Evolving Variability Tolerant Logic

Intrinsic variability occurs between individual MOSFET transistors caused by atomic-scale differences in the construction of devices. The impact of this variability will become a major issue in future circuit design as the devices scale below 50nm. In this thesis, the background to the causes and effects of intrinsic variability, in particular that of random dopant placement and line-edge roughness, is discuss. A system is developed which uses a genetic algorithm to attempt to optimise the dimensions of transistors within standard-cell libraries, with the aim of improving performance and reducing the impact of intrinsic variability in terms of the effect on circuit delay and power consumption. The genetic algorithm uses a multi-objective fitness function to allow a number of circuit characteristics to be considered in the evolution process. The system is tested using different standard-cell libraries from open-source and commercial providers, with developments and alterations to the system that have been made throughout the course of the experiments discussed. Comparisons of the performance with other optimisation techniques, hill climbing and simulated-annealing, are discussed. The optimisation process concludes with the use of e-Science techniques to allow for detailed statistical analysis of the evolved designs on high-performance computing clusters. The observed results for two-input logic gates demonstrate that the technique can be effective in the reduction of statistical spread in the delay and power consumption of circuits subject to intrinsic variability. The thesis finishes with the investigation of larger circuits which are assembled from the optimised cells. A proposed design methodology is introduced, in which the processes of logic design are broken into small blocks, each of which uses techniques from evolutionary computation to improve performance. This includes an investigation into the application of a multi-objective fitness function to improve the performance of logic circuits evolved using Cartesian Genetic Programming, which produces designs for logic multiplier and display driver circuits which are competitive with human-produced designs and other evolved designs. These designs are assessed for their variability tolerance, with the multiplier circuit demonstrating an improvement in delay variability.