Study of CMOS Process Variation by Multiplexing Analog Characteristics by Karen Mercedes Gonzalez-Valentin Gettings Bachelor of Science, Universidad de Puerto Rico-Recinto de MayagUez, June 2000 Master of Science, Massachusetts Institute of Technology, June 2002 Submitted to the Department of Electrical Engineering and Computer Science in partial fulfillment of the requirements for the degree of Doctor of Philosophy at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY May 23, 2007 ©Massachusetts Institute of Technology, 2007. All Rights Reserved. Author D~partment of Electrical-4 ngiering and Computer Sciente May 23, 2007 Certified by -. - ning Professor of Electrical Encineefrino ins] Comnnt-r Science ervisor Accepted by Smith Studies Graduate on Committee Chairman, Department Department of Electrical Engineering and Computer Science AU '20237 BARKER Study of CMOS Process Variation by Multiplexing Analog Characteristics by Karen Mercedes Gonzilez-Valentin Gettings Submitted to the Department of ElectricalEngineeringand Computer Science on May 23, 2007, in partialfulfillment of the requirementsfor the degree of Doctor of Philosophy Abstract Aggressive technology scaling raises the need for efficient methods to characterize and model circuit variation at both the front and back end of line, where critical parameters such as threshold voltage and parasitic capacitance must be carefully modeled for accurate circuit performance. This thesis addresses this need by contributing a test circuit methodology for the extraction of spatial, layout and size dependent variations at both device and interconnect levels. The test chip uses a scan chain approach combined with low-leakage and low-variation switches, and Kelvin sensing connections, providing access to detailed analog device characteristics in large arrays of test devices. Front end of line (FEOL) test structures include transistors of different sizes, number of polysilicon fingers, polysilicon fingers proximity, and orientation, for both NMOS and PMOS MOSFETs. Back end of line (BEOL) test structures include parasitic coupling, plane to plane and crossover capacitances, measured using a charge-based capacitive measurement (CBCM) methodology integrated with switches in the scan chain. The testing of the designed test chip has proven successful for both device and interconnect test structures. Different layout practices in both NMOS and PMOS transistors are seen to result in significant differences in mean and standard deviation of measured output current, with 95% confidence or more. The FEOL structure analysis shows strong dependencies between layout practices: orientation offers a consistent but opposite offset in NMOS and PMOS transistors and variation increases for gate lengths split among fingers. Variation due to sizing follows Pelgrom's model, showing that variation increases for smaller gate lengths and widths, in both NMOS and PMOS transistors. Threshold voltage extraction and variation analysis also demonstrate how variation increases for smaller features. BEOL capacitances were extracted and subfemto Farad changes were detected for capacitive test structures. Spatial analysis reveals a large die-to-die trend in device performance. The parameter extraction and variation analyses made possible by the variation test chip enable the identification of likely variation sources, quantification of circuit impact and sensitivity, and specification of layout practices for variation minimization. Thesis Supervisor: Duane S. Boning Title: Professor of Electrical Engineering and Computer Science 3 4 Acknowledgements When thinking back about my seven year career at MIT, I can only thank the God Almighty for having blessed me with the opportunity of attending such a prestigious institution, and rub shoulders with so many fantastic and impressive people. It has been a long, windy road, and its successful conclusion could have not happened without the help and support of many people and institutions. I would like to thank my advisor, Prof. Duane Boning, for giving me the opportunity to join his research group and guiding me through my master's and doctorate degrees. Duane, your infinite patience, big heart, enthusiasm, impressive knowledge, and attention to detail have made me feel comfortable at MIT for the last seven years. Thank you very much for all your help and guidance. Special thanks to Prof. Joel Dawson, whose friendship and selfless help with the analog part of my thesis have been invaluable. I also want to thank Prof. Anantha Chandrakasan for his help and guidance with circuit design and life in general since my first day at MIT, and more recently as a thesis committee member. Also, thanks to Prof. Luca Daniel, also a thesis committee member, whose enthusiasm, motivation and good ideas have been very helpful in completing this thesis. I was able to collaborate with very talented people during the design and analysis of the work presented in this thesis. I wholeheartedly appreciate the collaboration of JinKyu Park, from Samsung Electronics, for providing ideas and guidance with the CBCM test structures. I also appreciate the help of Dr. Robert Lefferts for his enthusiastic help with the design of the test chip architecture, particularly the low leakage switches. I appreciate the help I received during my master's thesis that also influenced this thesis, including the contributions of my M.S. thesis partner Joseph Panganiban, and IBM's Sani Nassif and his research group, whose contributions in the scan chain design and design of experiments were very helpful for the design of this new test chip. I also very much appreciate the help in the understanding of device behavior from Prof. Jesus del Alamo, Cait Ni Chldirigh, and Niamh Waldron, all from MTL. I very much appreciate the comradery, collaboration and friendship from all my group mates at the Metrology group at MIT, present and past: Aaron Gower-Hall, David White, Han Chen, Brian Lee, Vikas Mehrotra, Tamba Gbondo-Tugbawa, Tae Park, Joseph Panganiban, Allan Lum, Michael Mills, Nigel Drego, Xiaolin Xie, Hong Cai, Tyrone Hill, Hayden Taylor, Shawn Staker, Kwaku Abrokwah, Brian Tang, Mehdi Gazor, Karthik Balakrishnan, Daihyun Lim, Shyam Vudathu, Daniel Truque, Ali Farahanchi, Ajay Somani, and Edward Paul. Nigel, thank you for all your noble help and support with the design, tape out, and testing of the test chip, not to mention all the scripts and all the support when things were not going so well. I am so lucky to have you and Vidya as my friends. Daihyun and Karthik, thank you for all of your great ideas regarding the design and testing of the chip. To everybody else: thanks for all the collaboration, all the fun conversations, all the donuts, and all the good times. You will always be in my heart. 5 Many thanks to everybody at MTL for all their support and friendship. Special thanks go out to my dear friend Debb Hodges-Pab6n. Thank your friendship and for making me part of your life. You are the best ambassador MTL could wish for. Also, many thanks to Sharlene Blake, Duane's assistant, who has helped me with so many administrative matters, and with whom I have shared so many enjoyable conversations. I wish you and your family the very best in life. Thanks to all my friends everywhere for their continuous support, and for constantly reminding me of how much more there is in life besides work. Many thanks also to all of my previous educators who paved the way for my successful tenure at MIT, especially Dr. Efrain Rosario. Not many girls from Las Marifas, Puerto Rico, have the chance of getting a Ph.D. from MIT. There are two people who deserve a lot of credit for helping me get here: Dr. Susan Cohen, who has been my mentor and role model since 1999, when I had the privilege of working for her at IBM in a summer internship, and who was the first person to suggest that I go to graduate school. Susan, thank you so much for all your motivation, guidance and encouragement. I will always be grateful for all you have done for me, and for your beautiful friendship. The second person is Dr. Keith Whittingham, who held my hand through the process of applying and getting a full fellowship for graduate school. Thank you very much for your help and mentorship. Last but not least, my family deserves all the credit for this accomplishment, as they have supported me every step of the way with complete selflessness. It has been that support that has made me strong enough to deal with all that life has to offer, good and bad. Many thanks to my amazing husband Andrew Gettings, the love of my life and who has made my life happier than I could have ever imagined. Thank you for your unconditional support and love. Graduate school would have been unbearable without you. Betsy, thank you for all your love, your beautiful friendship and all your support. I am so lucky to have you as my sister. Tatito, thank you for being such a great brother, for all your love, and for the beautiful Marcos, Kristina and Waleska. Papi, there are no words to thank you enough for all that you have done for me. Thank you for being the best father one could wish for, for the upbringing you gave me and for all your encouragement and support in everything I do. Mami, I am sure you are enjoying this moment from Heaven. I thank the Lord for letting me have the very best mother on Earth for 26 years. I miss you, but I feel comfort in knowing that you wanted me to reach this milestone, and hopefully I am making you proud. To all my family, this milestone is dedicated to you. Special thanks also go to Andy's parents, D. Anne and D. Larry. Thank you for taking care of me here in New England like your own daughter, and for all your warmth and encouragement. This research has been funded in part by the Samsung Electronics, the MARCO Interconnect Focus Center and the Center for Circuits, Systems and Solutions, by the IBM Ph.D. Fellowship, and by the Bell Laboratories Cooperative Research Fellowship Program. 6 A la memoria de mi madre y el amor de mi esposo, padre y hermanos. To the memory of my mother and the love of my husband, father and siblings. 7 8 Table of Contents CHAPTER 1 1.1 1.1.1 1.2 1.2.1 1.2.2 1.2.3 1.2.4 1.3 1.4 1.5 CHAPTER INTRODUCTION............................................................................................19 M otivation............................................................................................... Our Previous Work................................................................................. Review of Test Structures from Literature ............................................. Dedicated Variation Test Structures ........................................................ Ring Oscillator Variation Test Structures............................................... M ethods to Extract Full I-V Curves......................................................... M ethod to Measure and M odel Interconnects ........................................ Thesis Contributions ............................................................................... Thesis Outline .......................................................................................... Sum m ary ................................................................................................. 2 TEST CHIP ARCHITECTURE ..................................................................... 19 20 22 23 24 26 28 30 31 32 33 2.1 2.2 2.3 2.4 2.5 2.5.1 2.6 2.7 2.8 M ethodology ............................................................................................. Tile .......................................................................................................... Row .......................................................................................................... Chip..............................................................................................................38 Feedback Im plem entation........................................................................ Operational Amplifier Design .................................................................. Implem entation of Low Leakage Switches............................................. Sim ulation Results ................................................................................... Sum m ary ................................................................................................. 39 40 43 46 49 CHAPTER 3 TEST STRUCTURES DESCRIPTION ............................................................ 51 3.1 3.1.1 3.1.2 3.1.3 3.1.4 3.2 3.2.1 Front End of Line Structures.................................................................... NM OS and PM OS Structures of Different Sizes.................................... Orientation .............................................................................................. Polysilicon Spacing .................................................................................. Number of Polysilicon Fingers ............................................................... Back End of Line Structures ................................................................... Charge Based Capacitive Measurement Methodology and Im plem entation........................................................................................ Plane Structures ....................................................................................... Comb Structures...................................................................................... Grid Structures ........................................................................................ Sum m ary ................................................................................................. 51 51 52 53 54 55 3.2.2 3.2.3 3.2.4 3.3 CHAPTER 4.1 33 34 37 55 59 59 60 60 4 RESULTS AND ANALYSIS FOR LAYOUT-DEPENDENT TEST STRUCTURES............................................................................................ 63 Functionality ............................................................................................ 63 9 Electrical Repeatability ........................................................................... 4.1.1 Leakage Currents ...................................................................................... 4.1.2 Variation Analysis ................................................................................... 4.2 W ithin Chip and Chip to Chip Components ............................................. 4.2.1 W ithin Chip Variance.........................................................................67 4.2.1.1 4.2.1.2 Chip to Chip Variance ...................................................................... 4.2.1.3 Total Variance ................................................................................... 4.2.1.4 Aggregated Variance ........................................................................ 4.2.2 Normalized Variation................................................................................68 Evaluation of Mean Differences ............................................................. 4.2.3 4.2.4 Evaluation of Variance Ratios ................................................................. Layout Dependent Variation Analysis.................................................... 4.3 Orientation ............................................................................................... 4.3.1 4.3.1.1 Variance Components and Variation................................................ 4.3.2 Polysilicon Spacing.................................................................................. 4.3.2.1 Variance Components and Variation..................................................80 Number of Polysilicon Fingers ............................................................... 4.3.3 4.3.3.1 Variance Components and Variation..................................................86 Summary ................................................................................................. 4.4 CHAPTER 64 65 65 66 67 68 68 69 70 70 70 77 77 81 86 5 RESULTS AND ANALYSIS OF TEST STRUCTURES OF TRANSISTORS OF DIFFERENT SIZES .................................................................................... 89 89 Functionality ............................................................................................ 5.1 93 Transistors of Different Sizes at Saturation ............................................. 5.2 99 Transistors of Different Sizes with Voltage Sweeps ............................... 5.3 Transistors of Different Sizes with Gate Voltage Sweeps........................99 5.3.1 5.3.2 Threshold Voltage Variation Analysis.......................................................102 NM OS Threshold Voltage...................................................................103 5.3.2.1 108 PM OS Threshold Voltage ................................................................... 5.3.2.2 114 Summary .................................................................................................... 5.4 CHAPTER 6 RESULTS AND ANALYSIS OF INTERCONNECT TEST STRUCTURES ........... 115 6.1 6.1.1 6.2 6.3 6.3.1 6.4 6.5 6.6 Measurement M ethod Review and Adjustments ....................................... 115 117 Electrical Repeatability .............................................................................. Results............................................117 Overview of M easured Capacitances 119 Comb Capacitances Results and Analysis ................................................. 120 ..................... and Thickness Spacing Finger Between Variation Analysis Plane to Plane Capacitances Results and Analysis .................................... 121 Grid Capacitances Results and Analysis....................................................122 125 Summary .................................................................................................... CHAPTER 7 SPATIAL ANALYSIS .................................................................................... 7.1 7.2 127 Chip to Chip Spatial Analysis....................................................................127 Within Chip Spatial Analysis.....................................................................130 10 7.3 CHAPTER 8.1 8.2 8.3 8.4 8.5 CHAPTER 9.1 9.1.1 9.1.2 9.2 9.2.1 9.2.2 9.2.3 9.2.4 9.2.5 9.3 Summary .................................................................................................... 134 8 SUGGESTIONS FOR FUTURE IMPLEMENTATIONS ..................................... 137 Gate Leakage in Newer Technologies ....................................................... 137 Design of the On-Chip Feedback...............................................................138 Design Review of FEOL Low Leakage Switches ..................................... 138 Off Tiles Leakage Currents........................................................................140 Summary .................................................................................................... 141 9 CONCLUSIONS............................................................................................143 Thesis Contributions .................................................................................. 143 Test Chip Design and Implementation.......................................................143 Testing Results and Analysis of the Test Structures..................................144 Future W ork ............................................................................................... 146 On-chip M easurements .............................................................................. 146 Study of Sub-threshold Variation .............................................................. 146 Charge Based Capacitance Measurement Structures.................................146 Coupling Capacitances Structures ............................................................. 147 Directly Probed Transistors ....................................................................... 147 Summary of Contributions.........................................................................148 REFERENCES...................................................................................................................149 APPENDIX I TEST STRUCTURES NOMENCLATURE AND Row LOCATION ..................... 153 APPENDIX 11 CALCULATION FOR CROSSOVER CAPACITANCE......................................161 APPENDIX III LAYOUT SCREEN SHOTS..........................................................................163 11 12 List of Figures Figure 1-1: Agarwal et al. VT mismatch test structure [8]. ........................................... 23 Figure 1-2: Bhushan, Ketchen at al. ring oscillator stage to measure AVT [14] ........... 25 Figure 1-3: Transistor array test structure proposed by Quarantelli et al. [18].............27 Figure 1-4: Lefferts architecture [4]...............................................................................28 Figure 1-5: C B C M structure [22]......................................................................................29 Figure 2-1: Traditional (a) method of testing vs. proposed (b). .................................... 34 Figure 2-2: Tile block diagram for FEOL test structures. ............................................ 35 Figure 2-3: Tile block diagram for BEOL test structures............................................. 35 Figure 2-4: Tile layouts. ................................................................................................ 36 Figure 2-5: Row block diagram.................................................................................... 37 Figure 2-6: Chip block diagram.................................................................................... 38 Figure 2-7: Die photo of full test chip .......................................................................... 39 Figure 2-8: Feedback network for VDD regulation. ....................................................... 40 Figure 2-9: Feedback implementation.......................................................................... 40 Figure 2-10: Opamp architecture................................................................................. 41 Figure 2-11: Opamp frequency response...................................................................... 42 Figure 2-12: Step response of the opamp. ..................................................................... 42 Figure 2-13: Low leakage switch (left), when enabled (center) and when not enabled (right)........................................................................................................ . . 43 Figure 2-14: RON resistance for enabled switch. ........................................................... 44 Figure 2-15: Schematics for low leakage switch...........................................................45 Figure 2-16: Simulated data VDS sweep for NMOS 0.50 jim /0.18 gm device.............46 Figure 2-17: Model for simulation of row.................................................................... 47 Figure 2-18: NMOS probed alone.................................................................................47 Figure 2-19: Simulated opamp limitations for NMOS 0.50 gm /0.18 gm device. ....... 48 Figure 2-20: Scan chain functionality . ......................................................................... 49 Figure 3-1: Different size FEOL test structures, with different channel lengths (L) and w idths (W ).......................................................................................... 52 Figure 3-2: Orientation FEOL test structures............................................................... 53 Figure 3-3: Poly finger spacing FEOL test structures. .................................................. 54 Figure 3-4: Number of poly fingers FEOL test structures.............................................54 Figure 3-5: Interconnect capacitances - 1. Fringing, 2. Coupling, 3. Plane..................55 Figure 3-6: CBCM structure with switches.......................................................................56 Figure 3-7: BEOL switch schematic. ........................................................................... 58 Figure 3-8: Layouts of BEOL tiles............................................................................... 58 Figure 3-9: BEOL plane structure.................................................................................59 Figure 3-10: BEOL comb structures. ........................................................................... 