A Logic Diagnosis Methodology for Improved Localization and Extraction of Accurate Defect Behavior

PADRE: Physically-Aware Diagnostic Resolution Enhancement

“…The candidate features no_pass_states, no_fail_states, no_incst_states, and candidate_entropy characterize the physical characteristics of a candidate when it is both activated and sensitized. Specifically, the neighborhood state of a candidate is defined to be the logic values driven on lines that are in physical proximity of the candidate for tests that detect the candidate (i.e., TPSF and TFSP patterns) [1]. The neighborhood of a candidate, as illustrated in Figure 2, includes:…”

“…In past work [1,4,10,11,12], it has been shown that the consistency check was very adept at identifying bad 1 Although simulation data is used in this work, it should be noted that it is not integral and is only used in order to verify accuracy.…”

“…Diagnosis is a fast and non-destructive approach to preliminarily identify and locate possible defects in a failing IC [1]. It is a software-based method that analyzes the applied tests, the failed IC tester response, and its netlist/layout to produce a list of diagnostic candidates that represent the locations and sometimes behaviors/types of defects in the chip.…”

“…In [8,9], it is suggested that candidates detected by more tester-passing patterns are less likely to capture the actual defective location(s). Other work reveals that the same neighborhood state of a good candidate should not be observed in both the Tester-Pass-Simulation-Fail (TPSF) and the Tester-Fail-Simulation-Fail (TFSF) patterns [1,4,10,11,12]. If a neighborhood state appears in both TPSF and TFSF patterns, the candidate is said to be inconsistent and is likely incorrect [1,4,10,11,12].…”

“…Other work reveals that the same neighborhood state of a good candidate should not be observed in both the Tester-Pass-Simulation-Fail (TPSF) and the Tester-Fail-Simulation-Fail (TFSF) patterns [1,4,10,11,12]. If a neighborhood state appears in both TPSF and TFSF patterns, the candidate is said to be inconsistent and is likely incorrect [1,4,10,11,12]. While these techniques are effective (e.g., the work in [10] reports a resolution improvement of 67% for 2,293 chips), they only One issue with creating a supervised classifier is the need for "training data" [13].…”