1 / 29

Statistical Methods for "Zero Defect” Quality in Nanometer Integrated Circuits

Statistical Methods for "Zero Defect” Quality in Nanometer Integrated Circuits. Adit D. Singh Electrical & Computer Engineering Auburn University. IEEE EWDTS2009. Moscow Sept 20 2009. Outline. DPM requirements for ICs – why the push to “Zero Defects”

alair
Download Presentation

Statistical Methods for "Zero Defect” Quality in Nanometer Integrated Circuits

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Statistical Methods for "Zero Defect” Quality in Nanometer Integrated Circuits Adit D. Singh Electrical & Computer Engineering Auburn University IEEE EWDTS2009 Moscow Sept 20 2009

  2. Outline • DPM requirements for ICs – why the push to “Zero Defects” • Perfect testing of all parts for zero DPM is impossible/cost prohibitive • Statistical Methods and Adaptive tests: - Test each part differently only to the extent it needs to be tested - Only “suspect” parts are extensively tested • How suspects can be identified • Results and Conclusions

  3. The IC Quality ChallengeApproximate order-of-magnitude estimates • Number of parts per typical system: 100 • Acceptable system defect rate: 1% (1 per 100) • Therefore, required part reliability 1 defect in 10,000 100 Defects Per Million (100 DPM) Typical Commercial Requirement ~100 DPM ~500 DPM for ASICs

  4. The IC Quality Challenge New Automotive industry target : “zero” defects! Why? • Electronics contributing to excessive warrantee repair costs • Offending parts can be identified through full part traceability over 4-5 year warrantee period -High unit volumes provide robust statistical data • US annual sales ~ 15 million

  5. Required Test QualityAssume 2 million ICs manufactured with 50% yield • 1 million GOOD >> shipped • 1 million BAD >> test escapes cause defective parts to be shipped • For 100 BAD parts in 1M shipped (DPM=100) Test must detect 999,900 out of the 1,000,000 BAD For 100 DPM: Required Test Coverage = 99.99%

  6. So How Do we Target Zero DPM? Test better, longer and harder OR Learn from Airport Security Screening and Test smarter! After 9-11 requirement for Airport Security Screening immediately went to “Zero” Defects First response: Test every passenger perfectly

  7. New Airport Security Check-in Procedures

  8. How to Target “Zero” Defects Smart Tests Exhaustive testing every airline passenger is cost prohibitive - Cannot strip search everyone Solution: Identify a relatively few “suspect” parts to test intensively (or discard) “Suspects” Identified Based on: Profiling: “company you keep” Outliers: “something doesn’t look right”

  9. Extreme cases! What happens to Suspects? More generally Suspects are discarded - bumped off the flight: no fly list Suspects are tested further Test Optimization - saves high cost testing of exhaustive testing of every passenger Adaptive testing - Appropriate tests are applied depending on “what looks different”

  10. Targeting Zero DPM in ICs Cannot perfectly test all parts –too expensive Solution: Intensively test “suspect” chips or even discard without conclusive evidence of fault “Suspects” Identified Based on: Profiling: “company you keep” - Die from bad neighborhoods Outliers: “something doesn’t look right” - Die that behave differently in some way

  11. Understanding Screening Methods Targeting Zero Defect IC Quality Profiling: “company you keep” - Statistical defect clustering in manufactured lots and wafers – Bad neighborhoods Outliers: “something doesn’t look right” - Abnormal, although within specification, test responses

  12. DPM Depends on incoming YieldTest Coverage: 99.99% (Escapes 100 per million defective) • 1 Million Parts @ 10% Yield 0.1 million GOOD >> shipped 0.9 million BAD >> 90 test escapes DPM = 90 /0.1 = 900 • 1 Million Parts @ 90% Yield 0.9 million GOOD >> shipped 0.1 million BAD >> 10 test escapes DPM = 10/0.9 = 11