60 Figure 3-11: Grid test structures................................................................................... 60 Figure 4-1: Gate voltage sweeps at 1.8V VDS for all FEOL test structures, NMOS and PMOS, for one row of a single test chip (728 structures total)...........64 Figure 4-2: Representation of chip-to-chip versus within-chip variation..................... 66 Figure 4-3: Orientation effect for NMOS and PMOS transistors..................................71 Figure 4-4: Orientation effect for NMOS and PMOS transistors with confidence intervals for the mean................................................................................. 73 13 Figure 4-5: NMOS orientation effect on a chip by chip basis......................................74 74 Figure 4-6: PMOS orientation effect on a chip by chip basis. ...................................... Figure 4-7: Current difference between adjacent NMOS vertical-horizontal test . 76 structures. ................................................................................................. Figure 4-8: Current difference between adjacent PMOS vertical-horizontal test . 76 structures. ................................................................................................. (right)..................78 Figure 4-9: Poly finger spacing effect for NMOS (left) and PMOS Figure 4-10: Mean(left) and normalized variation (right) in saturation current for NMOS transistors with different number of polysilicon fingers. Statistics for N=560 DUTs consisting of 16 replicates per chip, for 35 chips..........83 Mean(left) and normalized variation (right) in saturation current for for 4-11: Figure PMOS transistors with different number of polysilicon fingers. Statistics for N=496 DUTs consisting of 16 replicates per chip, for 31 chips..........83 85 Figure 4-12: M OSFET cross section ............................................................................. Figure 5-1: VDS sweep for VGS ranging from 0-1.8V in 0.1V intervals.......................90 Figure 5-2: Semi-log plot of a VDS sweep for VGs ranging from 0 to 1.8 V in 0.1V 91 intervals, showing the limitation of our measurements. ........................... V in 0.IV to 1.8 Figure 5-3: Semi-log plot of a VDS sweep for VGS ranging from 0 intervals. The opamp's ground voltage is set at -100 mV..........................91 Figure 5-4: Comparison between simulated and measured data for a VDS of 1.8 V, ranging from 0 to 1.8 V in 0.1 V intervals, for eight replicates of one 93 type of transistor in one chip (N=8). ......................................................... Figure 5-5: Variation in NMOS currents at VDS=1.8V and VGS=1.8V, N=280 (8 replicates in 35 chips)................................................................................. 94 Figure 5-6: Variation in PMOS currents at VDS=1.8V and VGS=0 V, N=248 (8 94 replicates in 31 chips)................................................................................. Figure 5-7: Relative current variation in NMOS transistors as a function of device 5 are a.................................................................................................................9 device of as a function transistors in PMOS Figure 5-8: Relative current variation 6 are a.................................................................................................................9 Figure 5-9: VGATE sweep for transistor in scan chain at VDS Figure Figure Figure Figure Figure Figure Figure Figure Figure Of 0.1 V to 1.8 V, and 99 opamp ground set at -I00mV ................................................................... 5-10: Normalized variation (/[t) of the output current at different VGATE values, for devices with different channel lengths. Measurements of one test chip, N=176 (16 replicates of 11 different gate widths at a fixed gate 10 0 len gth ). ......................................................................................................... 5-11: Variation of a simulated transistor. The variance is calculated from the curve of a transistor in a typical simulation corner and another transistor w ith a 5% variation in VT.............................................................................101 5-12: NMOS VT error-bar plots, ±1-y bars shown. N=480.................................103 5-13: Variation in NMOS threshold voltage........................................................106 5-14: Linearity of NMOS threshold voltage according to size............................106 5-15: PMOS VT error-bar plots, ±1-a bars shown. N=416..................................109 5-16: Variation in PMOS threshold voltage.........................................................112 5-17: Linearity of PMOS threshold voltage according to size.............................112 6-1: Measurements of interconnect test structures...............................................118 14 Figure 6-2: Capacitive loads to study automated simulation-based extraction tool crossover capacitance extraction..................................................................123 Figure 7-1: Location of test dies within the wafer. Shaded squares indicate the dies available for spatial measurement and analysis. .......................................... Figure 7-2: Spatial mean currents (left) and normalized standard deviations (right) per chip across wafer for NMOS 3-finger structure. N=32. ........................ Figure 7-3: Spatial saturation mean currents (left) and normalized standard deviations (right) per chip across wafer for all NMOS of different sizes. N = 5 2 8 ........................................................................................................... Figure 7-4: Spatial saturation mean currents (left) and standard deviations (right) for threshold voltage of NMOS transistors with 0.18 gm gate length. N=176. Figure 7-5: Within-row spatial analysis for most replicated NMOS, showing mean current values for the sixteen replicates per row for 35 chips. N=35. ......... Figure 7-6: Within-row spatial analysis for most replicated NMOS, showing 1-sigma current values for the sixteen replicates per row for 35 chips. N=35. ......... Figure 7-7: Within-row spatial analysis for most replicated NMOS, showing saturation current values for the sixteen replicates per row in a specific chip . N =1 ...................................................................................................... Figure 8-1: Feedback implementation for technologies with high gate leakage currents.........................................................................................................137 Figure 8-2: Revised NMOS and PMOS FEOL tiles........................................................139 Figure 8-3: Problem with FEOL switches.......................................................................140 127 128 12 9 129 131 132 134 15 16 List of Tables Table 2-1: Opamp dimensions and specifications........................................................41 Table 2-2: Variation comparison between probing DUT directly and with the switch .... 45 Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table Table 45 2-3: Sw itch transistors sizing............................................................................... 3-1: Sizes of test transistors, NMOS and PMOS.................................................52 3-2: CBCM measurements capabilities...................................................................57 58 3-3: BEOL switch transistors sizing.................................................................... 4-1: Bias conditions for layout dependent variation analysis.............................70 4-2: Tabulated orientation effect for NMOS and PMOS transistors...................71 72 4-3: Two-way ANOVA for NMOS with different orientation. ......................... 4-4: Two-way ANOVA for PMOS with different orientation ............................ 72 4-5: Variances and variation for orientation test structures.................................77 4-6: Tabulated poly spacing effect for NMOS and PMOS transistors................78 4-7: Two-way ANOVA for NMOS with different finger spacing......................79 4-8: Two-way ANOVA for PMOS with different finger spacing.......................79 4-9: Variances and variation for polysilicon spacing test structures...................80 4-10: Tabulated poly spacing effect for NMOS and PMOS transistors..............82 84 4-11 :Variance ratios of NMOS multiple fingers structures. ............................... 84 4-12: Variance ratios of PMOS multiple fingers structures. .............................. 4-13: Variances and variation for polysilicon-fingers test structures..................86 5-1: Variances and variation in saturation current, VDS=VGS=1.8 V, for NMOS 97 transistors of different sizes, N=560. ......................................................... 5-2: Variances and variation in saturation current, VDS=1.8 V and VGS=0 V, 98 for NMOS transistors of different sizes, N=496. ....................................... 5-3: Tabulated NMOS VT calculation for devices of different sizes. ................... 104 5-4: Three-way ANOVA for NMOS threshold voltage........................................105 5-5: Variances and variation for V1 in NMOS transistors of different sizes. 10 7 N =4 80 ........................................................................................................... 5-6: Tabulated PMOS VT calculation for devices of different sizes.....................110 5-7: Three-way ANOVA for NMOS threshold voltage........................................111 5-8: Variances and variation for VT in PMOS transistors of different sizes. 1 13 N =4 16 ........................................................................................................... capacitances. measured and by simulation extracted 6-1: Comparison between 1 19 N = 5 60 ........................................................................................................... 6-2: Two-way ANOVA for comb capacitances made with different metal 12 0 layers. ........................................................................................................... 6-3: Variation (1-sigma/mean) in comb test structures. ........................................ 121 6-4: Variance ratios of coupling test structures.....................................................121 6-5: Two-way ANOVA for plane to plane test structures.....................................122 6-6: Two-way ANOVA for grid test structures.....................................................125 17 18 Chapter 1 Introduction Aggressive technology scaling raises the need for efficient methods to characterize and model circuit variation at both the front and back end of line, where critical parameters such as threshold voltage and parasitic capacitance must be carefully modeled for accurate circuit performance. This thesis addresses this need by proposing a test chip that is capable of extracting parameters and distributions of transistors and interconnects. This introductory chapter will first look at the motivation of our research in Section 1.1. It continues by describing other work in the field in Section 1.2, focusing in four main areas: dedicated test structures for variation extraction in a specific parameter, ring oscillators to detect parameter variation by comparing frequencies, test circuits that enable generalized measurement of analog characteristics, and a description of the charge base capacitance methodology to measure interconnect parasitic capacitances. We present the contribution of this thesis in Section 1.3, and in Section 1.4 we outline the structure of the rest of the thesis. Finally, in Section 1.5 we summarize this chapter. 1.1 Motivation The study of process variations has greatly increased in importance due to aggressive technology scaling. Previous research has shown the substantial impact that process variations in front end of line structures have in reducing yield in integrated circuits [1], [2]. Robust circuit design depends on a more complete characterization of these variations and their impact on circuit-level parameters, as the degradation in the precision of device and interconnect parameters has elevated variability to a first order limitation to continue technology scaling [3]. 19 This thesis seeks to address this issue by developing a methodology capable of testing a large number of front-end-of-line (FEOL) and back-end-of-line (BEOL) structures, and modeling variations in threshold voltage and leakage currents, among other parameters. This is achieved by designing and implementing test circuits that include a large number of high performance devices (devices under test or DUTs) controlled by low leakage switches and sensors to ensure a nominal bias condition at the DUT terminals. This architecture is partially based on the test structure proposed by Lefferts [4]. By accessing analog characteristics of a large number of DUTs, it is possible to gather the statistics necessary to identify and model these variations, and prevent them from contributing to performance failure. Through this we expect to provide a replicable methodology so that the effect of variation sources may be quantified in different technologies. This thesis studies variations in circuits due to the one of the fundamental sources of variations in integrated circuits, as noted by Nassif [5]: physical factors (e.g. polysilicon dimension variation) arising from the fabrication process. Physical factors can fall into two categories: die-to-die physical variations and within die physical variations. Analysis includes separation of spatial, layout dependent, and random variation components within die and as a function of wafer location. 1.1.1 Our Previous Work Our group has previously built a test chip to study the impact of layout dependent variations on circuit performance [1]. The test chip consists of a range of differently designed ring oscillator circuits which are made sensitive to some of the most common and critical layout-dependent variations. Analysis of the ring oscillator output frequencies enables assessment of variation impact at both the device (FEOL) level and interconnect 20 (BEOL) level. FEOL variations studied are dense vs. isolated polysilicon gates, polysilicon finger proximity, global polysilicon density, orientation and others. BEOL variations studied include parasitic coupling, fringing and planar capacitance, among others. The testing of the ring oscillator based test chip has proven successful for both device and interconnect test structures. Different ring oscillator layout practices are seen to result in significant differences in mean and standard deviation of measured ring oscillator frequencies. Increases in variance of nominally identical structures due to layout practice are ascertained with 95% confidence level tests on variance ratios. ANOVA is performed to demonstrate that the means of different structures are different at a high (over 95%) confidence. The FEOL structure analysis shows strong dependencies between the layout practice and gate length variations: spacing between poly fingers can shift ring oscillator frequency by 4.4%, and polysilicon density can change frequency by 2.1%. BEOL structure analysis shows dependencies due to the metal geometry. Spatial analysis reveals both a large die-to-die (within-wafer) trend, and systematic within-die spatial patterns for particular test structures. The analysis of the ring oscillator test chip shows several areas where improvements are needed. For example, threshold voltage variations were suspected but not proven or quantified for the FEOL structures. A means to identify specific device variations which cause circuit variation is needed. In addition, BEOL structures need an improved technique for measurement so that we can improve accuracy and separate interconnect capacitance components, in order to characterize interconnect variations. In this thesis we target these and other issues inspired by the analysis of our previous test chip. 21 1.2 Review of Test Structures from Literature Semiconductor device or circuit test structures have been in use since the 1960's [6], with publications dating back to 1974 reporting the use of test structures to accept or reject wafers. Research in the area of characterization of technologies using more efficient methods than probing each transistor independently has increased greatly in the last few years [6], likely due to the non-idealities of device models because of the aggressive scaling that the semiconductor industry has kept following Moore's Law [7]. Broadly speaking, three approaches can be identified for front end of line variation analysis. In the first family of structures, dedicated test structures or circuits focus on an extraction of variation in a specific parameter of substantial importance, such as MOSFET threshold voltage. These test circuits are often large arrays to gather the statistics necessary and identify variation dependencies. A second approach is to use ring oscillators as the test structure, and study ring oscillator (RO) frequency variation in structures that are made to be more or less sensitive to particular layout or device parameters. These RO test structures may be used in large arrays, or may be small enough that they can be distributed as monitors in scribe lines or within product chips. Finally, the third test circuit family seeks to enable generalized measurement of device or interconnect response, using pad sharing or on-chip digitization approaches. In the following subsections, we illustrate each of these three approaches with examples from the literature. The work presented in this thesis pursues the third approach, to enable the relatively rich characterization of a large number of test devices through current-voltage (I-V) measurements using a scan chain based pad sharing architecture. The fourth subsection illustrates a methodology to measure interconnect 22 parasitic capacitances, which we also implement in this thesis. 1.2.1 Dedicated Variation Test Structures There has been extensive work through many years to model variation in front end of line (FEOL) devices. The accurate modeling of the variation in these devices is important to quantify design corners and not over- or under-estimate worst case scenarios, both within chip and between chips [5]. In this section we present a few examples of these types of test structures, all focusing on a specific variation extraction. Force drain - Force gate (V) IP 1 V. MUT Device Array - -w . HE k _d d T HE ----*------------- ---- - - Source fdlower sense drain Figure 1-1: Agarwal et al. VT mismatch test structure [8]. Agarwal et al. [8], [9] have modeled mismatches between adjacent SRAM devices, concluding that the large variations observed in the extracted threshold voltage statistics indicate that the random dopant fluctuation is the reason behind mismatch in the adjacent devices. Their further research [10] includes a test chip to measure threshold voltage mismatches of adjacent transistors. The structure is described in Figure 1-1. Once a row and column is activated, the applied current IDS flows through the device under test, and the threshold voltage shift of the device is measured by sensing the corresponding source voltage through the sense source line. This technique is powerful because the mismatch can be measured with a single measurement, but the reference VT (which is only one 23 device) must be fully characterized beforehand. Drego et al. [11] designed a highly dense test chip also, to measure scattered VT variation. With approximately 140,000 transistors, their structure can map variations in threshold voltage by comparing two currents, with which they can extract the needed parameters to isolate AVT in subthreshold, taking advantage of the subthreshold current characteristics where AVT can be dominant. One may also characterize the actual VT value of one device per chip for chip-to-chip comparison of AVT. 1.2.2 Ring OscillatorVariation Test Structures A ring oscillator continues to be an attractive type of test structure because of its simple implementation and single output. Back in 1984, Cassard [12] presented a ring oscillator test chip to measure within chip and chip-to-chip variation, and compared these test results with SPICE simulations. The test structures were accessed using a probe card, and the experiments were able to conclude that the polysilicon gate width was indirectly proportional to the frequency and was the primary cause of the frequency variation. Cassard also concluded that intra-chip parameter variations can influence SPICE sensitivity. Today ring oscillators are a commonly used test structure, as they only require one (start up) or no (self oscillating) input signals and can provide a simple frequency readout. A ring oscillator can be easily digitally controlled, making it a prime candidate for a large array, such as we did in our previous work [1]. Bhushan, Ketchen et al. [13], [14], [15], have done extensive work using ring oscillators to identify the effect of variations in a number of specific MOSFET parameters. The measurements represent an average of hundreds of identically designed MOSFETs, therefore being resistant to random statistical fluctuations in the parameters 24 of isolated single MOSFETs. For example, part of their work [14], [15] includes a test chip capable of isolating AVT. The circuit schematic of a single stage in the RO is shown in Figure 1-2. The full ring oscillator includes four of the stages pictured in Figure 1-2 and a two-input NAND gate as an enable, for a total of five stages. The structures were placed in scribe lines, not consuming any of the area already reserved for non-test chips in the wafer. The delay of the passgate is strongly modulated by threshold voltage, and, because of sizing, the switching delay is designed to be highly sensitive to the characteristics of the passgate. The variation of threshold voltage can then be obtained by measuring the difference in frequencies among several ring oscillators. This is an efficient method of modeling AVT as there is no need to individually characterize any of the individual devices. However, we can only gather the variation out of it, and not an actual threshold voltage value. PFET Keeper NFET Passgate Inverter Load Figure 1-2: Bhushan, Ketchen at al. ring oscillator stage to measure AVT [14]. The work from Bhushan and Ketchen extends to ring oscillators beyond standard MOSFET dc characterization [16]. They also use ring oscillator test structures that perform in the multiple GHz range, so that they are representative of today's microprocessors. With various ring oscillator designs they can measure a range of parameters, from resistance distributions to sub-picosecond delay measurements, all with 25 simple 1/0 requirements so that they are compatible with standard in-line parametric testers. Pang et al. [17] designed, implemented and tested a test chip similar to our previous work as presented in Section 1.1.1, with ring oscillators that enhance variation due to layout practices. The delay in the ring oscillators is characterized through the spread of ring oscillator frequencies and the transistor leakage is measured by using an on-chip analog to digital converter, which enhances the test structure ability when compared to [1]. With these functionalities they are able to conclude that regular layout reduces the systematic layout-dependent process variations, something that was not captured by automated simulator extractors. 