performance penalties. The second frontier concerns a dynamics verification scheme of the circuit behavior (i.e., study the chaotic vs. stochastic circuit behavior). It is based on a surrogate generation approach
and a statistical proof by contradiction technique using Gaussian Kernel measure in the state space domain. The last frontier focus on quantitative verification approaches to predict parametric yield for both a single and multiple circuit performance constraints.
The single performance approach is based on a combination of geometrical intertwined reachability analysis and a non-parametric statistical verification scheme. On the other hand, the multiple performances approach involves process parameter reduction, state space based pattern matching, and multiple hypothesis testing procedures. The performance of the proposed methodology is demonstrated on several benchmark analog and mixed signal circuits. The optimization approach greatly improves computational efficiency while locating a comparable/better design point than
other approaches. Moreover, great improvements were achieved using our verification methods with many orders of speedup compared to existing techniques.