  13. Exploiting Spatial Defect Statistics for Test • If we can bin die from wafers into bins with different yield: • We can get bins with lower DPM for the same test • Optimize further testing for each bin • Binning is possible because defects on wafers tend to cluster DPM=11 90% DPM=900 10%

  14. Manufacturing Defects and Die Yield • Two classes of Manufacturing Defects • Gross or area defects • Random Spot Defects • In mature well controlled processes, die yield is mostly limited by random spot defects

  15. Poisson Defect Statistics • The simplest defect distribution • model for semiconductor • wafers assumes that • random spot defects are • uniformly distributed • Die Yield x x x x x x x x x x x Where is the average number of defects per die

  16. Defect Clustering on Wafers • The Poisson model has been found to consistently underestimate yield • This suggests, defects on semiconductor wafers are not uniformly distributed but are clustered x x x x x x x x x x • For a given total number of defects on the wafer, defect clustering results in more die with multiple defects, and therefore more defect free die (higher yield)

  17. Defect Clustering on Wafers • Defect clustering has been observed in virtually every fabrication line in 40 years of semiconductor manufacturing experience • The causes of defect clustering are numerous and varied, and can be related to many different fabrication steps. • The extent of defect clustering can vary based on the product and process technology

  18. Binning Good Die Based on Neighbors 0 Bad Neighbors 1 Bad Neighbor 2 Bad Neighbors . . . 8 Bad Neighbors X X X X 4 4 0 X X X X 1 2 X X 4 X X X • Best Bin has dice with highest “a priori” yield • => Lowest defect levels from test escapes

  19. Burn-In Fail Probability for 77,000 Chips - Barnett, Singh VTS 2002 Fail Probability (0+) 1 2 3 4 5 6 7 8 Bin 1 => Bin 8

  20. Burn-In Fail Probability for 77,000 Chips - Barnett, Singh VTS 2002 Fail Probability Burn-in 1(+0) 2 3 4 5 6 7 8 Bin 1 => Bin 8 (Neighboring Bad Die)

  21. Binning for Low Defect Levels • Best bin defect levels (DPM) are typically 3-7X better than the lot average • Greatest benefit for high clustering and low yields • For extensively tested die needed for ultra reliable applications - virtually impossible to achieve DPM improvements from additional testing • Further testing can focus on the worse bins to optimize test costs – Adaptive Test

  22. Adaptive Test and DPM Optimization by Binning based on Local Yield • Binning die based on Local Region Yield gives die with various degrees of DPM and reliability • Further testing can focus on bins with die containing repaired defects and/or from defective neighborhoods for test optimization • Methodology covered by US Patents 7409306 and 7194366 • “System and Method for Estimating Reliability of Components for • Testing and Quality Optimization” - Inventors: Barnett and Singh

  23. Identifying “Suspects” from Abnormal Response as compared to Matched Parts • Something doesn’t “look right” on some analog measurement –e.g. IDDQ, speed etc. • Part still tests within functional specifications

  24. Identifying “Suspects” from Abnormal Response from Matched Parts Key Idea: • Analog IC performance measures should be similar for matched parts • Any anomalies, even if within functional specifications, indicate a defect which could be a test escape and fail in the field, or result in a reliability problem

  25. Eaample: minVDD Timing Tests - Madge et al VTS 2003 • minVDD Testing finds the lowest VDD for which the circuit passes a delay fault (TDF) test for a given clock speed • An abnormal minVDD value with respect to the expected value for matched parts indicates a defect that may be a test escape or reliability failure

  26. MinVDD vs Device SpeedTwo different lots showing min VDD outliers and lot-to-lot intrinsic variation. Ring Oscillator Frequency

  27. minVDD Testing Minimum VDD results for different functional tests clearly showing min VDD outliers (circled)

  28. Conclusion • Faulty die are more likely to be found near other faulty die. • Die that test good yet are near many faulty die are greater reliability risks: suspect parts • Suspect parts can also be indentified as outliers from analog test measurements • Adaptive tests focus on suspect parts to minimize test escapes at affordable costs

  29. Questions?

More Related