Finding deterministic solution from underdetermined equation: Large-scale performance modeling by least angle regression

The aggressive scaling of IC technology results in high-dimensional, strongly-nonlinear performance variability that cannot be efficiently captured by traditional modeling techniques. In this paper, we adapt a novel L1-norm regularization method to address this modeling challenge. Our goal is to solve a large number of (e.g., 104∼106) model coefficients from a small set of (e.g., 102∼103) sampling points without over-fitting. This is facilitated by exploiting the underlying sparsity of model coefficients. Namely, although numerous basis functions are needed to span the high-dimensional, strongly-nonlinear variation space, only a few of them play an important role for a given performance of interest. An efficient algorithm of least angle regression (LAR) is applied to automatically select these important basis functions based on a limited number of simulation samples. Several circuit examples designed in a commercial 65nm process demonstrate that LAR achieves up to 25× speedup compared with the traditional least-squares fitting.