1.2.3 Methods to Extract Full I-V Curves In the previous subsections we highlighted the importance of using test structures to extract variation. However, because they focus on variation in specific device parameters, those test structures are not able to provide full analog I-V curves, and only a few or no parameters can be measured from them. Other test structures, like the one presented in this thesis and the ones described below, can gather enough analog characteristics to not only measure variation in a test chip (due to large repeatability), but also extract the set of parameters needed to generate models for the transistors. Quarantelli et al. [18] proposed a test structure based on transistor arrays, as seen in Figure 1-3. The structure is small to minimize the effect of inter and intra die variation. The structure consists of columns of transistor arrays, with multiplexed gates and shared sources throughout the entire structure, and shared drains per column. The rows are multiplexed by a self-resetting scheme controlled by off chip circuitry. Saxena et al. [19] 26 built the proposed structure to analyze layout induced variation, for example, proximity between polysilicon fingers in stacked gates. They analyzed the I-V characteristics at only a few critical points: off state current, drive current and threshold voltage. They successfully measured and analyzed this data and arrived at conclusions on layout-based variation consistent with [1]. The structure provides a good assessment of layout induced variation, but it utilizes off chip control circuitry and the data gathering is limited to a given number of bias conditions. Col #1 - CoI#2 CoI#3 CoI#4 1A Row #2 I 022 CI aX CD LRo Reset__ #32F t CLK Figure 1-3: Transistor array test structure proposed by Quarantelli et al. [18]. Lefferts et al. [4] proposed a test structure architecture to characterize and monitor a CMOS 0.25 gm technology process. Their objective was to do this in a small area so it could fit in scribe lines, and avoid using the area reserved for non-test chips in the wafer. Lefferts' 0.25 jim test structure architecture consists of 340 test devices controlled by 14 probes, as seen in Figure 1-4. The probes provide access to the drain and source voltages of the DUTs, and to guarantee accuracy in testing they both have full Kelvin connections to eliminate the IR drop down the long test row. Testing was performed using an off chip 27 semiconductor analyzer for the drain current measurement and voltage sensing, and an off-chip operational amplifier (opamp) ensured a desired drain voltage (nominally ground). 4- 14 Probe Paas 0 000 -p 75 Devices - - 0 0 0 0 3 00O DUT 09* On Probe Card Figure 1-4: Lefferts architecture [4]. With his test structures, Lefferts successfully modeled both device and interconnect variation, with accuracy that goes beyond conventional test structure testing because of the Kelvin sensing. However, the several off-chip elements needed to perform the testing make this architecture inconvenient to the user. In addition, the design of the switch that turns the devices on and off is not disclosed. Wang and Shepard [20] designed a test chip to gather full analog characteristics onchip and with on-chip current measurements. They built a prototype in 0.25 Im technology as proof of concept, where they were able to measure currents in both saturation and sub-threshold for a given voltage range. This type of test structure architecture is powerful because they are able to do on-chip current measurements, but it is quite complex because of the additional circuitry needed for the current measurements. 1.2.4 Method to Measure and Model Interconnects Physical sources of variation also include wire or interconnect variations. Nassif [5] 28 --- Nxa- __ - ____ - __ __ - Mwmffl - --- concludes that wire variations are expected to increase in importance dramatically in future IC technologies, and appropriate methods of modeling these variations must be developed to keep up with technology scaling. Proper characterization of interconnects is increasingly important as technology scales, as interconnects are increasingly affected with scaling [21], and this can give rise to timing errors, voltage degradation and other problems likely to affect both analog and digital circuit design. Charge-based capacitance measurement (CBCM) techniques [22] provide sufficient accuracy to model many of the interconnect sources of variation. NWELL VDD(No Cap) AL VDD (Cap) A' V2 V1 Figure 1-5: CBCM structure [22]. Chen et al. [22] describe the CBCM structure, pictured in Figure 1-5, as follows: the structure on the left is identical to the one on the right in every manner except that it does not include the target capacitance to be characterized. Both structures are driven by two non-overlapping signals. When the PMOS transistor turns on, it will draw charge from VDD to charge up the target interconnect capacitance. This amount of charge will then be subsequently discharged through the NMOS transistor into ground. By measuring the 29 current at the source of the PMOSFET, the DC value of this charging current can be measured. The difference between the two DC current values is used to extract the target interconnect capacitance, determined by I'-I=Inet and Ine =C Vdd-f, where C is the target capacitance andf the frequency in which the structure charges and discharges. The CBCM technique has been extended to a fully differential approach and implemented by [23], [24], and [25]. We use the differential method in this thesis, and its design and implementation are explained in Section 3.2.1. 1.3 Thesis Contributions The contributions of this thesis focus in two main areas: the design of a test structure capable of efficiently gathering analog characteristics of devices and interconnects, and a statistical methodology to study variation for both devices and interconnects. In order to address need for variation analysis of large sets of data to properly model distributions of devices and interconnects, we propose a test circuit that can access output currents and provide a range of input voltages to a number of devices under test. The devices under test are accessed using a scan chain approach, which use low leakage and low variation the switches to turn the DUTs on or off. This allows us to share terminals of the DUTs and therefore save considerable chip area. The devices will not be probed directly, and therefore we need to assure that their terminals are at the nominal voltage we intend them to be. For this we implement Kelvin sensing, so that the intended voltage at the terminals of the DUT is the desired one. The architecture is flexible enough to allow us to measure both front and back end of line test structures. FEOL test structures focus on typical design issues (e.g. orientation) and different sizes, and BEOL test structures focus on interconnect capacitances, which allows a rich set of data that enables 30 us to perform various types of variation analysis. The variation analysis uses methodologies including ANOVA and variance confidence interval ranges to identify which design practices or parameters can affect circuit performance the most. We also separate the impact of inter and intra die variation using nested variances, to not overestimate the chip to chip variation. Spatial analysis to map wafer and die trends are also accomplished. With the analysis and the large sets of data one can identify variation sources, quantify circuit impact, and specify layout practices for variation minimization. These contributions are distributed across the thesis, the organization of which is described in Section 1.4. 1.4 Thesis Outline This thesis work focuses on creating a test chip methodology that can efficiently model both device and interconnect variation, its implementation, testing and analysis. In Chapter 2 we describe the architecture of the test chip, including the building blocks of the test chip (tile, row and chip), design of an on-chip Kelvin sensing methodology, and low leakage switches. Chapter 3 describes the test structures that are implemented in the test chip, including both front and back end of line. In Chapter 4 we present the testing results and analysis of the front end of line test structures that focus on layout induced variation, including a description of the statistical analysis that is used throughout the rest of the thesis. Chapter 5 also includes testing results and analysis, but in this case focusing in variation due to the sizing of transistors. Chapter 6 focuses on the testing results and analysis of the BEOL test structures, and Chapter 7 on within-die and within-wafer spatial analysis. In Chapter 8 we offer suggestions for the implementation of the test chip 31 methodology in other technologies. Conclusions and other suggestions for future work are presented in Chapter 9. 1.5 Summary In modem integrated circuits it is becoming increasingly important to detect and model process variations for critical parameters, at both the front and back end of line. This thesis contributes a methodology to gather full I-V curves for transistors and to measure interconnect capacitances using a differential CBCM approach, to not only model variation, but also extract device parameters. In this chapter we summarized and commented on the features of several different types of test structure architectures. With our design we intend to take advantage of the positive features that these architectures have taught us and contribute new extensions to the methodology for variation test circuits and variation analysis. The following chapter describes the architecture of this test chip, and further chapters discuss the analysis of our methodology and its test structures. 32 Chapter 2 Test Chip Architecture The architecture for the test chip, referred to as the "Analog Multiplexed Characteristics" chip, or AMC, is designed similar to the V-2 chip as in [1] and [2], emphasizing hierarchy, regularity, and repetition in a scan chain approach. This approach helps to transfer the test chip design to other technologies using automated tools. The hierarchical approach of the test chip also allows the user to choose pieces of it depending on the area they have available or other limitations. This chapter describes the architecture of the test chip, including a description of the methodology used (Section 2.1), and its building blocks: the tile (Section 2.2), row (Section 2.3) and the overall chip (Section 2.4). Two of the most important parts of the design are the implementation of low variation switches that turn the Devices Under Test (DUTs) on and off, and the implementation of a feedback network that allows the user to generate IV curves without having to probe the drain voltage of each DUT. Section 2.5 describes the feedback network, while Section 2.6 explains in detail the design of the switches. Simulations results are shown in Section 2.7, and an overall summary is presented in Section 2.8. 2.1 Methodology This work contributes a methodology to effectively model and characterize IC fabrication processes by measuring test structures using an architecture that multiplexes a small number of pads to measure a large number of devices under test. The DUTs are composed of front-end-of-line structures (MOSFETs) and interconnect structures attached to charge-base capacitance measurement (CBCM) structures. The architecture is 33 based in part on Lefferts' test chip architecture [4], and on our previous scan chain based architecture [1], [2]. Traditionally, detailed device I-V measurements are obtained by connecting each terminal to a pad, from which the appropriate signals and/or measurements will be applied, as illustrated in Figure 2-1 (a). This results in large amounts of wasted area in the test chip. The proposed methodology, illustrated in Figure 2-1 (b), uses a small number of pads for an entire row of test structures. The test chip has 1456 FEOL structures and 672 BEOL structures, for a total of 2128 structures, all controlled by 33 pads, and the user is still able to gather traditional DC analog characteristics. The chip is composed of mainly three building blocks: the tile, row, and chip blocks. Each of these are described in detail in the subsequent sections. (b) (a) Figure 2-1: Traditional (a) method of testing vs. proposed (b). 2.2 Tile The tile sub-block is the basic building block of the chip. There are two different types of tiles, the FEOL tile and the BEOL tile. Each of the FEOL tiles contains a DUT, switches, and a flip flop (DFF) that works as part of a scan chain, as shown in Figure 2-2. A shift register controls the switches that 34 turn the DUT on or off. The output of the DUT is wired to the output bus through a switch. Section 2.6 describes the design and implementation of the switches, to obtain low variation and low resistance control of each DUT. Vgate Switch en /en Output/Ground Switch en /en Data Scan Clock - en Reset Figure 2-2: Tile block diagram for FEOL test structures. Each of the FEOL tiles has the following input signals: scan-clock (the clock for the shift registers), data signal (the control signal for the DFF), reset (resets the registers), Vgate (gate voltage controller for DUT), and power. Each tile has one overall ground/output bus signal. Vii Switch V2i Switch neti netj Switch Vlj Switch V2j en /en en /en Output/Ground en /en Data Scan Clock D > Q -en -/en Reset Figure 2-3: Tile block diagram for BEOL test structures. 35 The BEOL DUTs have a different tile because its implementation requires many more switches, and the area must be bigger because of the metal capacitance structures that are attached to them. Figure 2-3 shows a block diagram of the BEOL tile. Each of the BEOL tiles has the following input signals: scan-clock (the clock for the shift registers), data signal (the control signal for the DFF), reset (resets the registers), four different Vgates (gate voltages controller for CBCM structure), and power. Each BEOL tile has have one overall output bus signal. FEOL Tile BEOL Tile Figure 2-4: Tile layouts. Figure 2-4 illustrates the layout of the tiles. The FEOL tiles have separate supply voltages for DUT, switches and DFF, and separate ground lines for the switch and DFF. The dimensions of the switches are 15 gm wide by 12 gm tall, the ones for the DFF are 16 pm wide by 6 gm tall, and the overall FEOL tile measures 18 gm wide by 21 jim tall. 36 The BEOL tiles also have separate supply voltages for the CBCM structure, switch and DFF. There are separate ground lines for the switch and DFF, and four separate gate voltage lines to control the CBCM structure. The overall BEOL tile measures 18 gm wide and 40 gm tall. The wires that complete the BEOL DUTs take the entire space of an additional tile. 2.3 Row A horizontal chain of tiles comprises the row sub-block. This array is shown in Figure 2-5. Tile Output/Ground Vsupply Fdck Feedb-ack - Vgate-~ Scan Clockp Reset Figure 2-5: Row block diagram. Each row has a feedback network in order to compensate through Kelvin sensing for the losses in the power line. There is a feedback network for the power and one for the ground line. The feedback network for the power line is entirely on chip, and its implementation is explained in detail in subsequent sections. The feedback network on the ground line is implemented off chip because the instrumentation needed to take current measurements (ammeter synchronized with Labview software) has a large input capacitance that would make our designed on-chip circuitry unstable. The FEOL and BEOL rows are different because BEOL rows need more signals to control the switches of the CBCM structures. There are two FEOL rows in the chip. There are 728 FEOL tiles in a row, including eight blank tiles to prove the functionality 37 of the scan chain. The row is subdivided in eight sub-rows. Each sub-row is of the horizontal length of the test chip (except for pad area) and contains one of each of the unique FEOL structures. All FEOL structures are repeated eight times per row (the three finger, minimum spacing test structure is repeated 16 times). The order in which the tiles are placed per sub-row varies, for a broader spatial analysis. There are two BEOL rows on the chip, each with eight sub-rows, as in the FEOL case. These sub-rows have 42 tiles each, and they are about as long as each FEOL subrow. Each BEOL DUT is repeated three times per sub-row. As in the FEOL case, the order in which the tiles are placed per sub-row varies, for a broader spatial analysis. Appendix I lists the sequence of test structures in each row. Appendix III shows screen shots of the layouts of both the FEOL and BEOL rows. 2.4 Chip The chip is simply four rows put together: two FEOL and two BEOL rows, intercalated as in Figure 2-6. Each chip can access 2128 test structures. We received 35 individual bonded chips in 40-pin ceramic DIPs, for a total of 74,480 tiles accessible in our set of packaged chips. We also received five un-bonded chips, for a total of 10,640 tiles that are accessible by a probing station. Tile Figure 2-6: Chip block diagram. The fabricated chip for the 0.18 gm CMOS logic process is shown in Figure 2-7. The 38 measurements of each die are 3 mm x 2.33 mm. Figure 2-7: Die photo of full test chip. 2.5 Feedback Implementation The power lines to control the power supply and ground voltages of the test rows are long, as they go from one end of the test chip to the other. The lines are parallel connected for each sub-row in order to minimize losses in the lines; however, a Kelvin sensing approach will yield more accurate voltages for the desired DUT. This is critical in this architecture, as the power voltage can not be directly measured for each DUT, but its value must be known accurately in order to create IDS vs. VDS and vs. VGS curves, and other analog characterizations. Figure 2-8 shows the feedback network needed in order to keep the desired voltage accuracy. 39 -- Vref .. VdasIrain .. t Vdrop Figure 2-8: Feedback network for VDD regulation. The goal is to achieve the following: Vdrain(s) a(s) (s 1+ a(s) Vej(s) Vdra-(S) Vdrop(s) 1 for very large a(s) (2.1) 1 0 for very large a(s) 1+ a(s) (2.2) For this one can use an operational amplifier, as illustrated in Figure 2-9. I a(s) Vref Vdrain td Vdrop I ~0A Power Supply VGATEr ----------------- A Off chip due to large cap. o on Ammeter mee Figure 2-9: Feedback implementation. 2.5.1 OperationalAmplifier Design The operational amplifier (opamp) used in the feedback network is a classic two stage differential-pair and common source amplifier, with an output buffer to provide the necessary current for the DUTs. The architecture of the opamp is shown in Figure 2-10 [26]. 40 4 Qbia s IQ5 I Q6 I LL Q8 V] Q1 Q2 I+ A Rbias Out Cc Qc Q3 Bias Diff Input Q9 7 Q4 CS Gain Out Buffer Figure 2-10: Opamp architecture. Thick oxide transistors are used to minimize variation and glitches in the opamp, and also to provide high enough voltages in the output to drive the DUTs to 1.8V. It has to be carefully designed to provide as much current as the DUTs will need, while staying stable and minimizing offsets. The opamp is sized in order to fulfill these conditions, as shown in Table 2-1. Transistors Q1, Q2, Q5, Q6, Q7, Qbias Q3, Q4 Q8, Q9 Qc Other specifications VDD Vss Cc Rbias Dimensions Width [pm] 24 12 600 Length [pn] 0.4 0.4 0.4 1 0.4 Values 3.3 V 0V 200 fF 4.15 KQ Table 2-1: Opamp dimensions and specifications. 41 10K 0 0) Ao Max = 3.35K 1K- 100 10 -A 180 160 CL (I) 0 140 120 100 on Unity Gain Phase Margin: 56.60 60- 1 10 ''100O K 100K Frequency (log) [Hz] 1M lOM Figure 2-11: Opamp frequency response. The opamp has a maximum gain of 3.3K, and phase margin of 56.6 degrees, as shown in Figure 2-11. The frequency response is important to ensure correct functionality in the feedback network. A phase margin near 600 provides the stability needed for feedback. The closer to 600, the faster it will settle in time and the less ringing it will suffer. Figure 2-12 illustrates how quickly the designed (simulated) opamp settles and how little ringing it shows. 2.0 1.8 1.6 > 1.4 & 1.2 - 0 1.0 0.8 0.6 0.4 0.2 0 0 500n lu Time (s) 1.5u 2u Figure 2-12: Step response of the opamp. 42 The large size of the input transistors minimizes problems with mismatch and potential offsets. The large output buffer provides the currents that the DUTs need to turn on all the way to full saturation. 2.6 Implementation of Low Leakage Switches In this design, low leakage switches turn the DUTs on and off in a precise manner, so that the input and output signals of the DUTs can be shared. New technologies have low and sensitive threshold voltages, and to assure a DUT on and off state it must be controlled by robust circuits. The switches designed here achieve both low variation and low resistance, to minimally perturb the DUTs. The switches for the FEOL test structures are expected to manage a large range of currents, from sub-threshold to saturation. This design is based in part on concepts by Lefferts [27]. /en en /en en /en VGATE VGATE en /en VREF en en VREF /en VREF Figure 2-13: Low leakage switch (left), when enabled (center) and when not enabled (right). The switches are built with thick oxide, high VT devices to minimize their variation, and they follow the schematic shown in Figure 2-13. When enabled, the gate of the DUT accesses the voltage VGATE through a transmission gate, and the source connects to ground, which also functions as an output bus. When not enabled, the gate connects to ground and the source connects to the reference voltage of the opamp, double ensuring that the DUT will be off by grounding the gate and short-circuiting the drain and source. The drain of the DUT is connected to the output of the opamp at all times. 43 A problem with these simple switches is that when large amounts of currents go through the drain, the "on" resistance in the switch that connects the source connection to ground results in a small voltage offset from the ground potential, perturbing the bias condition from the desired state. As a result, the output current can be different than if it had been measured by probing directly. This is diminished by enlarging the gate width of the switch device, as shown in Figure 2-14. This transistor was made as big as possible as long as it could still fit within the tile (i.e., not be wider than the control block), as shown in Figure 2-4. /en VGATE 9en .i1 .)VGATE . . . en /en VREF Figure 2-14: RON resistance for enabled switch. Any variation in the switch could also possibly have an effect on the variation observed in the DUT measurement. Using high VT (thick gate oxide) and large devices in the switch, this variation can be minimized. Table 2-2 illustrates this by comparing simulation results on the variation observed when the DUT experiences VT variation and it is not connected to any switches ("Variation DUT probed directly" column), and when the DUT is connected to the switch and the DUT has no variation but the switch does ("Variation Switch" column). As shown in Table 2-2, when the devices that form a specific DUT (NMOS of minimum length, in this case) have their nominal VT varied by, for example, 20%, its standard deviation of the output current is of 39.4 gA When the devices of the switch 44 have their VT varied by the same amount (and the DUT has its nominal VT value), the effect of variation is about 130 times less, with a standard deviation of the output current of 0.3 gA. VT DUT probed directly [pA] a Switch WA] +20 3.94E-05 3.OOE-07 -20 4.11E-05 2.36E-07 +10 1.99E-05 1.17E-07 -10 2.03E-05 1.87E-07 +5 9.97E-06 5.13E-08 -5 l.OlE-05 5.13E-08 % Change Table 2-2: Variation comparison between probing DUT directly and with the switch The sizes of the transistors in the switch are shown in Table 2-3, following the schematic of Figure 2-15. I kJ Q[ 1 Q2 Q3 Q4 5Q6 Figure 2-15: Schematics for low leakage switch. Transistor Q1 Q2 Q3 Q4 Q5 Q6 Gate Width [im 2 2 2 100 2 2 Gate Length [pm] 0.35 0.30 0.35 0.35 0.30 0.35 Table 2-3: Switch transistors sizing. 45 2.7 Simulation Results Simulations of the current at the Ground/Out node of a row tile, composed of an opamp, 727 "off' tiles and one NMOS "on" tile (consisting of a DUT of a single finger gate with W/L=0.50 gm /0.18 pm) are shown in Figure 2-16. We see that the design is capable of gathering a wide range of IDS-VDS curves for a given DUT. When comparing these measurements to the traditional method of probing the individual NMOS transistor, the plot shows how the test DUT closely mimics the traditional method. 3.5 x 104 Simulated Data: DUT in Row vs. NMOS Alone -e-. 