New Models for Burn-In of Semiconductor Devices Horst Lewitschnig - Infineon Technologies Austria AG, Daniel Kurz, Jürgen Pilz - Alpen Adria University Klagenfurt, Austria 2015-10-17 - Daejeon - South Korea
1 Burn-In today vs. new concepts 2 Countermeasure Model 3 4 5 6 Publication List Horst Lewitschnig - Infineon Technologies Austria AG, Daniel Kurz, New Jürgen Models Pilz -for Alpen Burn-In Adriaof University Semiconductor Klagenfurt, Devices Austria
1 Burn-In today vs. new concepts 2 Countermeasure Model 3 4 5 6 Publication List Horst Lewitschnig - Infineon Technologies Austria AG, Daniel Kurz, New Jürgen Models Pilz -for Alpen Burn-In Adriaof University Semiconductor Klagenfurt, Devices Austria
Bathtub curve Countermeasure Model Bathtub curve
Burn in - Elimination of Early life Early Fails Means to weed out early life fails: DD-screening Stress tests Outlier detection... Burn-In Burn-In is one method amongst many others to eliminate early life fails.
Burn-In today Countermeasure Model 2 Concepts Random sampling: 100 % Burn-In Optimization by Burn-In time reduction. Burn-In study Random samples are taken out of production and put into burn-in. In parallel 100 % Burn-In, as long as the Burn-In study is not passed. If the Burn-In study is pass, switch to BI-monitoring.
Burn-In today Countermeasure Model p @ 90 % CL Random sampling: 0/100000 = 23 ppm @ 90 % CL; 0/350000 = 6.6 ppm @ 90 % CL; 1 ppm @ 90 % CL = 0/2.31 Mio. Beyond Sampling Pure random sampling gives limited results. The idea is to include further information to the random sample. On the following pages some of these ideas are introduced.
Burn In - new concepts Extended approach The classical approach looks only at the random sample. The extended approach takes further information into account: Countermeasure model: The efficiency of countermeasures as well as the overall performance of BI-studies is taken into account. Multiple reference products: In case of two or more reference products, an exact area scaling is now possible. Synergy model: Information about equal chip subsets, that have already passed Burn-In studies, are used. Area specific scaling: If different chip subsets show different ppm-values, referencing is done on these area-specific levels.
1 Burn-In today vs. new concepts 2 Countermeasure Model 3 4 5 6 Publication List Horst Lewitschnig - Infineon Technologies Austria AG, Daniel Kurz, New Jürgen Models Pilz -for Alpen Burn-In Adriaof University Semiconductor Klagenfurt, Devices Austria
Idea: Burn In Fails and Countermeasures - Estimation of p 1 Burn In Fail, Countermeasure α % effective 0.1 = α B(0; n, p) + (1 α) B(1; n, p) Effectiveness The effectiveness of a countermeasure reflects expert knowledge, available data,... It is a prior distribution. We propose to use the 10 % - quantile of this prior distribution.
Idea: Burn In Fails and Countermeasures - Estimation of p Example n = 100 k pcs., 1 Burn In fail, countermeasure 80 % effective. 0.1 = 0.8 B(0; n, p) + 0.2 B(1; n, p). Solution without countermeasure: p = 39 ppm @ 90% CL, with countermeasure: p = 27 ppm @ 90% CL.
Burn In Fails and Countermeasures - full picture Burn In Fail with Countermeasure - full picture Let s assume we put 100 k devices to burn in with 1 fail, introduce a countermeasure with 80 % effectiveness, and burn again 100 k pcs. with no fails. Taking all information into account, this results in 0.1 = 0.8 B(0; 200000, p) + 0.2 B(1; 200000, p), 13.7 ppm @ 90 % CL.
