The linearized performance penalty (LPP) method for optimization of parametric yield and its reliability

Circuit tolerance design by differential evolution with hybrid analysis method

“…When the circuit scale and the number of uncertain parameters become large, the computational cost of such method is very expensive. Thus, approaches based on approximation have been proposed to reduce the computational cost, such as vertex analysis [3,5,6], polyhedral approximation [14] and quadratic approximation [15]. On the other hand, these approaches gain speed at the cost of accuracy which usually produces underestimated results especially when the variations are huge or performance surfaces are highly nonlinear.…”

An application-level synthesis methodology for multidimensional embedded processing systems

“…Yield estimation accounts for estimating the expected yield at the current design point (in the multidimensional synthesis case, estimation of the system accuracy) while yield improvement aims at obtaining a new design point with higher yield (higher volume of the perturbation space). In general, yield estimation (see [12] and [13] for an exhaustive review) is carried out either with a Monte Carlo sampling of the parameter space and the associated optimal experiment design methods for reducing the number of samples (response surface methods [13], [16], [17]) or by considering sophisticated boundary methods aimed at generating a description for the yield acceptability frontier (boundary and surface integrals [12], [14], [15]). Subsequent yield improvement techniques move the circuit parameters from an initial configuration toward a new point which maximizes some figure of merit (e.g., the distance from the point to the acceptability region border or its approximation [12]- [15]).…”

“…In particular, the optimization method can be applied to any Lebesgue measurable figure of merit (hence, comprising the yield one) without assuming the continuity and/or differentiability hypotheses as required by gradient descent methods [12]- [14] or the memoryless approximating functions assumption requested by surface response techniques methods [13], [16], [17]. In addition, the optimization procedure does not suffer from the presence of local minima in the "yield maximization" figure of merit which is a critical issue in local optimization methods based on gradient estimates.…”

Design of robust test criteria in analog testing

“…In order to obtain sample elements on both sides of the separating hyperplane and to retain the physically interpretability, we generate training and validation sample elements according to a normal distribution with an expected value x w , while keeping covariance matrix x . Here, x w is the worst-case point [22,12,13] of the considered specification. The worst-case point of a specification is defined as the set of statistical parameters that just satisfies the corresponding specification and has the highest probability density to occur.…”