3- DUT in Row SNMOS Alone 2.52-2- 1.5- 00 0 -0.5 0 0.5 1 1.5 2 VDS [V] Figure 2-16: Simulated data VDS sweep for NMOS 0.50 pm /0.18 gm device. In Figure 2-16, the "DUT in Row" curves refer to the curves taken by simulating the circuit shown in Figure 2-17, plotting the current through the "Output/Gnd" line as the "lout" current and the voltage between "Power Supply" and "Output/Gnd" as the "Vds" voltage. The signals shown, "Power Supply," "Vgate," "Data," "Scan Clock," "Reset," 46 "VddSwitch," "VddCtrl," "GndCtrl," and "Output/Gnd" are controlled off-chip. The short circuit between VddSwitch and the Power Supply is also made off chip. POWER SUPPLY Vgate Data Scan Clock Reset VddSwitch Vdd Ctr Gnd Ctri On Tile Off Tile Off Tile Off Tile Off Tile Off Tile Output/Gnd Figure 2-17: Model for simulation of row. The "NMOS Alone" curves refer to those simulated by simply sweeping the voltages of a single NMOS W/L=0.50 [tm /0.18 pm device in Spice, with each terminal connected to a specified supply, as shown in Figure 2-18, mimicking the traditional method of probing a transistor with a separate voltage probe per terminal. VSUPPLY VGATE Output/Gnd Figure 2-18: NMOS probed alone. There are some limitations in the on-chip voltage control for DUT measurement. First, the opamp cannot provide a full output swing because of the output buffer. Figure 2-19 illustrates how 100 mV is our limit for VDS voltage sweep. This simulation includes a single "on" tile (an NMOS with dimensions 0.50 gm /0.18 pm), and 727 other "off' tiles. The leakage current of these off tiles is approximately 2 nA, and Figure 2-19 illustrates the second limitation: it is not possible to accurately measure currents less than 47 this 2 nA limit. However, 2 nA currents and under only occur at the very lowest bias voltages, and we are able to gather the majority of desired I-V characteristics despite these two known design limitations. Simulated Data: DUT in Row vs. NMOS Alone -2 10 10 - 10 -- 0 10 - 10 10 AAAAAAAAAn, 102 10 00.1 0.5 1 Vds [V] 1.5 2 Figure 2-19: Simulated opamp limitations for NMOS 0.50 pm /0.18 pm device. The functionality of the scan chain is also tested using a model like that in Figure 2-17, but with three tiles with devices of size: 1.75/0.18, 1.75/0.36, and 1.75/0.54 tm. Figure 2-20 shows that when the Data-in signal is high at the falling edge of the clock, the first tile will be turned on. If the Data-in signal is low by the next fall of the clock, then the first tile turns off and the second one turns on. Had the Data-in tile stayed high, then both tiles would be on. The process repeats at every falling edge of the clock until all tiles are addressed. 48 Scan Chain Functionality 1.2m Tile 1 800U 600u 400u Tile 3 200u 0_1.8- I I I I 1.6 1.4- G) 1.2 ~ 1.00.87 0 0.6 0.4 0.2 I I I I I I I I I I ------ Data In Clock - 00 1OOm 200m Figure 2-20: Scan chain functionality. In the test chip there are a number of "blank" tiles, or tiles with no DUT in them, to test the functionality of the scan chain, by verifying that no current is measured at the expected blank tile point in the control sequence. 2.8 Summary The analog multiplexed characteristics test chip, or AMC, is composed of three main building blocks: the tile, row, and chip blocks. The tile includes a device under test (DUT) that is connected to low leakage switches in both the gate and drain terminals, which are carefully designed to ensure the on and off states of the DUTs. The tile also includes control circuitry (scan chain) to open and close the switches. A row is a group of tiles, with an on-chip Kelvin sensing feedback system to guarantee the drain voltage that each DUT receives. The chip is the group of all rows, and there are four rows per chip: two front end of line rows and two back end of line rows. Each FEOL row has 728 tiles, and each BEOL row has 42 tiles each, for a total of 2128 test structures per chip. The 49 chip was fabricated using a 0.18 pm CMOS logic technology, with 35 packaged chips in 40-pin ceramic DIPs, for a total of 74480 tiles accessible by measurement of packaged chips. The devices under test implemented in these tiles are described in Chapter 3. 50 Chapter 3 Test Structures Description One of the advantages of the analog multiplexed characteristics test chip is that its architecture is generic enough to include a large array of test structures as devices-undertest in its tiles. This chapter describes the particular DUTs implemented in this test chip. The design of experiments for this test chip include transistors that are designed to accentuate one source of variation, in order to extract that variation and link it to circuit performance, and also transistors of different sizes from which parameters like threshold voltage can be extracted. Section 3.1 describes these structures, while Section 3.2 considers back-end-of-line, or interconnect dominated test structures. Section 3.3 is an overall summary of the test structure design of experiments. 3.1 Front End of Line Structures Front end of line, or FEOL, test structures include an array of different NMOS and PMOS transistors designed to provide information about the behavior of a device at different bias conditions, and to model variation due to different layout practices, as in our previous work [1]. The FEOL structures are divided into those with different sizes, different orientation, different spacing between polysilicon fingers and different number of polysilicon fingers. 3.1.1 NMOS and PMOS Structures of Different Sizes An array of NMOS and PMOS transistors of different sizes provides information on how much variation affects a transistor given its size. The set of different size devices also enable one to extract parameters such as threshold voltage, as a function of channel length, width and area. The concept of these structures is shown in Figure 3-1. 51 i U Eli Im LmI U U U U IIIIIIIIIIIIIIII Figure 3-1: Different size FEOL test structures, with different channel lengths (L) and widths (W). The gates of all the size dependent transistors are single-fingered, and the sizes vary as shown in Table 2-3. There are 33 different sizes, for a total of 66 DUTs (33 NMOS and 33 PMOS). There are three different gate lengths (minimum, twice minimum and three times minimum), and 11 different gate widths (from 0.50 gm to 3 gm in 0.25 gm increments). DUT Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11 Width Lengt DUT Width Lengt DUT Width Lengt [p[ h [pm 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 Q12 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 Q23 Q24 Q25 Q26 Q27 Q28 Q29 Q30 Q31 Q32 Q33 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 [ ]jm Q13 Q14 Q15 Q16 Q17 Q18 Q19 Q20 Q21 Q22 Table 3-1: Sizes of test transistors, NMOS and PMOS. 3.1.2 Orientation Variation in direction dependent fabrication processes can be captured by varying the orientation of how the DUTs are laid out. The orientation test structures are identical 52 transistors laid out vertically and horizontally, as shown in Figure 3-2. Only the DUTs are placed in different orientation; the rest of the tile, including the switches and control circuitry, is always laid out with the same orientation (vertical). All orientation DUTs have a gate length of 0.18 gm and gate width of 1 ptm. There are four different structures in this category, two PMOS (one vertical and one horizontal) and two NMOS (one vertical and one horizontal). The pair of identical vertical and horizontal tiles are laid out side by side in the rows of the test chip to compare their behavior with minimum spatial impact. Vertical Horizontal Figure 3-2: Orientation FEOL test structures. 3.1.3 Polysilicon Spacing Variations due to lithography proximity can be studied by varying the distance between the polysilicon fingers of the gate of a transistor. The concept is shown in Figure 3-3. The distance between the fingers in these three-fingered structures are minimum spacing, 1.2x minimum, 1.5x minimum, and 3x minimum, similar to the structures in [1]. The dimensions of each finger are 0.5 gm/0.18 gm. There are eight structures in this category, four PMOS and four NMOS. The total drain to source resistance in these structures is kept constant (i.e., the distance between the drain and source contacts is kept constant), as shown in Figure 3-3, to minimize impact on output currents. 53 Figure 3-3: Poly finger spacing FEOL test structures. 3.1.4 Number of Polysilicon Fingers Transistors with different numbers of polysilicon fingers are laid out with the same effective gate length, but with the gate distributed across a different number of polysilicon fingers, as shown in Figure 3-4. These structures should have the same nominal current, if we neglect the device stack effect. The distribution of the gate length into different numbers of polysilicon fingers accentuates variation due to gate length variation, since AL may be x-times present in x-fingered transistors. Figure 3-4: Number of poly fingers FEOL test structures. There are a total of 12 structures in this category, six PMOS and six NMOS. These 12 test structures are divided into four different groups. The first group are NMOS transistors with gate widths of 0.50 gm and total gate lengths of 0.54 gm. There are three test structures in this group: a single finger device (0.54 gm per finger), a two fingered device (0.27 gm per finger), and a three fingered device (0.18 ptm per finger). The second 54 group has the same dimensions as the first group but with PMOS transistors. These two groups are illustrated in the left of Figure 3-4. The third group consists of NMOS transistors with gate widths of 0.50 tm and total gate length of 0.72 gim. There are three test structures in this group: a single finger device (0.72 gim per finger), a two fingered device (0.36 gm per finger), and a four fingered device (0.18 Rm per finger). The fourth group has the same dimensions as the third group, but with PMOS transistors. Both of these groups are illustrated in the left of Figure 3-4. 3.2 Back End of Line Structures Back end of line (BEOL) variation can also be studied with the analog multiplexed characteristics test chip. The BEOL analysis of this thesis focuses on capacitance extraction and variation measurements. Interconnect capacitance is mostly composed of three different capacitances: planar, fringing, and coupling capacitances, as illustrated in Figure 3-5. This thesis focuses on the extraction and analysis of the planar and coupling capacitances, and also the crossover capacitance which is a combination of the fringing and plane to plane capacitance. 2 T 1 2 T3 13 Figure 3-5: Interconnect capacitances - 1. Fringing, 2. Coupling, 3. Plane. 3.2.1 Charge Based Capacitive Measurement Methodology and Implementation Previous work [1] noted that measurements of variation in interconnects require 55 ----------------- accurate procedures, as these are harder to detect than device variation. This thesis uses a differential charge based capacitive measurement (CBCM) method [22], [23], [24], [25], [28] to detect these variations. The CBCM test structure is illustrated in Figure 3-6. It is integrated with switches so it can also be used with the scan chain approach. The back end of line (BEOL) tiles are then connected in a "row" sub-block, like the FEOL tiles are, except that these have a larger number of input signals, and therefore pads, to control them because of the additional required gate voltages (four different gate voltages, as opposed to one in the FEOL). The layout of this tile is shown in Figure 2-4 in Chapter 2. Vii- Switchli neti Switchj- Vlj V2i- Switch2i A Switch2j V2j Source Switch Vii - K- V2i ___L.VL Vlj -JhK-h - V2j .JhJh Step1 Step2 L--L hV Step3 Figure 3-6: CBCM structure with switches. With the CBCM method, the capacitance between nodes i and j is charged and discharged through the pseudo-inverters connected to the gate input signals during "Step 1" and "Step 2," as shown in Figure 3-6. In "Step 3," nodes i and j are at the same potential, and one can measure the current and calculate the capacitance due to the CBCM structure alone (without the interconnect load capacitance). The measurements of step three can then be subtracted from the measurements of steps 1 and 2. The average 56 source currents in these steps provide information to calculate C1, as described in [23] and [24], and as shown in (3.1): C= IstepI + Istep2 - Istep3 2Frequency VDD (3.1) Simulation of the CBCM structure gives measurements of sub-fF precision for test capacitors (ideal capacitors connected to their terminals) of 10 to 100 fF, for frequencies of 500 KHz. This data is shown in Table 3-2. Accuracy increases with increased speed, but we limit our measurements to 500 KHz for ease of measurement equipment selection. Test Cap CBCM Measured Capacitance (fW) 0 10 20 30 40 50 60 70 80 90 100 0.30 10.50 20.96 30.82 40.63 49.73 60.41 70.27 80.80 90.53 100.98 difference (fF) 0.30 0.50 0.96 0.82 0.63 -0.27 0.41 0.27 0.80 0.53 0.98 Table 3-2: CBCM measurements capabilities. The switches are the same as described in Section 2.6, except that here we need more per test structure (five switches instead of the two needed for FEOL test structures). Additionally, the RON in the source path because of the switch is not a problem, because what is taken into consideration is the difference between the currents; therefore a switch with a small RON connecting to ground is not needed here, allowing smaller switch transistors to be used. BEOL rows must be separate from FEOL rows, as they require a different setup of switches and of input pads. All of the switches in the BEOL tiles are the same size, and its schematic and dimensions are shown in Figure 3-7 and Table 3-3, 57 respectively. V1 i Switch )0.neti V2i Switch netj Switch IVij Switch V2j Q2 Q3 Q1 en /en en /en Output/Ground Switch BEOL switch en /en BEOL tile with switches Figure 3-7: BEOL switch schematic. Transistor Transistr Q1 Q2 Q3 Size Gate Length [pm] Gate Width [Wm] 0.35 2 0.30 2 0.35 2 Table 3-3: BEOL switch transistors sizing. The dimensions of the metal capacitances attached to the CBCM structure are chosen so they are as large as possible but that could fit in the area of a tile, as illustrated in Figure 3-8. PkanP Tilk Comb Tile Grid Tile Figure 3-8: Layouts of BEOL tiles. 58 3.2.2 Plane Structures The plane test structures measure the capacitance between two metal planes, enabling one to extract interlayer dielectric (ILD) thickness variation. The test structure consists of two large metal plates, one on top of the other, as seen on Figure 3-9. A large plate of metal-x is placed over another one of metal x-1; the x-l plate is made slightly larger in case they misalign during the mask process. The test chip includes four of these structures, with these combinations of metals: M1/M2, M2/M3, M3/M4, and M4/M5. All of the top planes have dimensions of 14.62 gm (wide) by 35 gm (tall). All of the bottom plates have dimensions of 16.62 jim (wide) by 37 gm (tall). CBCM Structure Figure 3-9: BEOL plane structure. 3.2.3 Comb Structures Comb structures enable one to detect variation in the spacing between the wires, using a finger-like structure, as seen in Figure 3-10. The difference between capacitances in different metal layers provides insight into the impact of variation in each metal layer. To determine if the variation is more heavily due to the variation in the lateral spacing or due to variation in the thickness of the wire, the structures are built with minimum width and minimum space, 1.2x minimum width and minimum space, and minimum width and 1.2x minimum space. Each of these comb structures are laid out in Ml, M2, M3, M4, and M5, for a total of 12 test structures. There are 28 fingers per structure, 14 attached to the node i and the other 14 to the node j of the CBCM structure. Each finger is 35.5 jim long, 59 and it is exposed to the next finger for 34.5 gm, as illustrated in Figure 3-10. Figure 3-10: BEOL comb structures. 3.2.4 Grid Structures The grid test structures focus on crossover capacitances, or the capacitance between a wire of one metal level over a wire of another metal level. The difference compared to the plane test structures is that the metals here are fingers of minimum size and at minimum distance, as shown in Figure 3-11, making the fringing capacitance from layer to layer a bigger factor than in the plane to plane case. The capacitive loads are built as a horizontally fingered metal layer under a vertical metal layer of higher level, all at minimum width and spacing. Coupling capacitance within the layer is not an issue, because the adjacent fingers of the same metal layer are at the same potential. The fingers have the same dimensions as those in the comb test structures, described in Section 3.2.3. There are four of these grid test structures: M1I/M2, M2/M3, M3/M4, and M4/M5. Figure 3-11: Grid test structures. 3.3 Summary This chapter has described the devices under test (DUTs) implemented in the test chip for characterization and investigation of process induced variation. By making each 60 structure, or each family of structures, primarily sensitive to a particular variation source, the pattern induced impact of these variations can be studied. Both front-end-of-line (FEOL) and back-end-of-line (BEOL) structures have been described. Chapter 4 will describe the measurement results of these test structures in fabricated chips. 61 62 Chapter 4 Results and Analysis for Layout-Dependent Test Structures This chapter analyzes the front end of line (FEOL) test structures based on electrical test data, focusing on structure functionality and variation due to layout practices. Section 4.1 first discusses the functionality of all the test structures in the test chip to determine which work as expected when tested. The statistical methodology used to identify chip-to-chip, within-chip, and structure-to-structure frequency variation components is presented in Section 4.2. In Section 4.3 we then examine the resulting measurements for various sets or families of structures in order to understand how structures within that family differ due to designed layout practices. First, the orientation test structures analysis studies the effect of layout orientation in the behavior of devices. The polysilicon spacing structures analysis focuses on the impact on channel length due to the local effect of polysilicon proximity (due to interaction between the different fingers of polysilicon in a device). The number of polysilicon fingers test structures analysis investigates how the channel length variation changes when different structures have the same total gate length but using a different number of fingers. Finally, the chapter is summarized in Section 4.4. 4.1 Functionality All FEOL and BEOL test structure designs have been tested successfully, as shown in Figure 4-1. Testing was performed using the 35 packaged chips with signals controlled through Labview and an NI PCI-6259 data acquisition card, and current measurements using a Keithley 2400 SourceMeter, also connected through GPIB to Labview. The 2400 ammeter can measure currents from 10 pA to 1.055 A. 63 Gate Voltage Sweep for All FEOL DUTs X 10. 1.6 1.4- N=8- / 1.2- 0.802 0 0 0.5 1 Gate Voltage [V] 1.5 2 Figure 4-1: Gate voltage sweeps at 1.8V VDS for all FEOL test structures, NMOS and PMOS, for one row of a single test chip (728 structures total). Each FEOL test row (720 tiles) takes approximately 30 minutes to measure for a single bias point, mainly due to measurement hardware and software setup/measuring times. Each tile only needs to be on for -50 milliseconds to gather a DC measurement; however, we leave the tile on for 100 ms, being cautious to make sure that all voltages and currents are settled in the feedback network. BEOL testing takes longer because the CBCM current must be averaged through extended periods of time for accuracy. With our setup each BEOL row (336 tiles) takes approximately 1.5 hours to measure. 4.1.1 ElectricalRepeatability It is important to have only minimal error in the electrical repeatability, so that we can trust that the source of variation seen is not due to measurement noise. FEOL test structures were all tested in DC, with time constants for the feedback network much 64 larger than what simulation calculations indicated were needed. This helped ensure a small measurement error, only 0.04% reproducibility and 0.02% repeatability errors for a single test chip tested multiple times, and gives confidence that most of the data measured is minimally affected by the measurement itself. 4.1.2 Leakage Currents When measuring a specific FEOL DUT in a test row, there are other 719 "off' tiles in the row that share the same output/ground line. The leakage current is measured by gathering the current in the share ground line after the scan chain has gone through all the test structures (after each DUT has been "on" once). The mean of the measured leakage current is 2.1 nA, which matches the simulated results and should not affect the saturation measurements. Sub-threshold measurements at very low voltages of currents near or below 2 nA will be affected by this current, and are therefore not used in our analysis. 4.2 Variation Analysis In this chapter, we focus on two key questions related to the observed variation in our test structures. The first goal is to consider the observed total variation in each type of test structure, and to decompose that variation into two components: "chip-to-chip" variation giving an indication of across wafer effects, and "within-chip" variation due to the effect of local or global layout environment on the test structure. We will see that some test structures are more prone to vary across the wafer than others. Similarly, we will also see that some test structures appear to be more sensitive to within-chip layout factors. The second goal of the statistical analysis in this chapter is to consider particular layout practices in more detail and identify their likely impact in the measured current. To accomplish this, sets of structures designed to investigate each layout effect are 65 considered. In general, we will be interested in two characteristics of the resulting DUT measurements: to what extent the mean of the transistor current has been affected, and to what extent the variance of the transistor current has been impacted, as performed in previous work [1]. 4.2.1 Within Chip and Chip to Chip Components As illustrated in Figure 4-2, there are two components of variation we would like to understand and separate based on measurements of currents across multiple replicates and multiple chips of a particular DUT. There is some spread in measured currents within each chip for a particular test structure type - the within-chip variance (across the replicates of that test structure within each chip, shown as a vertical set of points in the figure). Second, there may also be variation or differences between the mean currents (the line drawn within each chip group in the figures) from one chip to the next. We summarize these components of variation and their computation below using the nested variance components analysis method [29]. Chip 1 Chip 2 Figure 4-2: Representation Chip 3 .. " Chip N of chip-to-chip versus within-chip variation. 66 4.2.1.1 Within Chip Variance The within chip variance, Uac, refers to the variation of a specific test structure within a chip, and can be determined by: I 17C r1 ('i 1 2 )2 (4.1) In equation (4.1), r stands for the number of replicates of a specific DUT design in a chip, and I for the current measured in the device for a given bias condition. The subscript x indicates the chip number, and Ix is the average of the r replicate DUT currents within that test chip. After calculating all the variances within each chip, and because all c chips have the same number of replicates in them, the overall within chip variance can be calculated as the average of the variances in each chip, as shown in equation (4.2). 2X 2 4.2.1.2 (4.2) C IaWC 1 Chip to Chip Variance The chip to chip variance must be calculated carefully so that our calculation is not inflated by the within chip variance already calculated. The nested variance method compensates for exactly this. The observed chip to chip variance qocc is calculated as the variance across all chips of the mean current of a DUT in each chip, as shown in equation (4.3), where IAC is the average of Ix across all c chips. 2 JOcc = 1 (IL C - Ii=1 -1 )2 - IAC (4.3) To estimate the actual chip to chip variance (acc ) we subtract the within chip variance divided by the number of replicates to the observed chip to chip variance, as 67 shown in equation (4.4). (ECC 4.2.1.3 2_ ~ (4.4) c C CC Total Variance will then include both the chip-to-chip and within-chip The total variance a variance components, resulting in equation (4.5). This is under the assumption that the chip to chip and within chip variance sources are independent, e.g., due to across-wafer equipment non-uniformity vs. within chip layout dependencies, respectively. O72 2C _02 (4.5) T = JEcc - aWC Finally, we note that the within-chip variation may be greater than or less than the chip-to-chip variation - the sources of variation would physically be different and one or the other could well be larger. 