From Big Data to Real Time Manufacturing Intelligence Keith Arnold
Agenda Introduction Adaptive Test Background O+ System Architecture Data Feed Forward Questions & Discussion Optimal+ 2014 Company Confidential 2
Introduction Adaptive Test & Data Feed Forward the long awaited test methodology DFF leverages Big Analog Data to create predictive models Data Feed Forward is predictive Normal test methods are reactive Access to massive, comprehensive data sets is key for constructing good models Modeling can be daunting Global Data Infrastructure & integrity are KEY Optimal+ 2014 - All rights reserved 3
Adaptive Test A set of methods to automatically change test conditions, manufacturing flow, test content, test limits, and test outcomes in order to improve the effectiveness of the test operation. Goals Increase quality, reduce cost, increase throughput Insignificant test time overhead cost Fully automated and integrated with test process Optimal+ 2014 - All rights reserved 4
ITRS Architecture for Adaptive Test Optimal+ 2014 - All rights reserved 5
Adaptive Test - Types In-situ Data from the current operation/device insertion Speed grading, trim, sample diagnostics Post Test Data from statistics between operations to re-bin devices Spatial Outlier Detection, PAT, flow change Data Feed Back Data from previous device(s) in same operation Real Time DPAT, Dynamic TTR, Test Augmentation Data Feed Forward (DFF) Data source is from previous operation(s) for same or multiple lots Quality Grading/Indexing, Burn-In reduction, Die Pairing Optimal+ 2014 - All rights reserved 6
O+ System Architecture OSAT / CM / Factory OEM / IDM / Fabless MES CLIENT APPLICATIONS Analytics Queries Rules Simulations etest/wat Guidance & Requests Alerts & Linked Reports Wafer Test PROXY SERVER APPLICATION SERVERS OPTIMAL+ DATABASE SERVERS OPTIMAL+ CLOUD OR ON PREMISE Package Test Test Floor Monitor Optimal+ 2014 - All rights reserved 7
DFF Applications Applications Quality Grading / Indexing Escape Prevention & Outlier Detection Burn-In Reduction / Elimination Test Insertion Reduction Smart Die Pairing Optimal+ 2014 - All rights reserved 8
DFF Components Modeling Domain & Modeling Expertise Comprehensive DB of all cross operational data Modeling Tools Regression, SVM, ANN, Decision Trees Recipe Generation Rules Engine Virtual Operation Rule Dynamic model updates Global Data Infrastructure Direct Access to OSATs and Fabs Local Factory Operational DB Remote O+ Proxy Services Integration with OSAT workflow and MES Test Program Interface DFF Test Program API Optimal+ 2014 - All rights reserved 9
Cross Site DFF Flow Example Fab Sort Final Test System PCM WS WS Re-Binning FT1 Burn-In FT2 FT3 SLT PCM to WS DFF Local Proxy Server Local Proxy Server WS + FT to SLT DFF Incoming data from any operation Outgoing DFF Payload O+ DB Fabless/IDM Optimal+ 2014 - All rights reserved 10
DFF Data Flow - Considerations Challenges in Cross Site DFF When the DFF data needs to be available Minutes or days Where should the payload be sent One or many locations When the payload should be sent Immediately or triggered by an MES event Who is responsible for DFF data delivery O+ or Product Owner Dealing with missing data (ECID) Optimal+ 2014 - All rights reserved 11
DFF Assumptions Previous operation data exists in central DB ECID consistency across operations etest/pcm, Wafer Sort, Final Test, SLT Valid coordinate mapping PCM to Wafer Sort DFF payload is relatively small Sufficient storage on O+ Proxy at OSATs DFF API installed in Test Program Optimal+ 2014 - All rights reserved 12
DFF Data Types Raw Test Data Virtual Test Aggregated test data (mean, sigma, Cpk, ) Sequoia Scripted functions & algorithms Outliers: DPAT, Cluster, NNR, GDBN, ZPAT Transforms: Kriging, Z Score, Sigmoid Complex Model implementation Quality Index Performance Index Virtual Operations Contains copied test data and Virtual Tests Triggers generation of DFF file May trigger transmission of DFF to destination May encompass multiple operations Optimal+ 2014 - All rights reserved 13
Virtual Operation Rule 14
DFF Case Study - SLT Problem SLT costs were too high Dramatic impact on throughput Solution Perform modeling of all cross operational data and derive model that accurately categorizes devices likely to fail SLT Split devices by bin at FT step using DFF TP API Only devices that need SLT follow that path Dramatically reduces SLT cost and impact on throughput Optimal+ 2014 - All rights reserved 15
DFF Case Study Burn In Drift Problem IDDQ Drift from pre (FT1) to post BI (FT2) test Need a way to calculate cross operation result and re-bin questionable devices with no additional insertions Solution Create VOR for each drift parameter at FT1 Feed forward each drift parameter to FT2 DFF TP API performs the drift calculation and splits devices by bin at FT2 Questionable devices are split off from the population Significant improvement in quality for automotive customer Optimal+ 2014 - All rights reserved 16
DFF Case Study Quality Grading Problem High volume product required automotive grade option Need a transparent way to perform quality grading that is transparent to operations Solution Perform modeling of cross operational data and derive model that accurately categorizes devices with poor quality signature Create VOR (WSX) that pulls PCM and WS data Filter and transform 100 s of dynamic tests Calculate Model for current run Send DFF to FT destination facilitates User DFF TP API to re-bin questionable devices at FT Case study in implementation phase Optimal+ 2014 - All rights reserved 17
Conclusions DFF testing is PREDICTIVE, not REACTIVE DFF is an extremely powerful tool to improve quality, reduce cost and improve capacity DFF requires considerable domain experience and expertise in statistics and data science O+ is the only commercial supplier of DFF solutions Big Data Infrastructure is the barrier between the science projects and viable commercial solutions Optimal+ 2014 - All rights reserved 18
Questions & Discussion Optimal+ 2014 Company Confidential 19
Thank You 20