|
Improving Diagnostic Resolution Using Machine Learning
Osei Poku |
Shawn Blanton |
Improving yield and test quality often hinges on successfully identifying the causes of manufacturing imperfections. When these imperfections cause integrated circuits to fail, logic diagnosis is often used to localize the defect causing failure. However, in over 70% of the cases we investigated, diagnosis identifies more than one candidate (i.e., possible failure location), thus leading to ambiguity in any subsequent analysis such as physical failure analysis, where the failed chip is physically examined to uncover the root-cause of failure. Fig. 1 shows the cumulative distribution of the number of candidates that result from diagnosing over 8,000 failing GPUs (graphics processing units). As shown in the distribution, approximately 80% of these failures would be discarded using a one-candidate strategy. Any failure mechanisms that are mostly captured in that 80% portion with more than one candidate will simply go unnoticed.
In this work, we introduce a machine-learning-based framework called micro-learning, that combines various heuristics for ranking the candidates identified by diagnosis to significantly improve the likelihood of identifying the correct candidate. Each candidate is represented by a set of features that describe its location, impact on circuit behavior, and its behavior relative to other candidates. These features are used by micro-learning to predict actual defect locations from the set of candidates produced by diagnosis. Fig. 2 shows the results of an experiment measuring the improvement in prediction accuracy achieved by micro-learning when compared to conventional ranking metrics. For an entire diagnosis outcome to be correctly predicted, all candidates must be correctly predicted as either an actual defect location or an incorrect location. Using micro-learning, prediction accuracy is improved, on average, by 600% over conventional approaches for this very conservative definition of correct.
 |
 |
Fig 1. Cumulative distribution of the number of candidates resulting from the diagnosis of 8,874 failing Nvidia GPUs. |
Fig 2.Improvement in prediction accuracy for a few ISCAS85, ISCAS89 and ITC99 benchmark circuits. |
|
