Silicon Debug with Maximally Expanded Internal Observability Using Nearest Neighbor Algorithm

One of the most difficult challenges during the process of silicon debug is overcoming the bottleneck of limited visibility of internal states. Although the application of state restoration technique enhances the limited debug data available through on-chip trace buffers, yet the number of restored signal states are not significant. This paper proposes an approach which addresses the limited observability problem through a machine learning perspective. Based on training with pre-silicon buggy signatures on a relatively smaller design, a model is developed which identifies a set of neighbors for every flip-flop of the design. The application of nearest neighbors principle eliminates the obstacle of unknown signal values despite restoration because these values are obtained from the neighbors. Experimental results on benchmark circuits depict that the proposed approach is able to correctly discover 93% of the total signal values. The methodology is verified with the help of cross-validation of the debug data on designs injected with gate-level error models.