A Machine Learning-based Approach to Optimize Repair and Increase Yield of Embedded Flash Memories in Automotive Systems-on-Chip

Nowadays, Embedded Flash Memory cores occupy a significant portion of Automotive Systems-on-Chip area, therefore strongly contributing to the final yield of the devices. Redundancy strategies play a key role in this context; in case of memory failures, a set of spare word- and bit-lines are allocated by a replacement algorithm that complements the memory testing procedure. In this work, we show that replacement algorithms, which are heavily constrained in terms of execution time, may be slightly inaccurate and lead to classify a repairable memory core as unrepairable. We denote this situation as Flash memory false fail. The proposed approach aims at identifying false fails by using a Machine Learning approach that exploits a feature extraction strategy based on shape recognition. Experimental results carried out on the manufacturing data show a high capability of predicting false fails.