Exploiting defect clustering for yield and reliability prediction

An integrated yield-reliability model is verified using burn-in data from 77 000 microprocessor units manufactured by IBM Microelectronics. The model is based on the fact that defects over semiconductor wafers are not randomly distributed but have a tendency to cluster. It is shown that this fact can be exploited to produce dies of varying reliability by sorting dies into bins based on how many of their neighbours test faulty. Dies that test as good at the wafer probe, yet come from regions with many faulty dies, have a higher incidence of infant mortality failure than dies from regions with few faulty dies. The yield-reliability model is used to predict the fraction of good dies in each bin following a wafer probe as well as the fraction of failures in each bin following stress testing (e.g. burn-in). Results show excellent agreement between model predictions and observed data.