Statistical Learning Tools and Methodologies for Test Data Analysis

As manufacturing technologies continue to scale, increasing test cost has been a growing concern. In today's industrial practice, system test has been the standard method for measuring performance of high-performance microprocessors. The feasibility of using low-cost structural tests for this purpose has been studied for many years. In most of the past studies, the key idea was to find the structural test measurement that best correlates to target Fmax behavior. This dissertation takes a different approach to the problem. In this work, correlation to target Fmax behavior is treated as a data-learning problem and data-learning approaches are studied for correlating structural test measurement results to the Fmax behavior. Based on industrial data on two recent high performance microprocessor designs, we discovered that a data-learning approach can always deliver better correlation results than existing practice of finding the single best structural measurement. While these results were encouraging, we discovered that in practice, obtaining a good correlation model from silicon data was usually not convincing enough for test engineers to utilize the model in production test. Often, a good correlation model needs to be explained with domain knowledge before it can be accepted for application. To understand why a good correlation model exists is a very challenging task. In this work, we take the first step by developing a framework to first understand the cause behind target Fmax behavior or more specifically, to understand the causes behind speed limiting paths that determine the Fmax behavior. Hence, in the last part of this thesis, we develop a framework for studying speed limiting paths from structural Fmax measurements based on transition fault pattern set. Although this framework should not be considered as the final product that can be used to explain a correlation model, the result that it was successfully applied to explain the causes behind some of the speed limiting paths demonstrates its effectiveness and shows that the framework can serve as a solid foundation for building such a final product in practice.