1 Burn-In today vs. new concepts 2 Countermeasure Model 3 4 5 6 Publication List Horst Lewitschnig - Infineon Technologies Austria AG, Daniel Kurz, New Jürgen Models Pilz -for Alpen Burn-In Adriaof University Semiconductor Klagenfurt, Devices Austria
Countermeasure Model Let s assume two technologies which only differ in the metal block:
Countermeasure Model Let s say technology 1 had 250 k pcs. in Burn-In, technology 2 had 100 k pcs. in Burn-In. Products are assembled from all combinations of their subsets (combinatorial model):
Countermeasure Model Example - Comparison If we look only at technology 2: 1 / 100000 pcs. 39 ppm @ 90 % CL. If we look at technology 1 and 2: 1 / (100000 + 250000) pcs. for the substrate 0 / 100000 pcs. for the metal block 29 ppm @ 90 % CL.
1 Burn-In today vs. new concepts 2 Countermeasure Model 3 4 5 6 Publication List Horst Lewitschnig - Infineon Technologies Austria AG, Daniel Kurz, New Jürgen Models Pilz -for Alpen Burn-In Adriaof University Semiconductor Klagenfurt, Devices Austria
Reference products 1 and 2 are built up with elements of size A gcd 1 fail at reference product 2 can be caused by 1 or several failed elements of size A gcd failed on the same chip (probabilities based on hypergeometric distribution). (all possibilities to built up reference products 1 and 2, p gcd ) = 0.1, numerical solution for p gcd
Example Reference Product 1: 5 mm 2, 0/100000 pcs, 23 ppm @ 90 % CL; Reference Product 2: 12 mm 2, 1/100000 pcs, 39 ppm @ 90 % CL. A gcd = 1 mm 2, p gcd = 2.3 ppm @ 90 % CL, For a follower product with 7 mm 2 : p = 16 ppm @ 90 % CL
1 Burn-In today vs. new concepts 2 Countermeasure Model 3 4 5 6 Publication List Horst Lewitschnig - Infineon Technologies Austria AG, Daniel Kurz, New Jürgen Models Pilz -for Alpen Burn-In Adriaof University Semiconductor Klagenfurt, Devices Austria
Area scaling Countermeasure Model Likelihood for a defect per device scales with the die size.
Burn-in study Countermeasure Model Figure: Reference and follower products per technology
Follower product - same CMOS, 3 x bigger DMOS
Separate area scaling - example Follower product - same CMOS, 3 x bigger DMOS Reference Product: 1/100k 39 ppm @ 90 % CL Classical area scaling: 78 ppm @ 90 % CL, Separate area scaling: 65 ppm @ 90 % CL.
1 Burn-In today vs. new concepts 2 Countermeasure Model 3 4 5 6 Publication List Horst Lewitschnig - Infineon Technologies Austria AG, Daniel Kurz, New Jürgen Models Pilz -for Alpen Burn-In Adriaof University Semiconductor Klagenfurt, Devices Austria
Publications Countermeasure Model Publications Countermeasure Model: Decision-Theoretical Model for Failures Which are Tackled by Countermeasures; Kurz, Lewitschnig, Pilz; IEEE Trans. Reliability, Vol. 63, No. 2, June 2014. R-package GenBinomApps, Lewitschnig and Lenzi, 2014. : Modeling of chip synergies for failure probability estimation in semiconductor manufacturing; Kurz, Lewitschnig, Pilz; applied to: Journal of Applied Statistics. : Failure probability estimation with differently sized reference products for semiconductor burn-in studies; Kurz, Lewitschnig, Pilz; Applied Stochastic Models in Business and Industry, online since Dec. 2014.
Publications Countermeasure Model Publications : An advanced area scaling approach for semiconductor burn-in; Kurz, Lewitschnig, Pilz; Microeletronics Reliability, Vol. 55, Issue 1, January 2015 Acknowledgment: The work has been performed in the project EPT300, co-funded by grants from Austria, Germany, Italy, The Netherlands and the ENIAC Joint Undertaking. This project is co-funded within the programme Forschung, Innovation und Technologie für Informationstechnologie by the Austrian Ministry for Transport, Innovation and Technology.