Skip to main content
Open Access Publications from the University of California

UC Berkeley

UC Berkeley Electronic Theses and Dissertations bannerUC Berkeley

Variability Analysis and Yield Optimization in Deep-Submicron Mixed-Signal Circuits


Scaling of CMOS technology into the deep-submicron regime has made superior device performance and high density possible. However, achieving extreme performance is often limited by an increase in variability due to the aggressive shrinking of dimensions. As variability continues to rise, statistical modeling methods like Monte-Carlo and corner simulation that have been extensively used to predict circuit yield in the past, now become insufficient. It is now evident that models can no longer be developed solely by device variation measurements, but they need to exhibit flexibility and tunability to a specific design.

In this work, we address the problem of rising variability and insufficient variability modeling in two ways. Firstly, by characterizing variability in a deeply-scaled technology node, and secondly by developing a methodology for simple, fast model tuning for design specific yield optimization.

Technology characterization is achieved by designing a set of dedicated test structures in a 28nm FDSOI technology. Test structures include both device characterization as well as high-speed comparator characterization, and focus on design-dependent, layout-dependent and topology-dependent sources of variation. Worst-case measured within-die device variation goes up to 11 % while a 46% of current variation is measured across different dies. Layout-dependent systematic effects do appear to be significant in this technology. Several comparator topologies are also measured, showing a direct link between comparator sensitivity and measured offset.

Yield optimization is achieved by model customization to a specific design. A methodology that uses backward propagation of variance and sparse regression techniques is developed in order to achieve this. The methodology is shown to have the ability to tune models to variability structure measurements, decreasing the estimated prediction error from 30% to <4%.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View