Um Método Híbrido de Amostragem para Projeto Automático de Circuitos Integrados Analógicos Considerando Otimização de Produtividade

Abstract

Regarding the design automation level, digital integrated circuits are more advanced when they are compared with analog integrated circuits. As a result, the time-to-market of mixed-signal circuits is hindered due to the long time dedicated for analog blocks designs. Analog circuit sizing can be considered as an optimization problem and the task can be fulfilled by optimization heuristics. The obtained solution depends on the circuit modeling strategy and performance estimation, which is usually performed by electrical simulation. In general, the solutions are located in the design space threshold, and consequently, any variation on the fabrication parameters of the devices affects circuit performance. In this case, it is necessary the inclusion of the yield prediction in the design flow to ensure that parameters variability will not affect circuit performance. Yield estimation is done with the assistance of Monte Carlo analysis in the optimization flow, this results in na increase computational processing. In order to reduce the number of Monte Carlo runs and minimize the required time to estimate the resultant yield in each iteration, two different sampling methods are analyzed in this work. Latyn hypercube sampling method, used in the first iterations of the optimization process, it is capable of providing a good reliability in yield estimation even with a small number of samples. After a certain point, the sampling method used in the optimization process is changed to the conventional random sampling, in order to refine design space search. The proposed methodology was added to UCAF framework and used to size automatically a Miller operational transconductance amplifier, in which the design presented advantages in terms of processing time and yield optimization. It was demonstrated that the use of different sampling methods for Monte Carlo analysis in a single design flow is capable of providing a more efficient search in the design space.