DescriptionVariation-aware characterization for timing libraries (.libs) is a necessity for modern process nodes, due to the increased impact of process variability from shrinking feature size, increased device integration and low voltage operations.
A key hurdle to producing accurate variation model timing views is the runtime required to characterize Liberty Variation Format (LVF) .libs. Due to the millions of data points required over many process, voltage and temperature (PVT) corners, a full characterization run may require weeks or even months to complete.
However, the abundance of data points in .libs also present an opportunity to leverage machine learning (ML) methods to produce accurate characterization results while significantly reducing total characterization runtime.
In this paper, we present the usage and results of machine learning methods that complement SPICE characterization of LVF .libs to improve characterization runtime while maintaining equivalent output .lib accuracy. Results expected are on the order of ~100X speedup for direct runtime comparison between SPICE characterization and ML generation per .lib, and 40%-60% runtime reduction across a full set of .libs, depending on number of .libs generated using ML. We will present actual results from our testcase in this paper.