Diagnosis-Based Failure Analysis

Osei Poku

Rao Desineni

Project

The growing complexity associated with manufacturing nanometer scale integrated circuits (ICs) makes it increasingly difficult to successfully identify failing mechanisms in faulty ICs. The arduous task of chemically disassembling and physically localizing the problematic region in a failing chip is extremely time consuming and largely impractical in the quick debug loop necessary during a typical process ramping stage. This procedure, known as physical failure analysis (PFA), is also destructive and so the first hypothesis about the defect location must be correct. Hence, PFA, when unsuccessful, is extremely wasteful of time and resources. Traditional methods of improving yield by physically analyzing failing ICs must inevitably be supplemented or even replaced by faster and more automated ways of extracting useful information about the process-design interaction. We are investigating new ways of using diagnosis as an automatic failure analysis vehicle. In our diagnosis research [1], we extract defect models from failing ICs based on a few assumptions about the nature of the defect. Under these assumptions, a defect consists of a set of faulty signal lines, or victims, whose erroneous behaviors are influenced solely by their physical neighbors. We use these assumptions to identify the defective locations and, ultimately, generate a model of the defect that describes the conditions under which the defect causes faulty behavior. Analysis of the extracted defect models provides valuable insight into the underlying failure mechanism and can be statistically mined to provide a broader view of the process and design interaction, and an understanding of its prevalent failure mechanisms.

Selected Highlights

As a preliminary step, we performed diagnosis on 800 failing dice [2] to highlight some of the benefits of our approach when compared to a traditional diagnosis procedure. For each failing die analyzed, we extracted a defect model that contained the failing victim line(s) and the subset of their physical neighbors that were deduced to influence each victims' failing behavior. It is assumed that during traditional PFA, the complete set of physical neighbors of a victim will be considered as potential defect sites. However, using our methodology, a reduced number of physical neighbors is identified for PFA. In Figure 1, we show the extent of the reduction in the number of neighboring signal lines after our diagnosis. The x-axis shows the number of neighbors in the defect model extracted during diagnosis. On average, our methodology reduces the number of neighbors by 70%. In Figure 2, we show the distribution of the extracted models across successive metal layers. A defect model is assigned to a metal layer if one or more physical neighbors in that metal layer are determined to influence the behavior of the victims. The plot confirms an expected result: the first metal layer exhibits the majority of failures.