4.2.1.4 Aggregated Variance We choose to distinguish between our estimated total variance (UT), and the observed variance of all measurements of a particular test structure type across all chips. We define the aggregated variance (oA ) as in (4.6). 2 UA = C1 (Ix I cr - 1 X=1 i=1 2 - Ic A (4.6) This is simply the measured variance across all replicates of a given DUT type, across all chips, without the detailed decomposition into separate chip-to-chip and within chip variance components. This aggregate variance is the simplest variation indicator based directly on the data without variation structure assumptions. 4.2.2 Normalized Variation To get a better qualitative feel for the observed variations between different test 68 structure types, we present a percent aggregate variation, V%, as a scaled ratio of the aggregate standard deviation UA to the mean, IAC. The variations presented in this and subsequent chapters are the standard deviation of each test structure type divided by the overall current mean IAC for that test structure type, as shown in equation (4.7). A V%= 100% (4.7) I AC 4.2.3 Evaluation of Mean Differences In many cases, we will be interested in determining how significant an observed mean difference is between two or more samples. For the example shown earlier in Figure 4-2, we might ask whether the mean difference is "real" - that is, how unlikely is it that we would observe the indicated mean difference given the spread within each set of measurements. We will use the analysis of variance, or ANOVA, formulation in the work below [30], [31]. A two-way ANOVA analysis accounts for the difference in mean, per DUT type and per chip. A two-way ANOVA implies that there are two factors at work; in our case they are the DUT design type and chip number. The implicit model for the two-way ANOVA is shown in (4.8). y = YO + yf + yc + N(O,u 2 ) (4.8) In equation (4.8) yo represents the overall mean of the current of a test structure type, yj represents the deviations of each test structure type's current (from the mean current) that are attributable to a specific design practice, y, represents the deviations of each test structure type 's current (from the mean current) that are attributable to the specific chip 69 number, and N(O, a') accounts for random disturbances. The two-way ANOVA analysis has the advantage of separating out an additive mean shift due to the chip number from a mean shift due to layout factor of interest (e.g. spacing between polysilicon fingers). 4.2.4 Evaluation of VarianceRatios It is also important to understand if the spread or variance observed in two populations are "the same" or significantly different. For example, we might want to understand if the variance in chip 1 is different than in chip 2, or if the observed ratio of variances is instead likely to occur simply by chance given the spread and number of measurements taken within each chip. For this we compare the ratio of variances to an inverse of the F cumulative distribution function for a given probability and degrees of freedom. 4.3 Layout Dependent Variation Analysis Some dependencies studied in [1] were again studied here to examine layout variation effects at the transistor level. All results shown in this section are for saturation bias conditions, as shown in Table 4-1. VREF NMOS PMOS (VDS equivalent) [VI 1.8 1.8 VGATE [ I 1.8 0 Table 4-1: Bias conditions for layout dependent variation analysis. 4.3.1 Orientation Test structures with different orientation (vertical vs. horizontal) show a small but consistent offset due to orientation, as seen in Figure 4-3. In these plots, the mean across all 16 replicates of the W/L = 1/0.18 pm device, and across all chips (35 chips for the NMOS and 31 chips for the PMOS) is shown; the error bar is the one standard deviation 70 across all devices (560 for NMOS and 496 for PMOS) of this size around that grand mean. The results are tabulated in Table 4-2. 5.95 X10 Overall Orientation NMOS -4 N=560 5.9 2.45 W/L=1/0.18 um 5.85 a) N=496 W/L=1/ 0.18um 2.4 -,_2.35 2 5.8 ----- - --- - - - -- - a) PMOS X 10-4 2.5 -- 2.3 --- ) 2.25 ----- - - --- - - - - - - 0 5.75 0 U0 5.7 -- ---- -- - ---- - -- - - - 2.1 -- - - -- - - - - 5.65 --- -- -- - 5.6 V 2.05----2- H V H Figure 4-3: Orientation effect for NMOS and PMOS transistors. Test Structure Type Average Current Among Al Replicates and Chips IAC [mAI NMOS Vertical NMOS Horizontal PMOS Vertical PMOS Horizontal 0.575 0.584 0.228 0.219 Standard Deviation Among All Replicates and Chips (A [LAI 12.3 11.1 16.3 15.7 Table 4-2: Tabulated orientation effect for NMOS and PMOS transistors. Interestingly, NMOS and PMOS transistors have the opposite dependency: NMOS have higher currents when laid out horizontally in this particular test chip, and PMOS are higher when laid out vertically, as seen in Figure 4-3. The difference in means between the test structures are proven statistically significant with the two-way ANOVA tests shown in Table 4-3 and Table 4-4. These tables refer to the data plotted in Figure 4-3. In this case, and referring to the implicit model for the two71 way ANOVA shown in (4.10), y, represents the overall mean of the current of a test structure type (NMOS, PMOS, vertical or horizontal), y, represents the current deviations given the designed orientation (vertical or horizontal), and yc represents the current deviations per chip. N(0, a 2 ) accounts for random disturbances. (4.9) y = YO + y, + yc + N(0,u 2 ) NMOS Vertical Source Orientation Chips Error Total And Horizontal Prob>F Mean Sq. F d.f. Sum Sq. 0 1 2.10E-08 332.6998 2.1OE-08 0 39.8306 34 2.52E-09 8.56E-08 6.85E-08 1084 6.32E-11 1.75E-07 1119 Table 4-3: Two-way ANOVA for NMOS with different orientation. PMOS Vertical And Horizontal Mean Sq. d.f. Sum Sq. Source 1 2.15E-08 2.15E-08 Orientation 30 4.66E-09 1.40E-07 Chips 960 1.20E-10 1.15E-07 Error 991 2.76E-07 Total F 179.9218 39.0038 Prob>F 0 0 Table 4-4: Two-way ANOVA for PMOS with different orientation Another method to corroborate that the observed means are statistically different is with Student's t-distribution a t-test. Equation (4.10) describes the estimation of a confidence interval for the mean, where IAC is the calculated overall current mean per type of test structure (for all chips and replicates), s is the pooled sample standard deviation and n is the number of replicates (in our case, 560 for the NMOS and 496 for the PMOS). We use a=0.05 for a 95% confidence interval. S S S 'AC - ta/~ ' /1 ' AC + a2nI-T (4.10) The results of these confidence interval estimates are plotted in Figure 4-4, which 72 provides a graphical way to verify that the sampled means are different. The thicker, smaller error-bars on top of the one standard deviation error-bars that were already presented in Figure 4-3 represent the confidence interval of the mean, to 95% confidence, given our sample size (560 for NMOS and 496 for PMOS). We will continue in the rest of this chapter to highlight mean differences between test structures with different layouts, using plots of both the sample standard deviation, and of the confidence interval of the mean as a simple graphical test or indication of significance to complete ANOVA calculations. Overall Orientation 5.95 x10 4 x10 4 NMOS 2.5 2.45 N=560 5.9 -----------------------------2.4 W/L=1/0.18 um 2.35 5.85 ------- t -------------a, 2.3 N=496 W/L=1/0.18 um ----------------- 2.2 -- -------- -- --- 2.25 ------- 5.75 ------ --- -- -- --------- 23 - - - - - - - - - - - C~~~~~~~ -------------5.8--- 0 0 PMOS -- ----- 0 5.7 -------- ----------- -- u 2.15 2.1 ----------- 5.65 -----5.6 V H - ---- ----- 2.05 2 - -- - ---- ----- - --- V H Figure 4-4: Orientation effect for NMOS and PMOS transistors with confidence intervals for the mean. The offset in currents between the vertical and horizontal structures is seen both in the averages across all chips and for each individual test chip. Figure 4-5 and Figure 4-6 plot the average current per chip for the orientation test structures, and show an offset between the vertical and horizontal test structure types for both the NMOS and PMOS cases. The sampled data includes all replicates per FEOL row, with a sample size of eight 73 per data point plotted 6 C: Orientation per Chip: NMOS Row 1 x10 -4 5.8 0 5.6L U) -- *--- -,A - Vertical Horizontal N=8 5.4 Chips 6 X1O -4 Orientation per Chip: NMOS Row 3 5.8 - 0 5.6 0 -- - Vertical -A - Horizontal U) N=8 5.4 Chips Figure 4-5: NMOS orientation effect on a chip by chip basis.. 2.4 X10-4 Orientation per Chip: PMOS Row 1 N=8 - 2.3 2 0 U) 2.2 2.1 - - - Vertical --A - Horizontal 2 Chips 2 3 x10 Orientation per Chip: PMOS Row 3 I I FwI-I-I -- A 2.5 F - Id I ver tica Hor izontal N=8 0 !LA U) A 2 Chips Figure 4-6: PMOS orientation effect on a chip by chip basis. 74 The consistency prompts us to look at each device and explore its behavior. In the FEOL test rows, all vertical structures are placed right next to a horizontal structure (see Appendix I for location information within row). We subtract the current mean of each vertical structure to its neighbor horizontal structure, calling this our orientation delta. There are 16 orientation deltas per chip, as there are 16 vertical and 16 horizontal test structures per chip. The 16 orientation deltas per chip (eight deltas per FEOL row) are averaged out and plotted in Figure 4-7 for the NMOS and Figure 4-8 for the PMOS. The delta of the current between the horizontal and vertical test structures is relatively consistent for all chips and for both NMOS and PMOS, as shown in Figure 4-7 and Figure 4-8. The consistent offset is seen in most structures placed right next to each other, and it hints at a systematic variation during the fabrication process. Furthermore, the fact that the offset is consistent but in opposite directions with NMOS and PMOS suggests that the variation may be related to the stress in the substrate. It has been reported [32], [33] that the electron and hole mobilities depend on the current flow direction relative to crystal orientation because of an anisotropic property of the effective carrier mass with strain. PMOS and NMOS transistors are affected in opposite ways by strain, as NMOS mobility increases with biaxial tensile strain and PMOS mobility increases with uniaxial compressive strain. The process we used is not purposely strained, but there may be some residual strain from the making of the devices, from processes such as shallow trench isolation (STI). Other possible sources of variation that may also affect the transistor's behavior due to orientation are mask bias and ion implantation directionality. 75 NMOS Orientation Deltas for Both Rows x10 5 =3 0 N=16 Chips Row 1 X 10-5 2_ 2 X10 -5 Row 2 1 C C C 0 U) 0 a) N=8 -1 N=8 0 -1 Chips Chips Figure 4-7: Current difference between adjacent NMOS vertical-horizontal test structures. 0 U) x 10-5 PMOS Orientation Deltas for Both Rows -0.5 C C -1 U) N= 16 0 -1.5 Chips 5 0 x 10- Row 1 Row 2 -11pu' I -0.5 -0.5 (D -1 x10 5 0 PFri' I'll1 -1.5 Chips C: N=8 N=8 0 -1.5 Chips Figure 4-8: Current difference between adjacent PMOS vertical-horizontal test structures. 76 . .......... ------------ 4.3.1.1 Variance Components and Variation Table 4-5 breaks down the sources of variation for the orientation test structures. The aggregate variation in these structures is approximately 2% for the NMOS and 7% for the PMOS, as shown in the last column of the table. The variation for all the structures was split almost evenly between within chip and chip to chip, with both vertical structures having a slightly larger chip to chip component. Within Chip a* V Varation -100% [(IiA)2} [(PA)2] 57.9 97.1 155 37.4% 62.6% 2.15% 62.6 62.4 125 50.1% 49.9% 1.90% 122 150 272 44.9% 55.1% 7.16% 122 129 251 48.6% 51.4% 7.18% T Vertical PMOS Horizontal Chip to Chip % [(MA)2] ECC Vertical NMOS Horizontal PMOS Within Chip % W"r .100% OT WC NMOS Estimated Estimated Total Chip-Chip - 1 00% IAC T Table 4-5: Variances and variation for orientation test structures. 4.3.2 Polysilicon Spacing Structures with the same number of polysilicon fingers but with different spacing between the fingers show a mean shift, but do not show much change in variation for either NMOS and PMOS devices, as seen in Figure 4-9. In these plots, the mean across all 16 replicates of the W/L =0.5/0.54 gm devices, and across all chips (35 chips for the NMOS and 31 chips for the PMOS) is shown; the error bar is the one standard deviation across all devices (560 for NMOS and 496 for PMOS) of this size around that grand mean. These results are tabulated in Table 4-6. The thicker, smaller error-bars on top of the one standard deviation error-bars represent the confidence interval for the mean, at 95% confidence, given our sample size 77 (560 for NMOS and 496 for PMOS). 1.66- X10 -4 Polysilicon Finger Spacing PMOS NMOS 1f 5 4.5 N=560 W/L=0.5/0.54 ______ ___ 1.62 -------------- 0 CO) 4.2 4.1 ------------------- 1.58 a) 4.3 --- - ------ ---------------------- .0 0) =3 0 ----------- 1.56 N=496 W/L=0.5/0.54 -- -- - - - - - 4.4 1.64 1.6 x --------- - - -- - - - -- - ----- -------- - 4 3.9 1.54 CO) 1.52 -------------- - ---- --~--- 1.5 F---- -------- ------------ 3.7 3.6 1.48 1x 3.5 1.2x1.5x 3x 1x 1.2x1.5x 3x Spacing Spacing Figure 4-9: Poly finger spacing effect for NMOS (left) and PMOS (right). Test Structure Type Average Current Among all Replicates d CChips [mAj 0.159 0.156 0.155 0.153 0.0402 0.0392 0.0388 0.0384 IAC NMOS NMOS NMOS NMOS PMOS PMOS PMOS PMOS Ix Spacing 1.2x Spacing 1.5x Spacing 3x Spacing Ix Spacing 1.2x Spacing 1.5x Spacing 3x Spacing Standard Deviation Among al Replicates CA [pA_ 3.60 3.56 3.51 3.49 3.22 3.10 3.09 3.13 Table 4-6: Tabulated poly spacing effect for NMOS and PMOS transistors. 78 A strong trend is observed in Figure 4-9: larger spacing or separation between poly fingers leads to lower currents, at both NMOS and PMOS test structures. Table 4-7 and Table 4-8 demonstrate by two-way ANOVA that the mean differences between the different polysilicon finger spacing is statistically significant, for both NMOS and PMOS transistors. These tables refer to the data plotted in Figure 4-9. In this case, and referring to the implicit model for the two-way ANOVA shown in (4.12), y, represents the overall mean of the current of a test structure type (NMOS or PMOS with spacing of minimum size, 1.2x minimum size, etc.), y, represents the current deviations given separate spacing between the fingers due to design practice, and y, represents the current deviations per chip. N(0, U 2 ) accounts for random disturbances. y = yO + ys + yc + N(0,. NMOS Poly Spacing Sum Sq. Source 1.18E-08 Spacing 1.60E-08 Chips 1.20E-08 Error 3.98E-08 Total 2 (4.11) ) Prob>F Mean Sq. F 0 3 3.92E-09 717.3836 0 85.9738 34 4.70E-10 2202 5.47E-12 2239 d.f. Table 4-7: Two-way ANOVA for NMOS with different finger spacing. PMOS Poly Spacing Mean Sq. F d.f. Sum Sq. Source 3 2.77E-10 57.8831 8.31E-10 Spacing 70.625 30 3.38E-10 1.01E-08 Chips 9.33E-09 1950 4.79E-12 Error 2.03E-08 1983 Total Prob>F 0 0 Table 4-8: Two-way ANOVA for PMOS with different finger spacing. The trend suggests that the channel length is somewhat reduced (leading to increased saturation currents) as we pack the transistor fingers more closely together, consistent with a lithographic or etch proximity effect that shortens or narrows these features as they 79 7:7 % approach each other. An additional theory is that the "stack effect" may be at work, in which voltage drop due to the diffusion layer resistance between the fingers has an adverse effect on the transistor behavior. What this means is that the resistance between the fingers will change the source voltage of the inner nodes, and this will change the VBS voltage of each transistor, which can lower the current for both NMOS and PMOS. However, for a diffusion layer sheet resistance of 5 Q/square, the largest resistance between nodes for the 3x minimum spacing structure would be under 20 Q, and simulation results show that with such a resistance one would not see a change larger than 0.5% between the output currents. In Figure 4-9 we see a change of 1.6%-2.5% in the measured data, which is larger than we expect to see with the stack effect only. 4.3.2.1 Variance Components and Variation Within Estimated Estimated Total Chip-Chip Chip Within Chip % (T~clic--100% [(pA)2] NMOS . I x Spacing NMOS 1.2x Spacing NMOS . 1.5x Spacing NMOS . 3x Spacing PMOS . Ix Spacing PMOS . 1.2x Spacing PMOS . 1.5x Spacing PMOS . Chip to Chip % -- -100% Varation ILC [(pA)2] 6.11 [(A)2] 7.04 13.15 45.1% 54.9% 2.26% 5.21 7.67 12.88 39.5% 60.6% 2.28% 5.1 7.41 12.51 39.7% 60.3% 2.26% 5.22 7.16 12.38 41.1% 58.9% 2.28% 5.22 5.33 10.55 48.0% 52.0% 8.02% 4.96 4.81 9.77 49.2% 50.8% 7.91% 4.72 4.98 9.7 47.2% 52.8% 7.96% 5.07 4.86 9.93 49.5% 50.5% 8.14% Q2 SpacVrg Tb3x Table 4-9: Variances and variation for polysilicon spacing test structures. 80 Table 4-9 breaks down the sources of variation for the polysilicon spacing test structures. The variation in these structures is approximately 2.3% for the NMOS and 8% for the PMOS, as shown in the last column of the table. The variation for all the structures is slightly more heavily due to chip to chip variation. The fact that the split is close to even suggests that both level process variation, perhaps due to film thickness or related variation, and die-level effects due to lithographic or etch proximity effect, are important. 4.3.3 Number of Polysilicon Fingers NMOS structures with same drawn gate length but with the gate split among a different number of polysilicon fingers show a strong dependency on variation due to the number of fingers, as seen in Figure 4-10. The top plots in Figure 4-10 show the analysis for NMOS structures with equivalent gate length of 0.54 gim. These gate lengths are divided into one-0.54 gim finger, two-0.27 gim fingers and three-0.18 gm fingers. The bottom plots show the variation for a structure with equivalent gate length of 0.72 im. These gate lengths are divided into one-0.72 gm finger, two-0.36 pim fingers and four0.18 gim fingers, as explained in more detail in Section 3.1.4. On the left hand side of Figure 4-10 we plot the current means of the NMOS test structures with different number of fingers but equivalent gate length (1, 2 and 3 fingers for the top plot, with W/Lequivaient=0.5/0.54gim, and 1, 2 and 4 fingers for the bottom plot, with W/Lequivalent=0.5/0. 7 2 gim). The current means are calculated across all 16 replicates of each test structure type and across all 35 chips, for a total of 560 samples. The error bar is the one standard deviation around that grand mean. On the right hand side of Figure 4-10 we plot the variation of each test structure type. 81 The variation is calculated by dividing the aggregated standard deviation of all 560 devices (with a specific number of fingers and equivalent gate length) by their overall mean, or -IAC 100%. The error-bars on top of the variation bar plots represent the normalized confidence variance interval. The interval formula for the estimate of variance is shown in (4.12), where n is the number of replicates (560 in this case) and s2 is the pooled sample variance. We use a=0.05 for 95% confidence interval in the Chisquare distribution. (n- 1)s 2 2 - %a/2,n-1 - 2(n 2 A ( )s2 2 (4.12) X-a/2,n-1 The bar plots show how variation increases steadily as the numbers of fingers increase. Figure 4-11 shows the same analysis for the PMOS transistors, with analysis of all 16 replicates of each test structure type and across 31 chips, for a total of 496 samples. However, even though variation is larger for the PMOS transistors, their variation deltas are much smaller. The overall means and standard deviations for the NMOS and PMOS transistors of different number of fingers are tabulated in Table 4-10. Test Structure Type Type Leq ivaint # of Fingers 1 NMOS 0.54 gm 2 3 1 NMOS 0.72 gm 2 4 1 PMOS 0.54 gm 2 3 1 PMOS 0.72 gm 2 4 Average Current IAC [mA] 0.143 0.148 0.159 0.115 0.114 0.127 0.0360 0.0371 0.0400 0.0276 0.0269 0.0302 Standard Deviation 'TA [pA 2.13 2.57 3.47 1.69 1.78 2.74 2.68 2.76 3.14 2.06 2.03 2.44 Table 4-10: Tabulated poly spacing effect for NMOS and PMOS transistors. 82 -777-77=--- N=560 W L=_Th 1.6 0 1.4 1 1.3 _4_.___ _ _ 0.01 F-- 0 3 2 1 2 3 Number of Fingers -4 0.03 2.16% C 0.02 1.2 1.47%1.56% 1 0.01 0 1.15 C/) - 1.74% 0.02 F 1.49%+ -AA N=560 W/L=-0.5/0.72 1.25 - 2.18% Number of Fingers Xl1O - 0.03 EU Cl) -- C CO 1.5 W - NMOS of Different Number of Fingers -4 x1O 1.7 ;;;-, - -- - --- 1.1 0 2 4 1 Number of Fingers 1 2 4 Number of Fingers Figure 4-10: Mean(left) and normalized variation (right) in saturation current for NMOS transistors with different number of polysilicon fingers. Statistics for N=560 DUTs consisting of 16 replicates per chip, for 35 chips. 4.5 xl1O PMOS of Different Number of Fingers -5 S4 0 C CO -- 7.%744-%7.85% 0.05 I - 3.5 E 0) 3 65) 0.1 N=496 -W/L--0.5/0.54 ---- 0 2 3 1 Number of Fingers 3.5 x10 C 3 1 2 3 Number of Fingers -5 - N=496 W/L=0.5/0.72 - 0.1 + (U cc; --- 2.5 - - 7.46%7.52%8.08% C 0.05 E M) 0 W/ 2-- 4 1 2 Number of Fingers 0 4 2 1 Number of Fingers Figure 4-11: Mean(left) and normalized variation (right) in saturation current for for PMOS transistors with different number of polysilicon fingers. Statistics for N=496 DUTs consisting of 16 replicates per chip, for 31 chips. 83 -.- . - -QM If we consider the ratio of the overall variance of the NMOS three-fingered, minimum length structure to the variance of the single finger, 3x minimum length structure, we get r = (4.75e-12/1.26e-1 1) 2 = 7.01. The inverse of the cumulative distribution function (cdf) for F gives a critical ratio corresponding to the largest ratio of variances one would expect to observe with a given probability, and for specific degrees of freedom in the numerator and denominator variances. In this case we choose a 95% probability, and the nominator and denominator degrees of freedom are the number of the specific type of test structure (for all chips) minus one. If we compare the ratio 7.01 to FO. 9 5, 559, 559= 1.15 (in this case, we have 16 of these structures in each chip, times 35 chips), we see that the observed expansion in variance is indeed significant at the 95% confidence level. We perform similar tests for the other types of structures, NMOS and PMOS, and tabulate the results in Table 4-11 and Table 4-12. NMOS NMOS 0.54 jim drawn gate length Number Fingersof 1 2 rr Estimated Total (Y '95s9,s 4.75E-12 6.91E-12 7.01 structure 3 1.26E- 11 0.72 jim drawn gate length 1 2 2.97E-12 3.32E-12 structure 4 7.82E-12 1.15 1.15 6.94 Table 4-11:Variance ratios of NMOS multiple fingers structures. Number of Fingers Estimated Total c2 0.54 jm drawn gate length 1 2 1.09E-1 1 1.01E-11 structure 3 9.99E-12 0.72 jm drawn gate length 1 2 4.44E-12 4.31E-12 structure 4 6.23E-12 PMOS Table 4-12: Variance ratios of PMOS multiple r F_ 1.89 1.16 1.97 1.16 fingers structures. 84 Referring to Table 4-11 and Table 4-12, NMOS test structures with equivalent gate lengths of 0.54 jim and 0.72 jim show strongly that there is a significant change in variance between the different numbers of fingers. PMOS test structures also show that their variances are different to 95% confidence, but the ratio is not as large as the NMOS. A possible explanation for why the transistors show larger variation given a larger number of fingers is that the effective gate length can change with variation, as ALeff can be three times larger in a three finger device than in a single finger device, as partially illustrated in the simplified MOSFET cross section [34] shown in Figure 4-12. This three fold increase in opportunity for variation suggests that variance may be proportional to number of fingers. Potential AL D Z GS Eff ective Gate Length P Figure 4-12: MOSFET cross section. However, the factor that seems to dominate the variation in our test structures with different number of fingers is the size of the poly fingers. The normalized variation of the NMOS test structures with the smallest sized fingers (0.18 jim long) show almost identical variation for the three-finger (2.18%) and four-finger (2.16%) structures, as shown in Figure 4-10. This means that even though the four-finger test structure is submitted to additional AL when compared to the three-finger test structure, it did not show more variation, but it shows much larger variation than its counterpart test 85 structures with larger fingers, regardless of the equivalent gate length. This highlights the need of studying variation due to transistor sizing, which we focus on in Chapter 5. 