| As IC technologies scale to finer feature sizes, it becomes increasingly difficult to control the relative process variations, particularly due to sub-wavelength photo-lithography. The relatively increasing fluctuations in manufacturing process have introduced unavoidable and significant uncertainty in circuit performance. Hence, modeling and analyzing these random process variations to ensure manufacturability and improve yield has been identified as a top priority for today's IC design problems.; In this thesis, we propose a robust analog design flow for statistical analog/RF optimization. Our proposed design flow starts from a novel performance centering approach to optimally compare different system architectures and propagate the performance specifications from system level down to block level. Unlike many other optimization formulations that typically minimize one cost function by pushing many performance constraints to their boundaries, our performance centering approach adapts the idea from the traditional design centering method. It simultaneously maximizes the design margins for all performance specifications. Taking advantage of this unique optimization formulation, successful system-level designs can be achieved even using simple performance models. Most importantly, the proposed performance centering approach centers each circuit block in a weakly nonlinear analog design space, thereby producing a good starting point for further post-tuning.; Next, we propose a novel robust post-tuning algorithm for optimizing individual circuit blocks and further improving the design accuracy and product yield. Since the first-step performance centering places the initial design in a weakly nonlinear design space, only local optimization is required in the second-step post-tuning, thereby significantly simplifying the post-tuning problem and reducing the computation cost. In addition, the statistical post-tuning is facilitated by a number of novel algorithms, including a projection-based approach for both polynomial and polynomial fitting and an asymptotic probability extraction method to estimate the non-Normal performance distributions.; Based on our knowledge, the proposed optimization flow is the first academic tool that can handle large-size statistical analog optimization and enable the top-down design from system level to block level. |
