Graphical IDDQ signatures reduce defect level and yield loss
IEEE Transactions on Very Large Scale Integration (VLSI) Systems
| Hi-index | 0.00 |
The contributions of this dissertation are new IDDQ current testing features and corresponding application methodologies. In the graphical IDDQ current testing research, we discovered that noise, in the entire set of current measurements for a chip, is a vastly superior feature for determining whether a chip is good or bad, when compared with present methods. A single IDDQ current threshold, whether absolute or differential, cannot separate good/bad chips with any desirable accuracy or confidence for the SEMATECH data. The reason is that the good chip IDDQ signature can be one or many well-defined bands of measurements. A bad chip IDDQ always shows noise and glitches in the band structure. The essence of the graphical IDDQ test is to look at the shape of the waveform defined by the total set of the IDDQ measurements, to see: (1) The number of bands that all of the current measurements cluster into, (2) The width and separation of the bands, and (3) Current glitch or noise detection among all IDDQ measurements. It is different from other IDDQ testing methods in that it allows good circuits to have multiple bands inside the entire measurement set; and, it differentiates chips by detecting noise (glitches) among the bands that the IDDQ measurements cluster into, and these glitches may not exceed the allowable thresholds of conventional IDDQ testing. This testing method is generalized and improved using data mining techniques so as to be customized to different digital CMOS chips and different manufacturing technologies. It shows very high accuracy for both SEMATECH data with a test escape rate of 5.97% and an overkill rate of 1.2%, (the method also corrects probing errors), and 0.18 μ m process ASIC data with a test escape rate of 6.2% (the overkill rate is not presented here because of insufficient data), compared to a test escape of more than 7.5% for the ΔIDDQ method and 7.6% for the current difference method. For SEMATECH data, the graphical IDDQ method had a 1.2% test overkill, compared with 2.3% for the single threshold method, 6.1% for current difference method and 7.6% for the ΔIDDQ method. Practical application guidance, including the application procedure, the classifier generation and the test limitations, are given in the thesis. Correlation analysis is conducted on the feature set of graphical IDDQ testing using SEMATECH data. Results show that there is a strong correlation between some features that should not be causal. This proves that the choice of a mixed parametric and non-parametric graphical IDDQ testing classifier to relax the correlation among these selected features is worthwhile in this dissertation.