A MATLABĀ®-Based Technique for Defect Level Estimation Using Data Mining of Test Fallout Data versus Fault Coverage

Kanad Chakraborty
Cypress Semiconductor


Abstract

To achieve progressively lower defective parts per million (PPM) levels in silicon, we need to target a wide variety of fault and defect models, such as stuck-at, at-speed and bridging faults and IDDQ/IDD failures, and apply tests targeting each such model. This paper describes a novel MATLABĀ®-based methodology for quantifying PPM improvements based on fallout data during manufacturing test application. It is seen that tests that target a range of defect models, each with moderately high levels of coverage, may be better in terms of lowering PPM than those that target a single fault model with high levels of coverage. This analysis is explained using regression models for PPM yield versus fault/defect coverage. This approach is beneficial to semiconductor companies for calibrating their fault coverage goals to meet PPM requirements from automotive and other customers.