4.3.3.1 Variance Components and Variation Similar to the analysis of the other test structures, Table 4-13 shows the breakdown of the sources of variation for the test structures with split gate lengths. As stated earlier, the variation in these structures increases as the size of fingers decrease, for both PMOS and NMOS, but more markedly for NMOS. The variation for all the structures is split almost evenly between chip to chip and within chip. Within Chip Or2 L m I C [A)21 [qLA Estimated Within hip- Total Chip % 2[(A) 1i L [(A2; Chi c 0.72 NMOS 1 2 3 1 2 4 2.26 3.64 5.82 1.18 1.41 3.12 2.34 3.04 6.38 1.71 1.82 4.50 4.75 6.91 12.6 2.97 3.32 7.82 0.54 PMOS 1 2 3 4.05 4.15 5.01 3.21 3.59 5.02 7.51 8.00 10.3 0.72 PMOS 1 2 4 2.41 2.28 2.96 1.88 1.89 3.08 4.44 4.31 6.23 0.54 NMOS -100% T-100% 14&491I Chip to Chip % T 47.7% 52.7% 46.3% 39.9% 42.5% 39.9% 53.9% 51.9% 48.5% 54.2% 52.9% 47.5% Variation .-100% 1AC C 52.3% 47.3% 53.7% 60.1% 57.5% 60.1% 46.1% 48.1% 51.5% 45.8% 47.1% 52.5% 1.49% 1.74% 2.18% 1.47% 1.56% 2.16% 7.43% 7.44% 7.85% 7.46% 7.52% 8.08% Table 4-13: Variances and variation for polysilicon-fingers test structures. 4.4 Summary All FEOL test structures are successfully measured, with high electrical repeatability (0.02% error), and leakage currents from the "off' tiles of 2 nA per FEOL row, as expected from simulation. Structures designed to have layout dependent variation are analyzed, and to make sure that the means of different structures are different we perform 86 ANOVA on them. We use a nested variance calculation method to separate the variance due to within chip and between chip sources. With this method we conclude that most layout-dependent structures have similar magnitudes of variance components due to within chip versus between chips. Structures with different orientation present an offset that is consistent between chips and between most structures laid out next to each other. Interestingly, NMOS and PMOS are affected inversely with orientation, probably because they are also affected inversely by residual strain. Structures with different spacing between fingers show a mean shift consistent with a lithographic effect. Structures with different number of fingers increase their variability as the number of fingers increase and the size of individual fingers decrease. Our findings are consistent with the ones found in our previous work [1] using a ring oscillator based test structure. Chapter 5 studies the transistors at different bias conditions, gathering I-V curves, something that our previous test chip was incapable of. 87 88 Chapter 5 Results and Analysis of Test Structures of Transistors of Different Sizes A key feature of our test chip is its ability to measure a large range of currents across a range of structure types, and with enough replication of structures to gather variation statistics. The test chip includes NMOS and PMOS transistors of three different gate lengths (0.18 tm, 0.36 Itm, 0.54 itm) and 11 different gate widths (from 0.50 ptm to 3.0 [tm in 0.25 [tm intervals), for a total of 33 unique NMOS DUTs and 33 unique PMOS DUTs, as explained in more detail in Section 3.1.1. In this chapter we will first discuss in Section 5.1 the functionality of the test chip in terms of voltage sweeps, and pinpoint its limitations. Section 5.2 studies the transistors in saturation, while Section 5.3 studies the transistor response to voltage sweeps, including threshold voltage extraction and analysis. Section 5.4 summarizes the findings presented in the chapter. 5.1 Functionality All of the transistors of different sizes are subjected to both reference voltage (theoretically equivalent to VDS, via a feedback network) and gate voltage sweeps. An example of a measured device is shown in Figure 5-1, where we see that the measurements closely resemble transistor curves up to small values of VDS, when our system reaches its limits as discussed in Section 2.7. In this plot we show a drain voltage sweep at different gate bias points for an NMOS transistor of dimensions of 0.50 gm/0. 18 pm channel width and length. 89 VDS Sweep, NMOS 0.50u/0.18u x 10-4 3.5 32.5- 2(1.5- 0.5 0 -0.5 0 VREF Figure 5-1: VDS 0.5 (controls VDS 1.5 1 [V] feedback) voltage through 2 sweep for VGS ranging from 0-1.8V in 0.1V intervals. Figure 5-2 shows the semi-log plot of the measured I-V gathered curves for a VDS sweep of the same device, an NMOS transistor of 0.50 [tm/0.18 gm channel width and length, at different gate voltages. As discussed in Section 2.7, our design does not gather data under 2 nA because of the leakage of the "off' tiles in the row, but we are able to measure currents above that point. Our output swing from the feedback network is limited more than expected in simulation (refer to Section 2.7), and our system reaches a limit in regulated VDS supply at a minimum of about 0.2 V. We are able to gather measurements for VREF Of 100 mV if we lower the opamp's ground to -100 mV, as shown in Figure 5-3. We believe that this is due to the bias of the output buffer, which limits the output swing of the output buffer. By lowering the ground voltage, Q9 from Figure 2-10 in Chapter 2 can go lower in voltage, hence allowing measurements at 100 mV, vs. the 200 mV limit observed before in Figure 5-2. 90 10 -4 ------- 10 10-4 10 VDS Sweep, NMOS 0.50u/0.18u -- -- -- - - - - - - - - --- -- ~- -- -_------------ - -- - -- --- - - - -V -- -_Gate=O -0.6-1.8V " 8 -5 Gate=O.5V 10 o--------- VGate=.4V V 0 , 10 10I 7 ---8- 1o 0 0.2 0.4 VREF VGate=0, .V Gate=O.3V 0VGate=O.2V 10 0.6 0.8 1 1.2 1.4 1.6 1.8 (controls VDS voltage through feedback) [V] - ranging from 0 to 1.8 V in - a VDS sweep for VGS - of 1 9 -Figure 5-2:- Semi-log plot showing the limitation of our measurements. 0.1V intervals, VDS Sweep, NMOS O.50u/O.18Bu 10-- 10 - - ---- VGateO*2 10 S16 V 10 -teO5V Gate=O0.4V 10 1 10 0.1 Figure VREF (controls VDS voltage through feedback) [V] 5-3: Semi-log plot of a VDs sweep for VGS ranging from 0 to 1.8 V in 0.1V intervals. The opamp's ground voltage is set at -100 mY. 91 We also observe a small inconsistency in the currents at approximately VDS= 0.7 V at low VGS. This seems to be due to the "off' PMOS transistors, which share the same "ground" off gate line with the NMOS (see switch description in Section 2.6). At ground (0 V), the gate of the PMOS transistors would be "on," but the device is still off because its source (through a switch) is at the same potential as its drain (supplied by the opamp). However, even a very small offset between the drain and source of the "off' PMOS can make it conduct current. This current is not connected to the output current we measure, but if the offset makes the drain voltage to be at a slightly lower potential than the source voltage, then the current can flow into the opamp, and it can slightly offset its output voltage, particularly at low voltages. This problem does not present itself at higher currents. It is noticed only when VDS is small (under 0.7 V), and even there it is not very pronounced for currents over 10 nA. To address this issue, in Section 8.3 we propose to have a separate "off' gate line for the NMOS and PMOS in future implementations. Despite these limitations, Figure 5-4 illustrates how closely our measured data matches with that of simulation at large drain voltages. In this plot, VDS is set at 1.8 V. This graph matches closely with simulated data until the output currents approach the leakage currents of the off tiles (2 nA). Most of our variation studies reported in this chapter use these VGS sweeps. The IourVDS curves are valid for most values VGS, Of VDS and and additional device parameter extractions are possible using these curves. 92 10 10 10 C Vgs Sweep for 1.8V Vds, NMOS 0.50u/0.1 8u - Simulated -Measured -4 -5 - - ---- - -- -6 -- - - -------- - - -- - --- -- - - 10 2 10 0 (.) -- ------------- ---- ----- - -- - - -7 -8 -10 10 10 Vgs Figure 5-4: Comparison between simulated and measured data for a VDS of 1.8 V, ranging from 0 to 1.8 V in 0.1 V intervals, for eight replicates of one type of transistor in one chip (N=8). 5.2 Transistors of Different Sizes at Saturation The NMOS DUTs at bias voltages of VDD = VGATh = 1.8V show a larger variation in devices with smaller gate length and gate width features, as seen in Figure 5-5. The bar graph shows the normalized variation oA /Ac value of the transistor currents, i.e., the standard deviation of the DUT set divided by the mean for each type of transistor. These measurements are taken for the eight replicates of the DUT type in each chip, for all 35 packaged chips, hence the total DUT in each case consists of N=280. For each channel length of 0.18, 0.36 and 0.54 tm, statistics for devices with widths from 0.5 to 3 tm in 0.25 [tm increments are shown (corresponding to the partial annotations in the figure). 93 In Figure 5-5 there is a clear dependency of variation on the size of transistors. The larger relative variation is seen not only for smaller gate lengths, but also for smaller gate widths. A similar, although noisier pattern, is also found for PMOS transistors, as shown in Figure 5-6. Normalized NMOS Variation per Size 0.0180.0160.0140.0120 0 C C 0.010.008- 0.006E C/5 0.0040.0020 L=0.18u L=0.36u L=0.54u W=0.50u-3.Ou W=0.50u-3.Ou W=0.50u-3.Ou Figure 5-5: Variation in NMOS currents at VDS=1.8V and VGS=1.8V, N=280 (8 replicates in 35 chips). Normalized PMOS Variation per Size 0.025 C: - 0.02- 0 U) 0.0150 C) CO () 0.01 F 0) E5 0.005 F 0 L=0.18u W=0.50u-3.Ou L=0.36u W=0.50u-3.Ou L=0.54u W=0.50u-3.Ou Figure 5-6: Variation in PMOS currents at VDS=1.8V and VGS=O V, N=248 (8 replicates in 31 chips). 94 The measurements for the PMOS devices are taken for the eight replicates in 31 chips, hence N=248. The entire set of 35 chips is not available for this study, as some chips were either damaged in the testing process or were rejected as outliers. The variation seen in both NMOS and PMOS transistors follow Pelgrom's model [35] for dependence in total channel area, as seen in Figure 5-7 and Figure 5-8. The fit of the NMOS transistors is closer to the ideal slope of 1 than the PMOS. The PMOS measurements are very linear with respect to channel area, but their extrapolation to infinite area (zero on the x axis) does not converge to zero variation. With the x coordinate set to zero (or infinite area), the linear fit for the NMOS devices is R2 = 0.96, and R 2 = 0.83 for the PMOS devices. Linearity of NMOS Transistor Variation According to Size 0.020 - 0.018 -I SI I 2 =0.96 0.016 -=R S0.014 E 0.0120.008 S0.006 0.002 - - - 0.006 ~t I I I I I 0.000 0 1 2 3 4 1/Square Root of Area Figure 5-7: Relative current variation in NMOS transistors as a function of device area. 95 LineariF y of PMOS Transistor Variation According to Size -, - - -, -, , - -, - ----0.030 7 I -I- I - 7 I -1 T - 0.025 i 0| -7 I I I ,R2= R2=0.83 '0 (D 0.020 Xa ~0.015 I I I I - -1T - I1 1 | - I - - I I - I7 F 1- Ir - II - TI-I- - - 10.010 AAIC ----T 0.005 - T Ci000A .UUU - I F I 0 I | 1 T I I - -I T I I I 2 3 I I 4 1/Square Root of Area Figure 5-8: Relative current variation in PMOS transistors as a function of device area. The variation values and variance breakdowns (chip to chip and within chip components as percentage of the total) are shown in Table 5-1 for the NMOS transistors and in Table 5-2 for the PMOS. NMOS structures showed aggregate uA/IAC variation ranging from 2.4% for the smallest test structure to 0.89% for the largest. The variation in the PMOS is larger, ranging from 7.3% to 6.9%. The NMOS devices with gate lengths of 0.18 pm show, for the most part, a larger variation component due to within chip variation. As most structures become larger, especially the PMOS, the relative proportion of within chip and chip to chip variation start to become more even. However, when we look at all of the structures as a whole, the within chip variation in the NMOS is 27.4% and in the PMOS it is 42.9% of the total variation oU . 96 .- - Gate Within Chip -. - Z.1-1 -- --- - - -, -" - --- Estimated Estimated Total Chip-Chip - -'. - - -- M ,- Within Chip % S-ECC100% . Chip to Chip% .100% t IA 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.54 0.54 0.54 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 27.5 43.4 51.7 78.0 82.1 111 127 145 185 212 199 5.30 7.95 9.52 13.8 17.4 20.1 24.6 29.2 32.9 39.2 43.0 2.16 3.19 4.38 26.3 70.4 117 162 236 318 382 451 545 653 738 4.00 7.54 11.3 15.9 23.9 26.7 29.9 40.3 45.5 47.9 63.2 2.68 4.42 6.02 55.5 117 172 245 323 436 516 605 742 878 949 9.63 16.0 21.4 30.5 42.4 48.1 56.0 71.3 80.5 89.5 109 4.97 7.81 10.7 49.5% 37.3% 30.0% 31.9% 25.4% 25.6% 24.6% 24.0% 25.0% 24.1% 20.9% 55.1% 49.7% 44.5% 45.1% 41.1% 41.9% 43.8% 40.9% 40.9% 43.8% 39.5% 43.4% 40.8% 41.1% 50.6% 62.7% 70.0% 68.2% 74.6% 74.5% 75.5% 76.0% 75.0% 75.9% 79.1% 44.9% 50.3% 55.5% 54.9% 58.9% 58.2% 56.2% 59.1% 59.1% 56.2% 60.5% 56.6% 59.2% 58.9% 2.38% 2.42% 2.25% 2.19% 2.11% 2.13% 2.04% 1.98% 1.99% 1.99% 1.91% 1.65% 1.48% 1.31% 1.27% 1.25% 1.15% 1.09% 1.10% 1.06% 1.02% 1.03% 1.60% 1.39% 1.23% 0.54 1.25 6.61 7.28 14.3 46.2% 53.8% 1.15% 58.9% 57.4% 1.05% 1.01% 0.54 0.54 1.50 1.75 7.01 9.08 9.60 11.7 17.0 21.3 41.1% 42.6% 0.54 2.00 11.0 15.1 26.8 41.1% 58.9% 1.00% 0.54 2.25 13.3 18.1 32.3 41.3% 58.7% 0.98% 0.54 2.50 13.8 19.3 34.0 40.6% 59.4% 0.90% 0.54 0.54 2.75 3.00 15.1 18.3 19.9 27.0 36.0 46.4 42.0% 39.5% 58.0% 60.5% 0.85% 0.89% 1 8 Table 5-1: Variances and variation in saturation current, VDS=VGS= . V, for NMOS transistors of different sizes, N=560. 97 Gate Within Chip Estimated Chip-Chip Estimated Total Within Chip % 100% W [11m] L [11 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 ] 2 36.0 73.8 131 184 244 350 453 515 626 753 846 9.46 19.4 29.7 48.1 65.0 87.1 112 143 178 211 235 4.66 9.11 15.6 23.7 32.4 43.2 55.8 68.8 85.8 103 121 2 40.9 85.2 154 233 331 468 591 758 906 1120 1300 7.86 15.6 26.0 40.6 58.1 80.5 103 129 158 188 226 3.47 6.92 12.1 18.2 27.7 36.5 46.2 58.4 71.6 88.6 104 79.1 164 293 429 590 840 1070 1310 1570 1920 2200 17.9 36.2 57.6 91.7 127 173 222.0 281 348 412 475 8.42 16.6 28.6 43.5 62.1 82.4 105 132 163.0 198.0 232 Chip to Chip % 100% Variation IAC YT 45.5% 45.1% 44.7% 43.0% 41.4% 41.7% 42.3% 39.5% 39.8% 39.2% 38.5% 52.8% 53.5% 51.6% 52.4% 51.1% 50.3% 50.4% 51.0% 51.3% 51.3% 49.3% 55.3% 54.9% 54.3% 54.6% 52.1% 52.4% 52.9% 52.3% 52.7% 52.0% 52.1% 54.5% 54.9% 55.3% 57.1% 58.6% 58.3% 57.8% 60.5% 60.2% 60.8% 61.5% 47.2% 46.5% 48.4% 47.6% 48.9% 49.7% 49.6% 49.0% 48.7% 48.7% 50.7% 44.7% 45.1% 45.7% 45.4% 47.9% 47.6% 47.1% 47.7% 47.3% 48.0% 47.9% 7.30% 7.16% 7.27% 7.03% 6.92% 7.02% 6.95% 6.82% 6.71% 6.75% 6.64% 7.12% 7.04% 6.82% 6.91% 6.82% 6.80% 6.76% 6.74% 6.75% 6.65% 6.57% 7.18% 7.04% 7.14% 7.07% 7.11% 7.00% 6.96% 6.88% 6.89% 6.89% 6.85% Table 5-2: Variances and variation in saturation current, VDs=1.8 V and VGs=O V, for NMOS transistors of different sizes, N=496. 98 Transistorsof Different Sizes with Voltage Sweeps 5.3 Each DUT of the "transistors of different sizes" category is measured at different bias voltages to characterize variation dependencies. Figure 5-9 shows an example set of current versus VGATE curves, at different reference voltage bias conditions. These measurements can be carried out across the large set of structures with different device sizes, and can reveal interesting information about device variations. x 10-4 VGS Sweep, NMOS 0.50u/0.18u VDS=1.8V 32.5 / VDS=0.5V 22 1.50 VDS=0.1V VD=O 1.5 0.501 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 VGS [V] Figure 5-9: VGATE sweep for transistor in scan chain at VDs of 0.1 V to 1.8 V, and opamp ground set at -100mV. 5.3.1 Transistorsof Different Sizes with Gate Voltage Sweeps Figure 5-10 plots the overall variation in output current across all device replicates in a single chip, for transistors designed at a given gate length. The bias conditions are 1.8 V for VDD and a voltage sweep of VGATE values. The figure clearly illustrates how variation in sub-threshold is substantially greater in devices with smaller gate lengths. Figure 5-10 99 shows the normalized variation in output current at different gate voltages for NMOS transistors of different sizes, calculated as the square root of the aggregate variance per gate length divided by the aggregated mean per gate length, UA/IAC . The plot covers different gate voltages (from 0 V to 1.8 V) at 0.01 V intervals, with drain voltages set at 1.8 V. Current Variation for all NMOS Structures 0.25 I I I I I I i I I I I I I I I I I I I I I I I I I I 0) I I I I I I I I I i I I I I I I I I I I I I I I I r I I I I I 0.15F I I I r *-~Iength=O.18u--f 0.2 I I I I I I - r I I I I I I I I I I I I I I I I r I I I I I I I I I I I I I I I Iength=0.36u I I I I E >) 0.1 I I I I 0 0 0.2 0.4 I 0.6 0.8 I I I I I I I I 0.05 I I I I ength=0.54u 1 1.2 1.4 I I I I I I I I I I 1.6 1.8 2 Gate Voltage Figure 5-10: Normalized variation (a/p) of the output current at different VGATE values, for devices with different channel lengths. Measurements of one test chip, N=176 (16 replicates of 11 different gate widths at a fixed gate length). The current equations for both the sub-threshold and saturation regimes for a large VDS voltage, as provided by [36] and [37] respectively, are: VGS -VT ISub = ION = Joe s K(VGS -VT)' (5.1) (5.2) 100 The sub-threshold slope behavior [38] indicates that in this region, ideally, there can be up to a lIx increase in currents for a 60 mV increase in voltage. This slope is the steepest when the device is deep into the sub-threshold region, and therefore the most variation can be found there. Because this is the maximum point of the slope, the variation is expected to decrease as VGATE decreases, as the currents of the MOS transistor become very small as the transistor turns off [26]. However, our measured data peaked earlier and decreased faster and more than expected from simulation, as seen in Figure 5-11. Simulated Transistor 0.50um/0.18um 0.45 0.4 ---------------------------------------- 2C 0.35 ---------------------0 0.5 ------------00.5---------- -------- --------------------------------------- -- -- - Cl) iL-------------- L-------------- .2b-------L-------------- 0' 0 E CD 0 .1 - - - - - - - ---- 0 .05 0 0 - - - - - - -- - - - -r --- - - 0.5 - - -- - - - - - - - - - - -7 - - - - - - - 1 Gate Voltage 1.5 2 Figure 5-11: Variation of a simulated transistor. The variance is calculated from the curve of a transistor in a typical simulation corner and another transistor with a 5% variation in VT. Figure 5-11 shows the simulation of an NMOS transistor with dimensions of 0.50 ptm/0. 18 tm. The transistor is simulated with a typical model (acquiring the range of 101 currents called Ii), and then with a modified model with 5% variation in VT(acquiring the range of currents called I2). We then determine the standard deviation of these I] and I2 currents with respect to each other, at each gate voltage. In Figure 5-11 we plot this standard deviation divided by the mean current at each gate voltage. For this simulation experiment, the sub-threshold variation peaks at -0.1 VGSThe discrepancy between the simulated and measured data is due to the very low currents measured at less than 0.3 V of gate voltage. These currents are on the order of a few nA and below. Because the system's leakage currents are approximately 2 nA (see Section 4.1.2), we are unable to measure currents lower than this, and hence our variation analysis hits its limit. The peak of the variation was likely higher than that seen in Figure 5-10, because the peak likely occurs at gate voltages under 0.3 V. Regardless of the limitation of the measurement, Figure 5-10 clearly shows that smaller length devices have much larger relative variation than larger devices. It also shows much larger relative variation in sub-threshold, likely due to current exponential dependence on VT, as suggested in [26]. Equation (5.3) shows the derivative of the subthreshold current with respect to VT. ai Sub _ TV -e VGS~ VT GS (5.3) s In equation (5.3) it is evident that as gate voltages decrease, the current can be more affected due to variations in VT, which agrees with our findings in the sub-threshold region. This shows the high impact of VT variation, which we study in the next section. 5.3.2 Threshold Voltage VariationAnalysis Threshold voltage variation is of increasing concern in advanced technologies [39], as its continual scaling requirements pose several technology and circuit design challenges. 102 For this analysis, threshold voltages (Vr) are obtained using the linear extrapolation method in the saturation region [25] with the alpha power-law MOS model [37], [40]. 5.3.2.1 NMOS Threshold Voltage All the NMOS devices of different sizes have their threshold voltage calculated with the linear extrapolation method, as mentioned above. For this study, the parameter x is set to 1.32 for 0.18 gm gate lengths, 1.62 for 0.36 gm gate lengths and 1.69 for 0.54 gm gate lengths, by fitting them to the measured data. VT voltages are calculated to be 0.53 V for the smallest NMOS devices and 0.49 V for the largest, which agrees with values for this technology (0.18 gm CMOS Logic), according to the manufacturer's test data for a logic process (typical library). 11 0.55 ~VIdIMl .,lIJ m.. Al Ml ~afr~V - ---------- OTS A% 0 O0 0.54 0.53 INI 0.52 CO, 0i 7- ------- 0.51 - L=0.18u 0.5 - I --- W=O.SOu-3.Ou - -W' L=0.54u 50u 0.49 L=0.36u W=0.50u-3.Ou AR Different NMOS Transistors Figure 5-12: NMOS VT error-bar plots, ±1-a bars shown. N=480. Figure 5-12 plots the calculated threshold voltage. In this plot, the mean across all 16 replicates of the NMOS devices of different sizes, and across 30 chips is shown. The 103 error bar is the one standard deviation across all 480 devices per size around the aggregated mean. These results are tabulated in Table 5-3. Gate W L Average VT VAC [V] 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.529 0.532 0.532 0.532 0.529 0.527 0.525 0.524 0.522 0.521 0.519 0.498 0.501 0.502 0.502 0.501 0.500 0.498 0.498 0.496 0.495 0.494 0.488 0.493 0.494 0.494 0.493 0.492 0.491 0.490 0.489 0.488 0.487 Table 5-3: Tabulated NMOS Standard Deviation 12.0 9.36 9.05 7.94 6.97 6.68 6.67 6.02 5.27 5.45 5.00 7.40 5.65 5.03 4.57 4.29 3.97 3.79 3.30 3.33 3.45 3.05 5.33 4.71 3.97 3.61 3.14 3.13 3.05 2.85 2.77 2.65 2.55 o [MV) VT calculation for devices of different sizes. 