Reliability improvement and burn in optimization through the use of die level predictive modeling

Abstract

A die-level, predictive defect model is presented which identifies sub-populations of a die with varying infant mortality. The model is shown to be twice as efficient in identifying unreliable material as traditional wafer or lot level methods. Applications include reduced burn in requirements while continuing to meet infant mortality goals. Key results include: 1) a predictive model for latent defects was developed using wafer sort yield data; the model was validated with production burn in data; 2) optimum nearest neighbor effect is two die in each direction (24 nearest neighbors), regardless of die size; 3) near optimum efficiency for local yield function can be obtained by generic weighting coefficients rather than product specific weights; 4) extreme edge die have a different neighborhood than center die; 5) die level low yield criteria are twice as efficient as traditional wafer or lot level scrap criteria; using a die level approach can increase both reliability and output; 6) the most efficient use of burn in resources is by identifying die level subpopulations for burn in, rather than wafer or lot level; sub-populations of die with high failure rates occur even on high yield wafers, thus die level segregation is needed; 7) a methodology for setting a generic die level scrap criterion based solely on die size is proposed; 8) the influence of cache repaired die in predicting a defect was found to be less pronounced than other failure bins.