Managing Variability in VLSI Circuits.

Over the last two decades, Design for Manufacturing (DFM) has emerged as an essential field within the semiconductor industry. The main objective of DFM is to reduce and, if possible, eliminate variability in integrated circuits (ICs). Numerous techniques for managing variation have emerged throughout IC design: manufacturers design instruments with minute tolerances, process engineers calibrate and characterize a given process throughout its lifetime, and IC designers strive to model and characterize variability within their devices, libraries, and circuits. This dissertation focuses on the last of these three techniques and presents material relevant to managing variability within IC design. Since characterization and modeling are essential to the analysis and reduction of variation in modern-day designs, this dissertation begins by studying various correlation models used within Statistical Static Timing Analysis (SSTA). In the end, the study shows that using complex correlation models does not necessarily result in significant error reduction within SSTA, and that simple models (which only include die-to-die and random variation) can therefore be used to achieve similar accuracy with reduced overhead and run-time. Next, the variation models, themselves, are explored and a new critical dimension (CD) model is proposed which reduces standard deviation error in SSTA by ~3X. Finally, the focus changes from the timing analysis level and moves lower in the design hierarchy to the libraries and devices that comprise the backbone of IC design. The final three chapters study mechanical stress enhancement and discuss how to fully exploit the layout dependencies of mechanically stressed silicon. The first of these three chapters presents an optimization scheme that uses the layout dependencies of stress in conjunction with dual-threshold-voltage (Vth) assignment to decrease leakage power consumption by ~24%. Next, the second of the three chapters proposes a new standard cell library design methodology, called “STEEL.” STEEL provides average delay improvements of 11% over equivalent single-Vth implementations, while consuming 2.5X less leakage than the dual-Vth alternative. Finally, the stress enhanced studies (and this document) are concluded by a new optimization scheme that combines stress enhancement with gate length biasing to achieve 2.9X leakage power savings in IC designs without modifying Vth.Ph.D.Electrical EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/75947/1/btcline_1.pd