104 Threshold voltages follow a pattern of decreasing for larger gate widths for a given gate length, in agreement with simulations and with the reverse short channel effect (RSCE) [34]. Table 5-4 shows by a three-way ANOVA that the difference in VT between devices of different gate widths and gate lengths are statistically significant. The three-way ANOVA accounts for the difference in mean per DUT gate length, gate width, and also per chip. A three-way ANOVA implies that there are three factors at work; in our case they are the DUT designed gate length, the DUT designed gate width, and the chip number. The implicit model for the three-way ANOVA is shown in (5.4). V =V,+VL +V + V, + N(O,u2 ) (5.4) In the model shown in (5.4), V, represents the overall mean of the threshold voltage of a test structure type, VL represents the threshold voltage deviations given the designed gate length, Vw represents the threshold voltage deviations given the designed gate width, and V, represents the threshold voltage deviations per chip. N(0,U ) accounts for random disturbances. The test shows that the differences in VT mean are significant. NMOS Threshold Voltage Sum Sq. Source 1.1265 Chip 0.1483 Gate Width 3.746 Gate Length Error 0.8218 5.8425 Total d.f. 29 10 2 15798 15839 Prob>F Mean Sq. F 0 0.0388 746.7216 0 0.0148 285.002 0 1.873 3.60E+04 5.20E-05 Table 5-4: Three-way ANOVA for NMOS threshold voltage. Our measurements also show that variation -A/7AC is larger in smaller devices for both gate length and width, as seen in Figure 5-13. The measurements shown are for 105 NMOS transistors in both FEOL rows (16 replicates per device) in 30 test chips, hence N = 480. The variation is illustrated in Figure 5-13. Figure 5-14 shows the linear correlation of relative variation to the square root of the area, following Pelgrom's model [35], [41]. Normalized Variation Vt NMOS 0.025 0.02 - -- - - - - - -8 - - - - - 0 .0 15 - - - - - | I| - - - -- - - -- - -- - -- - - - I I ------ - - - -- - S 0.01 E 0.005 -MEEEE 0 L=0.54u W=0.50u-3.Ou L=0.36u W=0.50u-3.Ou L=O. 18u W=0.50u-3.Ou Figure 5-13: Variation in NMOS threshold voltage. Linearity of NMOS Threshold Voltage Variation According to Size 1A E-02 1.2E-02 - - -I I - - - 02=0. , 1.OE-02 -- 8.OE-03lo - 6.OE-03-4.OE-03 -- - , - - - - - - 2.OE-03-O.OE+00 0 0.5 1 1.5 2 2.5 3 3.5 1NArea [gm-1] Figure 5-14: Linearity of NMOS threshold voltage according to size. 106 MMw_ ___Qww Within Chip Gate W L [g[ 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 Estimated Estimated Total Chip-Chip 2CC UEC V]n 150 104 93.8 75.8 64.2 59.0 63.6 47.7 41.9 45.6 36.8 67.6 46.1 37.3 37.4 29.9 27.4 27.5 25.0 25.0 25.1 25.6 44.8 33.8 27.9 22.9 23.9 24.1 22.1 21.5 21.2 20.5 19.0 [2 V] 217 216 252 221 217 232 226 209 208 204 203 51.4 53.8 56.4 58.7 60.3 60.4 55.4 62.8 59.1 59.5 63.4 44.6 48.8 50.9 51.1 48.4 57.5 54.7 56.0 56.4 51.9 53.7 T V]2 367 319 346 296 282 291 290 257 249 250 240 119 99.9 93.7 96.1 90.2 87.9 82.9 87.8 84.1 84.5 89.0 89.4 82.6 78.8 74.0 72.3 81.6 76.8 77.5 77.6 72.4 72.7 Within Chip % Chip to Chip % -10% -YE.100% 40.8% 32.5% 27.1% 25.6% 22.8% 20.2% 22.0% 18.6% 16.8% 18.2% 15.3% 56.8% 46.1% 39.8% 39.0% 33.2% 31.2% 33.2% 28.5% 29.7% 29.7% 28.8% 50.1% 40.9% 35.4% 31.0% 33.1% 29.5% 28.8% 27.8% 27.3% 28.3% 26.1% 59.2% 67.5% 72.9% 74.4% 77.2% 79.8% 78.0% 81.4% 83.2% 81.8% 84.7% 43.2% 53.9% 60.2% 61.1% 66.8% 68.8% 66.8% 71.5% 70.3% 70.4% 71.2% 49.9% 59.1% 64.6% 69.0% 67.0% 70.5% 71.2% 72.3% 72.7% 71.7% 73.9% OCy . Vaation VAC .Ecc 100 2.29% 1.77% 1.69% 1.50% 1.33% 1.28% 1.27% 1.15% 1.01% 1.04% 0.96% 1.48% 1.13% 1.00% 0.92% 0.86% 0.78% 0.76% 0.66% 0.67% 0.69% 0.62% 1.09% 0.95% 0.80% 0.73% 0.64% 0.63% 0.62% 0.58% 0.56% 0.55% 0.53% Table 5-5: Variances and variation for VT in NMOS transistors of different sizes. N=480. 107 The variation values and variance breakdowns (between chip to chip and within chip) for the NMOS VT measurements are shown in Table 5-5. For almost all structures (except for the smallest NMOS, which was the one showing the most variation) the chip to chip component is the largest contributor to the overall variance. On average 64% of the variation seen in the VT structures is due to the chip to chip component, suggesting that it has a large spatial trend within the wafer. However, a large within chip level in VT is also present. 5.3.2.2 PMOS Threshold Voltage The PMOS test structures of different sizes also have their threshold voltage calculated with the linear extrapolation method, as stated in Section 5.3.2. For this study the parameter c is set to 2 for all gate lengths, based on a fit to our measured data. The absolute values of threshold voltages are calculated to be 0.57 V for the smallest NMOS devices and 0.54 V for the largest, which is slightly larger than the expected values for this technology (0.53 V for a typical library in 0.18 tm CMOS logic), but still within the 10% range of the provided manufacturer's corners. The threshold voltages results are plotted in Figure 5-15. In this plot, the mean across all 16 replicates of the PMOS devices of different sizes, and across 26 chips is shown. The error bar is the one standard deviation across all 416 devices per size around that grand mean. These results are tabulated in Table 5-3. 108 Threshold Voltage PMOS 0.6 0.59 L=0.18u W=0.50u-3.Ou 0.58 0) 0) 0.57 0 ~0 0.56 L=0.54u -5 C,) 0) -C 0.55 I- 0.54 --- -----------q ---- ----- W=0.50u-3.Ou -po-- p;-- 0.53 0.52 L=0.36u W=0.50u-3.Ou Different PMOS Transistors Figure 5-15: PMOS VT error-bar plots, ±1-(y bars shown. N=416. 109 Gate Average VT Standard Deviation W L VAC [VI UA [mV] 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.573 0.575 0.577 0.577 0.578 0.577 0.577 0.577 0.577 0.577 0.577 0.538 0.542 0.543 0.544 0.545 0.545 0.546 0.546 0.547 0.547 0.548 0.530 0.533 0.535 0.536 0.537 0.538 0.539 0.539 0.539 0.540 0.541 14.0 12.7 11.4 11.0 10.2 9.97 9.78 8.92 8.64 8.62 8.90 8.77 8.57 7.73 7.65 7.01 7.14 6.66 6.59 6.50 6.33 6.11 7.45 6.78 6.48 6.08 6.27 5.89 5.71 5.51 5.61 5.41 5.29 Table 5-6: Tabulated PMOS VT calculation for devices of different sizes. 110 Threshold voltages follow a pattern that agrees with regular transistor simulations. Table 5-7 shows by a three-way ANOVA that the difference in VT between devices of different gate widths and gate lengths is statistically significant. This analysis is similar to the three-way ANOVA explained in Section 5.3.2.1. The implicit model for the threeway ANOVA is the same as in the NMOS study, as previously given in equation (5.4). PMOS Threshold Voltage Source Sum Sq. Chip 0.9779 Gate Width 0.0751 Gate Length 4.0109 1.47 Error Total 6.5339 d.f. Mean Sq. 25 0.0391 10 0.0075 2 2.0055 13690 1.07E-04 13727 F 364.3013 69.9255 1.87E+04 Prob>F 0 0 0 Table 5-7: Three-way ANOVA for NMOS threshold voltage. Our measurements also show that variation uA /IAC is larger in smaller devices in both gate length and width, as seen in Figure 5-16, but the trend is not as consistent as with the NMOS devices. The Pelgrom [35] [41] relation fit, shown in Figure 5-17, is also not as good as with the NMOS. The relationship between variation and square root of area is in fact quite linear (R2= 0.90), but with an offset of 0.005 in the y axis, due to measured variation that is independent of size. The R 2 for the fit when the y intercept is set to zero is only 0.66. An offset, although smaller, is also seen in the PMOS saturation study shown in Section 5.2. 111 Normalized Variation Vt PMOS 0.025 0 F F F F-F-F- - ----- - ------ L--- - - - -- $- - 0.02 F FFC) F 0.015 0 -C C: 0.01 E M) Fn) 0.005 0 L=0.54u W=0.50u-3.Ou L=0.36u W=0.50u-3.Ou L=0.18u W=0.50u-3.Ou Figure 5-16: Variation in PMOS threshold voltage. Linearity of PMOS Threshold Voltage Variation According to Size 1.8E-02 T 1.6E-02 -F 1~ T -- T - - - -L 1.4E-02 R2=0.66 - 1.2E-02 1.OE-02 FU 8.OE-03 6.OE-03 4.OE-03 2.OE-03- - T- -- -fI -- -|F ' - - r -F1 - -- 7 F i T - 1 - I7T F l I ---- - I + - -I- - ---I I I --- 1 _______________| 0.OE+00 0 0.5 1 1.5 2 2.5 3 3.5 1NArea [gm-1] Figure 5-17: Linearity of PMOS threshold voltage according to size. 112 Gate Within Chip - Estimated Chip-Chip Estimated Total Within Chip % 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.18 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.36 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.54 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 Table UT yy [ 231 185 164 166 134 135 134 112 113 108 109 104 106 97.2 90.7 80.6 82.9 78.4 76.6 75.1 75.5 67.0 85.6 78.2 72.6 70.4 71.7 63.7 65.5 61.5 63.6 61.4 58.5 132 111 151 158 161 163 162 178 166 175 168 48.6 52.0 46.6 45.0 52.7 49.9 56.8 47.6 54.3 50.1 52.2 35.6 34.4 39.9 39.6 37.2 37.7 35.6 39.5 36.1 35.8 35.4 363 296 315 324 295 298 297 289 279 283 277 152 158 144 136 133 133 135 124 129 126 119 121 113 113 110 109 101 101 101 99.6 97.2 93.9 63.6% 62.4% 52.2% 51.2% 45.4% 45.4% 45.3% 38.6% 40.7% 38.1% 39.4% 68.1% 67.2% 67.6% 66.8% 60.5% 62.5% 58.0% 61.7% 58.1% 60.1% 56.2% 70.6% 69.5% 64.6% 64.0% 65.9% 62.8% 64.8% 60.9% 63.8% 63.2% 62.3% % 1. .100% L [Wn] [pn] Chi Chip % UT Vatai V AC % O 36.4% 37.6% 47.8% 48.8% 54.6% 54.6% 54.7% 61.5% 59.3% 61.9% 60.6% 31.9% 32.9% 32.4% 33.2% 39.5% 37.6% 42.0% 38.3% 41.9% 39.9% 43.8% 29.4% 30.5% 35.5% 36.0% 34.2% 37.2% 35.3% 39.1% 36.2% 36.8% 37.7% 2.44% 2.20% 1.98% 1.92% 1.76% 1.73% 1.69% 1.54% 1.50% 1.49% 1.54% 1.63% 1.58% 1.42% 1.41% 1.29% 1.31% 1.22% 1.21% 1.19% 1.16% 1.12% 1.41% 1.27% 1.21% 1.13% 1.17% 1.09% 1.06% 1.02% 1.04% 1.00% 0.98% 5-8: Variances and variation for VT in PMOS transistors of different sizes. N=416. 113 The variation values and variance breakdowns (between chip to chip and within chip) for the PMOS VT measurements are shown in Table 5-8. For most structures, the within chip component is the largest contributor to the overall variance, but the within and chip to chip components are not very far apart. On average 58 % of the variation seen in the PMOS VT structures is due to the within chip component, suggesting that neither inter or intra die variation is dominant in this study. 5.4 Summary FEOL test structures that consist of transistors of different sizes are tested over a range of bias voltages, producing measurements for reference voltage values (drain voltage of DUTs) down to 0.2 V for the regular setup, and currents as low as 2 nA. At currents under 2 nA the measured currents are the leakage currents of the "off' tiles. If the ground of the operational amplifier is lowered to -0.1V, then we can gather currents for VREF down to 100 mV. The gate voltages sweep closely match simulation for VDS of 1.8V, and with those curves we show analysis of variation due to sizing, and threshold voltage extraction and variation. The variation of both NMOS and PMOS transistors follows a Pelgrom model, adding confidence in our findings. The variation in the threshold voltage for the NMOS transistors also fits Pelgrom's model, but the PMOS devices do not have a very close fit. The variance seems to split almost evenly between inter and intra die variation for threshold voltage, with a slightly higher intra die component for the current variation. This concludes the analysis of the FEOL test structures, and Chapter 6 discusses the testing and results of the BEOL test structures. 114 Chapter 6 Results and Analysis of Interconnect Test Structures The goal in the testing and analysis of the interconnect or back end of line (BEOL) structures is to be able to systematically and efficiently extract capacitance values using a scan chain, and to understand changes in capacitance due to process induced variation in the interconnect structures used for the capacitive load. This analysis is important for modem circuit design, in which the interconnect variations are no longer negligible; although interconnect variations tend to have a smaller impact than the device variations on circuit performance, the relative impact of interconnect variation in future technologies is expected to increase [21]. In this chapter we first look at how the measurements are calculated in Section 6.1. The three types of structures that are considered in this study, each of which is dominated by a different component of capacitance, are the coupling capacitance load, plane to plane capacitance load and a grid (or overlap) capacitance load, all measured using a CBCM approach, and an overview of the calculated capacitances is presented in Section 6.2. Each of these interconnect parasitic loads is tested in metals 1, 2, 3, 4 and 5 (metal 6 was reserved for signals). The findings of testing these structures are presented in Sections 6.3 (for the coupling capacitance), 6.4 (for the pane to plane capacitance), and 6.5 (for the crossover capacitance). We summarize the analysis of the interconnect test structures in Section 6.6. 6.1 Measurement Method Review and Adjustments Interconnect capacitances are measured using a modified differential CBCM method [23], [24], [25], integrated with switches so it can also be used with the scan chain 115 approach, as illustrated in Figure 3-6 in Chapter 3. As explained in more detail in Section 3.2, with this method, the capacitance between two nodes of interest is charged and discharged through the pseudo-inverters connected to the gate input signals during "Step I " and "Step 2," as shown in Figure 3-6. The average source currents in these steps provide information to calculate C1 as described in [23] and [24]. In "Step 3," the two nodes of the capacitor of interest are kept at the same potential, and one can measure the current and calculate the capacitance due to the CBCM structure alone (without the interconnect load capacitance), which can then be subtracted from the measurements of steps 1 and 2. However, reviewing out testing results we find that the blank capacitance test structures in the scan chain, or structures that have no capacitance connected to the CBCM structure, measure a non-zero capacitance value. "Step 3" should subtract whatever capacitance is present at these blank test structures, and its measured capacitance value should be zero. This offset may to be due to the limitations of our testing setup, where the particular board used to connect the signals seems to induce a measurement offset. To correct this offset we subtract the value of the blank test structure to the other capacitance values calculated in each chip. We therefore determine the interconnect capacitances using equation (6.1). C IstepI + Istep2 - Istep3 V DD *Frequency - B (6.1) 2 The constant B is the average capacitance calculated at the blank test structures in each chip, which accounts for the additional fixed error in capacitance extraction that may be included in the measurement. This fitting leads to capacitance values close to 116 what were expected by the layout extraction tools, as will be shown in Section 6.2. The use of the fitting factor comes at the expense of being unable to do the variance component separation and conclude if the observed variation is mostly due to its within chip or chip to chip component. This is because we subtract the value of the blank capacitance B for each chip, and with it we subtract off part of the chip to chip variance component. This subtraction allows us to understand structure to structure, and within chip variations, but comes at the cost that we lose chip to chip variation, and so the variance decomposition into these two components is not possible. 6.1.] ElectricalRepeatability As already addressed in Section 4.1.1, it is important to have only small errors in the electrical repeatability so that we can trust that the source of variation seen is not a measurement error, particularly in the interconnect structures that are measured at a relatively high frequency (500 KHz). BEOL test structures give a measurement error of only 0.08% for a single test chip, tested multiple times using "Step 1" biasing, providing confidence that the data measured is indicative of structure difference rather than electrical measurement noise. 6.2 Overview of Measured Capacitances Results All BEOL test structures are successfully tested, as partially illustrated in Figure 6-1. In this plot, the mean across all 16 replicates of a row of the different BEOL test structures, and across 35 chips is shown. The error bar is the one standard deviation across all 560 devices per size around the aggregated mean. 117 Interconnect Capacitances per Type 110 N=560 100 - -- --- 6 -- ---- -T ------ - -- - - - -- 7 90 -- 80 a) 0D 70 -- - - -7 - ---- -T ------ 7- - Comb fl-I T -- - - -- -- - -- - - --- ------ 0 - - ------- 60 050 40 -- - - - - -- - - - - - - -- Plane --- --------- Grid 30 nf Figure 6-1: Measurements of interconnect test structures. Trends shown in Figure 6-1 are as expected. The "comb" structure capacitances, which focus on within metal layer coupling capacitance at minimum spacing, are much higher than those of the layer to layer plane or grid structures. We also see that the metal 1 coupling capacitance is higher than in other metal layers, as expected, because the wires are closer together (the minimum allowed spacing is smaller) in this first layer metal layer. The "plane" capacitance concentrates on layer to layer capacitance, and is consistent with expected values given plate size (62 gm by 35 gim) and respective metal layer dielectric thicknesses. The "grid" test structures concentrate on the crossover capacitance, which is greatly impacted by both the layer to layer and fringing capacitance components. 118 Test Structure Combs Plane-Plane Grid Metals Used Ml M2 M3 M4 M5 M2-M1 M3-M2 M4-M3 M5-M4 M2-M1 M3-M2 M4-M3 M5-M4 Extracted Capacitance from Diva [fF] 98.82 87.65 88.08 88.29 88.49 24.50 24.45 24.49 24.48 66.88 60.82 60.48 59.88 Measured Capacitance [fF] 102.68 96.52 96.22 96.06 97.23 27.05 26.06 27.34 27.45 24.66 24.07 25.06 25.25 Table 6-1: Comparison between extracted by simulation and measured capacitances. N=560. Table 6-1 compares our measured data with the data extracted from the Diva layout extractor for this technology. The measurements of the combs and plane to plane capacitances are close to the extracted values, giving us confidence in our measurements. Measurements for the grid structures are substantially different than the extracted value, apparently because of an extraction error by the automated simulation-based extraction tool, which is explained in more detail in Section 6.5. 6.3 Comb Capacitances Results and Analysis The test structures that focus on the coupling capacitances, referred to as comb test structures, have the highest measured capacitances, as expected from the extraction. We are able to measure the difference in capacitance for all metal layers. The calculated values from the measurements are near, although larger by approximately 9%, to those predicted by the automated simulation-based extraction tool, as seen in Table 6-1. The difference in means between the different comb test structures are statistically significant based on the two-way ANOVA tests shown in Table 6-2. In this case, and referring to the implicit model for the two-way ANOVA shown in (6.2), y, represents 119 the overall mean of the calculated capacitance of a comb test structure type, ym represents the capacitance deviations given the designed metal layer (metal 1, 2, 3, 4 and 5), and y, represents the capacitance deviations per chip for all 35 chips. N(O, a2) accounts for random disturbances. y = YO + y, + yc + N(O,a 2 (6.2) ) We note that, while the subtraction of the blank structure capacitance B for each chip removes much of the chip to chip variation, the two-way ANOVA indicates that a small but statistically significant chip to chip variation in the comb and other capacitance test structure values still remains. As discussed in Section 9.2.3, future work might consider improvements to the CBCM structure to remove the need for this B correction, enabling more detailed analysis of chip to chip spatial trends. Comb Capacitances Sum Sq. Source Metal 1.75E-26 Chip 2.67E-26 3.79E-26 Error Total 8.21E-26 Mean Sq. 4 4.38E-27 34 7.85E-28 2761 1.37E-29 2799 1 d.f. F 319.3364 57.2375 Prob>F 0 0 1 Table 6-2: Two-way ANOVA for comb capacitances made with different metal layers. 6.3.1 VariationAnalysis Between FingerSpacing and Thickness We test the comb test structures with different spacing between the fingers and different wire widths, as explained in Section 3.2.3. Table 6-3 shows that the variation does not change much in our test structures as a function of line spacing or wire width. Furthermore, Table 6-4 shows that the difference in the variance between the structures is not statistically significant, when we look at the aggregate variance ratio r between the test structure type with larger spacing between the lines and minimum line width vs. the 120 test structure type with larger line width but minimum spacing. We could not determine with our data if variation was more heavily due to spacing or wire width. Future work could revise the experiment and use a larger range of spacing and/or widths of the metals. Metal Layer Variation in Metals of Minimum Width And Spacing = Variation in Metals of Minimum Width and 1.2x Spacing 100% 100% .100% CAC CAC CAC 4.85% 4.51% 4.67% 4.79% 4.78% 5.25% 5.43% 4.93% 4.76% 5.68% 5.32% 4.72% 4.72% 4.79% 4.94% 1 2 3 4 5 Variation in Metals of Minimum Spacing and 1.2x Wider Table 6-3: Variation (1-sigma/mean) in comb test structures. 22 Ml M2 M3 M4 M5 Minimum Width And Spacing 0.2985 0.2072 0.2066 0.2114 0.2308 Minimum Width and 1.2x Spacing 0.2159 0.2159 0.1790 0.1675 0.2345 MinItnum Spacing and 1.2x Wider 0.2614 0.2024 0.2142 0.2284 0.2293 r between s and w 0.68 1.14 0.70 0.54 1.04 Fo.9,5w, 2 1.19 Table 6-4: Variance ratios of coupling test structures. 6.4 Plane to Plane Capacitances Results and Analysis Plane test structures were successfully measured, and their measurements are also near although somewhat larger than those predicted by our automated simulation-based extraction tool, as shown in Table 6-1. The difference between the measured capacitance of the different metal layers is small, but it is statistically significant as indicated by the two-way ANOVA in Table 6-5. 121 Comb Capacitances Sum Sq. Source 6.71E-28 Metal 2.69E-27 Chip 2.79E-26 Error 3.13E-26 Total d.f. 3 34 2202 2239 Mean Sq. F 17.6498 2.24E-28 6.2497 7.92E-29 1.27E-29 Prob>F 2.54E-11 0 Table 6-5: Two-way ANOVA for plane to plane test structures. Similar to the two-way ANOVA for the comb test structures, in the implicit model for the two-way ANOVA shown in (6.3) y, represents the overall mean of the calculated capacitance of a plane test structure type, yrepresents the capacitance deviations given the designed metal layer (metal 1-2, metal 2-3, metal 3-4, and metal 4-5), and y, represents the deviations per chip for all 35 chips. N(0,u2 ) accounts for random disturbances. y 6.5 = YO + ym + yc + N(0,a 2 ) (6.3) Grid Capacitances Results and Analysis In contrast to the coupling and plane capacitances, the measurements obtained from the grid capacitances differ greatly from the extractions from the automated simulationbased extraction tool using our design kit (0.18 im CMOS Logic). To investigate the automated simulation-based extraction tool process, we studied the simulation of the two capacitive loads shown in Figure 6-2. The capacitive load labeled (a) features a crossover capacitance with fingers that are 30 times further apart than the fingers in the capacitive load labeled (b). These are minimum width wires, and therefore have a strong fringing component. The metal lines are metal 2 (yellow) and metal 1 (blue). The metal 1 line is labeled as node j, while the metal line 2 is labeled as node i. 122 Node i (a) Node j (b) Figure 6-2: Capacitive loads to study automated simulation-based extraction tool crossover capacitance extraction. The extractor calculates the Cij of structure "a" and Cij of structure "b" to be identical (64.6 aF). It also calculates half of this value for a single crossover (32.3 aF). This is incorrect according to Wong et al. [42], because the fringing component of the overlap capacitance is reduced when lines are closer together, as the flux from Ml and M2 is reduced with less spacing. Using Wong's [42] closed loop equations for the crossover capacitance between nets i and j (not to substrate), we can estimate the total capacitance expected for our crossover structure for metals 1 and 2 to be 7.5 aF, using the thickness of the wires and between 123 metal layers from the design kit. The calculations are shown in Appendix II. We have a total of 76x25 crossover capacitances, and an additional 8 fF capacitance in a connector line, for a total of 22.5 fF for our metal 1 to metal 2 grid test structure, which is much closer to our measurement than the 66.88 fF extracted by the automated simulation-based extraction tool. The extraction tool's measurement can be estimated by its one-crossover capacitance (32 aF), multiplied by 76x25 crossover capacitances and an additional 8 fF capacitance in the connector line, for an approximate total of 68 fF. Based on these results, it appears that the simulation-based extraction tool suffers from an extractor error in this (admittedly) unusual tightly packed wire geometry. This study shows the value of having accurate capacitance measurements on chip, to enable detection of miscalculations by automated extractors, particularly if the capacitance geometry or structure is not a commonly generated one. It is not common to have minimum length wires of the same potential at minimum spacing from each other, but as technology advances this practice may become more common, in uses such as regular interconnect fabrics [43]. It should be noted that this is an older design kit that is no longer supported by the manufacturer, and that this problem might have been fixed in more recent design kits. The difference between the measured grid capacitance of the different metal layers is small, but it is statistically significant as indicated by the two-way ANOVA in Table 6-6. Similarly to the two-way ANOVA for the comb and plane test structures, in the implicit model for the two-way ANOVA shown in (6.4), y, represents the overall mean of the calculated capacitance of a grid test structure type, y, represents the capacitance deviations given the designed metal layer (metal 1-2, metal 2-3, metal 3-4, and metal 124 4-5), and y, represents the current deviations per chip for all 35 chips. N(0, 2 ) accounts for random disturbances. y =YO + y, + yc + N(0, a 2 ) _ Comb Capacitances_ Source Sum Sq. d.f. 6.71E-28 3 Metal Chip 2.69E-27 34 Error 2.79E-26 2202 Total 3.13E-26 2239 _ (6.4) _ Mean Sq. 2.24E-28 7.92E-29 1.27E-29 F 17.6498 6.2497 Prob>F 2.54E-11 0 Table 6-6: Two-way ANOVA for grid test structures. 