Abstract
The circuit performance and the parametric yield are important for VLSI design, but it is not enough to consider only these factors. One of the other significant factors is the performance spread which should be as small as possible. Due to the manufacturing process variation, the circuit performance has some variations around its nominal performance. In order to decrease this performance spread efficiently, a novel algorithm based on the steepest-descent method and some statistical methods is introduced. At first, a linear relationship is assumed between the parameter values and the circuit performance, and then only tens of simulations need to be run to identify coefficients in the performance expression. In order to save the cost, novel analytical means are used to calculate the performance spread and its gradient. This algorithm decreases the spread efficiently and one example is shown to demonstrate its capability.