6.6 Summary The BEOL test structures were successfully tested, and we were able to map trends and measure interconnect capacitances in an effective and multiplexed way. We were unable, however, to pinpoint which source of variation, spacing or metal width, was more prevalent for the variation in the capacitances that accentuate coupling. All measurements are tested using a two-way ANOVA, and all of the different means reported are statistically different. However, the mean differences within structure types or from layer to layer are generally quite small. All test structures except for the grid measured values that are close to those expected from an automated extractor. The grid structure is seen to be miscalculated in the automated extractor, and hand calculations give us confidence in our measured data. This concludes the testing and analysis of individual test structures; Chapter 7 studies the spatial behavior across the wafer and within a die of various device and interconnect test structures. 125 126 Chapter 7 Spatial Analysis The wafer spatial location of each of our chips is tracked during manufacture. This, along with the within-chip location from the layout, provides data to perform chip-to-chip and within-chip spatial analysis. This chapter focuses on analyzing the variations due to spatial location at the wafer and die level. First, the wafer level trends based on patterns in the chip-to-chip variation are examined in Section 7.1. Next, in Section 7.2 within-chip trends are considered for a number of different test structures. Finally, Section 7.3 summarizes the results of the spatial analyses conducted. 4.5 6 7 6 9 07 Figure 7-1: Location of test dies within the wafer. Shaded squares indicate the dies available for spatial measurement and analysis. 7.1 Chip to Chip Spatial Analysis Spatial analysis throughout the wafer is possible for our 35 packaged chips, given the spatial location of each die within the single wafer used for all packaged chips, as shown in Figure 7-1. The means and normalized standard deviations for each chip of the most replicated structure, the 3-finger single spaced NMOS transistor, are shown in Figure 7-2. In this case the statistics are calculated with a set of size N = 32 (16 replicates in each of two rows on the chip), and the bias condition are in saturation (VDS = VGS = 1.8 V). 127 MR- --- Chip Means for - -- - - Chip Std for 3-Finger NMOS 3-Finger NMOS 164 x 10-6 3 1.62 2.8 1.58 2.4 1.56 2.2 *1.62. A 2 r1.54 1.8 1.52 Currents Std of Currents Figure 7-2: Spatial mean currents (left) and normalized standard deviations (right) per chip across wafer for NMOS 3-finger structure. N=32. The trend is not particularly clear across the wafer, but we observe that some of the higher current dies show less within chip variation for this particular test structure. To investigate if there is correlation we use equation (7.1), where we have two sets of data: X refers to the set of data of the mean currents in a chip and Y refers to the set of data of the standard deviation of the currents in a chip. In (7.1), x is the mean of the population X, y is the mean of the population Y, x, and yj are the values of the samples X and Y for chip i, sx and s, are the observed standard deviations for the samples X and Y, and n is the number of chips. n - x)(y_ - y) I(xi r= ( (7.1) x 100% (n - 1) sx s This correlation between the mean currents and their standard deviations for the 3finger single spaced NMOS transistor is not strong, only -24%. That is to say, although there are some spatial trends in both mean and within chip variation in Figure 7-2, the mean and variation are not strongly correlated with each other. The overall variation for the chip to chip means across the wafer is 1.65%. Figure 7-3 shows the spatial location for the combination of all "NMOS of different 128 :! -- -- W4 ~ -~-~ sizes" test structures combined. There are 33 different structures in this category with gate lengths from 0.18 gim, 0.36 pm and 0.54 gm, and gate widths of 0.5 gm to 3 gm (in increments of 0.25 gm), as explained in more detail in Section 3.1.1. There are 16 replicates for each one of the 33 different NMOS test structures in a chip, for a total of 528 samples per chip. The bias conditions are in saturation (VDS = VGS = 1.8 V). These plots show a correlation between the current means and the standard deviations per chip of -26%, again indicating only a weak relationship between within chip variation and chip mean. The trend in Figure 7-3 shows that most high current structures are in the north-east portion of the wafer, and most low current structures are found towards the center of the wafer. We find a normalized variation of 1.3% in the mean of the different dies. Chip to Chip Means for all Chip to Chip Std - NMOS Sizes for all NMOS Sizes 10 10-6 6.75 7.2 7 6.7 6.8 .45 6.6 6.66. 6.2 6.55 6 5.8 5.6 6.5 6.45 5.4 Currents Std of Currents Figure 7-3: Spatial saturation mean currents (left) and normalized standard deviations (right) per chip across wafer for all NMOS of different sizes. N=528. x10 3 Chip to Chip Means for NMOS VT 0.545 Chip to Chip Std for NMOS VT 10 0.545 0 53 0.525 7 0.52 0.- 6 0.05 .495 Voltage [V] 4 Std of Voltage [V] Figure 7-4: Spatial saturation mean currents (left) and standard deviations (right) for threshold voltage of NMOS transistors with 0.18 gm gate length. N=176. 129 Figure 7-4 reaffirms the trend observed in Figure 7-3, showing that the threshold voltage of the minimum sized NMOS transistors also follows the same pattern: the largest threshold voltages are towards the center of the wafer (the south-south west corner of our data) and some of the smallest threshold voltages are towards the north-east corner of the wafer. The threshold voltages plotted in Figure 7-4 are for all of the 0.18 Rm gate length NMOS test structures per chip. There are 11 different 0.18 pLm gate length NMOS test structures per chip, each with gate widths of 0.5 gm to 3 gm (in increments of 0.25 tm), as explained in more detail in Section 3.1.1. There are 16 replicates of each of the NMOS test structures per chip, hence the sample size is 176. The plots in Figure 7-4 show a correlation between the threshold voltage means and the standard deviations per chip of -33%. We plot fewer chips in these plots than in Figure 7-3 because some test chips are damaged in the testing process, and others were outliers. VT calculations are more prone to measurement glitches than the saturation measurements because of the voltage sweeps (many more data points sampled). The trend of having the "slowest" (lower current, highest VT) chips towards the center of the wafer may be due to wafer scale non-uniformity with typical edge versus center differences, which typically occurs on CMP and other processes. A possibility is a variation in oxide thickness across the wafer, which has a first order effect on threshold voltage variation. We find a range of 2.7 % in the mean of the different dies across the wafer. 7.2 Within Chip Spatial Analysis Repetition in the test chip enables spatial analysis per row. Figure 7-5 shows the spatial trends of the mean output saturation current (at VDS = VGs = 1.8 V) of the most 130 ENIENNNIENNNEEMN - replicated test structure, the 3-finger, minimum spacing NMOS transistor, which is replicated 16 times per row. These plots are the die spatial maps averaged across all 35 chips. Within Chip Current Mean for Both Rows X1 0-4 I1.600 1.595 1.590 1.585 1.580 N=70 Within Chip Current Mean for Row 1 X1 0-4 1.605 1.600 1.595 1.590 1.585 1.580 1.575 1.570 N=35 Within Chip Current Mean for Row 3 1.605 1.600 1.595 1.585 1.580 1.575 N=35 Figure 7-5: Within-row spatial analysis for most replicated NMOS, showing mean current values for the sixteen replicates per row for 35 chips. N=35. At the top of Figure 7-5 we show the aggregated data, averaged across both rows, and the bottom two plots show each of the two FEOL rows. In the individual rows, there is a soft trend of current extremes (highest and lowest) towards the edges. This may be 131 because of the low layout density around the edges, that can affect the etching or other processes. The low pattern density can be seen in the full chip photo, previously shown in Figure 2-7. The within row means (using the average "both rows" data) have a 0.38% variation, compared to 1.65% for the die to die variation of the same structure, as found in Section 7.1. Figure 7-6 shows the standard deviation of these test structures, which have no clear spatial trend. Within Chip Current Std for Both Rows 0-6 3.9 3.8 3.7 3.6 I N=70 Within Chip Current Std for Row 1 3.5 3.4 3.3 3.2 3.1 3.0 4.6 4.4 4.2 4.0 3.8 3.6 3.4 3.2 N=35 Within Chip Current Std for Row 3 X1 0 6 4.2 4.0 3.8 3.6 3.4 3.2 3.0 2.8 2.6 N=35 Figure 7-6: Within-row spatial analysis for most replicated NMOS, showing 1-sigma current values for the sixteen replicates per row for 35 chips. N=35. 132 Because we note no obvious trends in the within die spatial map, we look at individual dies to see if there is a small within die trend that differs depending on the location of the chip in the wafer. If there is, for example, a "slanted plane" within die trend arising from a wafer level "bowel" shaped uniformity, and we average all chips (e.g., including both left tilting planes with right tilting planes), the within die spatial information may be lost. We thus consider study the within die trends of three particular dies: one at the upper right hand side wafer edge (chip 17), one at the left-center edge (chip 38) and one that is towards the center of the wafer (chip 41), as illustrated in Figure 7-1. Figure 7-7 shows the plots for the spatial location of the most replicated NMOS test structure (three finger, minimum spacing device) at saturation bias conditions (VDS = VGS = 1.8 V). Each row presents the output current of each device in one row, with no averaging of any kind. The plots in Figure 7-7 also lack a clear spatial pattern, and we conclude that there is no obvious spatial trend within die. Because of the long dimensions of the test row, a trend showing a systematic voltage drop across the rows is a possibility because of voltage drops in the line. We use Kelvin sensing connections to prevent this systematic trend. Because such a trend is not observed in our analysis, it gives us confidence of the functionality of the feedback network to prevent voltage drops in the long interconnect lines. 133 Within Chip Currents Row I for Chip 17 10-4 1.505 1.6 1.595 1.59 1.585 1.58 1.575 1.57 Within Chip Currents Row 1 for Chip 38 10-4 1.605 1.6 1.595 1.59 1.585 1.58 1.575 1.57 Within Chip Currents Row I for Chip 41 10-4 1.505 1.6 1.595 1.59 1.585 1.58 1.575 1.57 Figure 7-7: Within-row spatial analysis for most replicated NMOS, showing saturation current values for the sixteen replicates per row in a specific chip. N=1. 7.3 Summary In this chapter we presented spatial trends across a wafer and across the FEOL rows of our test chip. The chip-to-chip spatial variation for aggregated NMOS structures shows a trend of increasing current diagonally, from the center of the wafer (which is the bottom edge of our sample) to the top right corner of the wafer. The spatial analysis of VT for NMOS transistors of minimum gate length also shows a similar trend, confirming that on 134 this particular wafer the slowest transistors tend to be towards the center. Within-chip analysis shows a subtle trend of extreme currents towards the edges of rows, but the lack of a strong trend helps to verify that the feedback network to prevent power losses works well. This chapter concludes the analysis of our test structure methodology and devices under test; Chapter 8 discusses how the architecture can be re-implemented or enhanced for use in other technologies. 135 136 . ... ...... Chapter 8 Suggestions for Future Implementations Our test chip methodology is efficient for use in measuring and characterizing transistors and interconnects, and future work can explore this methodology further. The test chip architecture can be more easily migrated to other technologies because of its hierarchical approach. In this chapter we offer some suggestions on how the design can be migrated and improved in future implementations, and for future technology generations. 8.1 Gate Leakage in Newer Technologies The 0.18 gm technology used for our test chip does not have the problem of high gate leakage currents. More modem technologies that suffer from this problem will also need a Kelvin sensing connection in the gate voltage terminal, as illustrated in Figure 8-1. This will ensure that the voltage at the gate nodes of the switches and DUTs is the intended one, and that it has not been decreased due to losses in the line. In addition, means for measurement of gate currents may be desirable, when these become appreciable in some bias conditions. VDD - VGATE Figure 8-1: Feedback implementation for technologies with high gate leakage currents. 137 8.2 Design of the On-Chip Feedback The failure of the on-chip feedback jeopardizes the reliability of all the measurements in the test chip, and it must be carefully designed. The sizing of the analog opamp will likely need to change for other technologies to guarantee proper feedback. Because of this, it is not recommended to build the on-chip feedback in a very young technology where basic parameters are unknown. Some preliminary characterization must be done before designing the on chip feedback in a given technology, so that it can be designed with phase margin, step response and gain that fulfill the specifications to ensure that the feedback will work. In a young technology that is not yet robust enough for on-chip feedback, the feedback can be implemented off-chip. On-chip feedback is, however, preferred to off chip as Kelvin sensing works best when the feedback is as close to the DUTs as possible, and it allows rapid measurements without additional setup. The present design has some trouble providing low enough drain voltages to our DUTs. We believe that the output stage of the operational amplifier can be eliminated. The output stage used in this design, a source follower, has a complicated active bias that can threaten the full output swing of the opamp, as seen in Figure 2-10. The revised opamp needs to have a large enough common source gain stage to provide the currents needed for all the DUTs. The resistive load is not an issue here as the opamp is connected, by feedback, to its input stage, which is a very high resistance node. 8.3 Design Review of FEOL Low Leakage Switches In the present design, we have only a single designed version of our FEOL switches. In future designs, it may be best to have two: one for PMOS and one for NMOS DUTs. 138 We turn the NMOS DUTs off in two ways, by having the source and drain at the same potential and by grounding its gate voltage, as shown in Figure 8-2. The PMOS off tiles, however, are only turned off by making their source voltage equivalent to their drain voltage. The PMOS tile can also be double turned off if the gate voltage were to be set to VDD in the off state, as shown in the revised tile in Figure 8-2. /en VGATE en /en VDD en en VGATE /en en VREF FEOL NMOS Tile /en VREF FEOL PMOS Tile Figure 8-2: Revised NMOS and PMOS FEOL tiles. The problem of not turning a tile off both ways is highlighted in Figure 8-3. If PMOS tiles have their gate voltages connected to ground and there is an offset between the source and the drain voltage of the off PMOS transistors at low values of Vref, the transistor will conduct current. That current will go to the output of the opamp as illustrated in Figure 8-3. This additional current can create an offset in the output of the opamp, hence failing to provide the accurate current. This can be easily fixed by having a separate gate voltage input line for the PMOS transistors, that will be kept at a high voltage instead of ground for the off tiles as shown in Figure 8-2. 139 Vref Il -+ len en /en VREF Figure 8-3: Problem with FEOL switches. 8.4 Off Tiles Leakage Currents The leakage current of the off tiles can affect the functionality of the test chip, because these currents add to the measurement of the currents of the desired DUT. In our case, these currents only affect the lower end of the weak inversion region, and we are able to gather the data we desire with the rest of the inversion region. However, new technologies may have higher "off' currents, or the designer may want lower "off' currents so that the weak inversion region can be more widely studied. One way to reduce the amount of leakage is by shortening the length of the test row. In our technology, each off tile has approximately 3 pA of leakage current (in simulation). This adds up to about 2 nA of leakage current for the 720 tiles when off at the same time. If we had only use a "subrow" which includes 90 tiles, for example, the leakage would be only 0.25 pA. Another way to reduce leakage is by increasing the size of the switch transistor, but in simulation we find that a 100% increase in the size of the switch (both width and length) will reduce the leakage by only 7%. 140 8.5 Summary Much has been learned from the design of this test chip, and in this chapter we provide several suggestions for its future implementation. We address the issue of gate leakage currents in more aggressive technologies, and we suggest adding a Kelvin sensing connection at the gate voltage terminal. We also address the issue of the redesign of the feedback opamp, which is likely the architecture component that will need the most change when migrated to other technologies. The low leakage switches were discussed and suggestions are given for improving their utilization and to minimize offstate leakage. Now that all aspects of our methodology have been discussed, we proceed to discuss and summarize the thesis contributions in Chapter 9. 141 142 Chapter 9 Conclusions This thesis focuses on the design, implementation, testing and analysis of a test chip that can extract analog measurements without directly probing a test structure, while multiplexing both the input and output signals of a large array of devices under test. It provides a methodology that emphasizes hierarchy so the architecture can be easily transported from one technology to another. In this last chapter we will summarize the contributions of the thesis, including a summary of the findings, and present suggestions for future work. 9.1 Thesis Contributions We have successfully designed, implemented and demonstrated a test chip that can gather analog (I-V) characteristics by multiplexing the use of pads. It enables tracking of trends and variations in transistor and interconnect structures, with results that show dependencies to common layout practices and process variation. We can summarize the contributions of the thesis in two areas: the design of the test chip and the analysis of the measurements taken. 9.1.] Test Chip Design and Implementation We presented the design of a test chip that can gather a large amount of analog characteristics in a much more limited area than traditional methods of testing. In it we tested a total of 2128 test structures per chip (FEOL and BEOL), as opposed to approximately 140 test structures that can be measured using traditional methods of testing in the same area (in a 3 x 2.33 mm area, with 50 pm pads and using five pads for two transistors, sharing one terminal with other devices). 143 The test chip is mainly composed of three blocks: the tile, row, and chip blocks. This methodology is powerful because its hierarchy allows the user to make hierarchical changes when migrating it to a new technology, and to modify the test chip (number of test structures, area, etc.) at the user's convenience. There are two different types of tiles, the front end of line (FEOL) tile and the back end of line (BEOL) tile. The FEOL tile includes a device under test (DUT) that is connected to two simple, novel low leakage pass gate switches, one in its gate and one in the drain terminal, ensuring the on and off states of the DUTs within an entire row. The tile also includes control circuitry (scan chain) to open and close the switches. The BEOL tile includes a CBCM structure connected to a metal capacitor; the tile input signals turn each of the four transistors in the tile on and off through low leakage switches. The BEOL tile also has a low leakage switch to connect its output current to the output line. Each row, a group of tiles, includes a carefully designed on-chip Kelvin sensing feedback circuit to guarantee the drain voltage that each DUT receives, which is critical to gather the desired VDS curves without probing each DUT directly. The feedback system is composed of an operational amplifier built in 3.3 V devices that can provide the output voltage and current needed to submit the front end of line DUTs to drain and gate voltage sweeps. It can also provide a steady voltage for a CBCM structure. The chip was fabricated using a 0.18 gm CMOS logic technology. 9.1.2 Testing Results andAnalysis of the Test Structures All test structures were successfully measured and tested. Layout dependent test structures show clear dependencies due to the layout practice used to implement them. Unlike previous work [1], we are able to more clearly identify where the source of 144 variation is, because we measure only one device or interconnect structure at a time, and because we can gather I-V curves from which parameters such as threshold voltage can be extracted. Structures with different orientation present an offset that is consistent between chips and between most structures laid out next to each other. Interestingly, NMOS and PMOS are affected inversely with orientation, probably because they are also affected inversely by residual strain due to shallow trench isolation (STI). Structures with different spacing between fingers show a mean shift consistent with a lithographic effect. Structures with different number of fingers and equivalent gate length increase their variability as the number of fingers increases and the size of each finger decreases. We are also able to study an array of transistors of different sizes to characterize their behavior and susceptibility to variation. We are able to show that variation changes according to transistor sizing, following Pelgrom's model [35], [41], and that variation in sub-threshold is much larger than when a transistor is on, particularly in transistors with smaller gate lengths. We extract threshold voltage from the voltage sweep current curves, and its variation also follows the Pelgrom model. Interconnect capacitances are measured, with values that match closely the values extracted from a parasitic extractor for two types of test structures (focusing on coupling and plane to plane capacitances). For the third type of structure, we are able to determine that the extractor overestimated the parasitic capacitance value, highlighting the need of having on-chip measurements to ensure proper calculation of parameters. Spatial analysis is able to highlight wafer-wide trends for FEOL structures. Within chip trends are not dominant, adding assurance that the voltage feedback did its work in 145 assuring a constant power voltage for all test structures within a row. 9.2 Future Work The successful testing and implementation of the test chip opens doors to use it for different analyses in future work. The methodology can be transported to a different technology by following the guidelines presented in Chapter 8. Below we suggest a few ideas on future work using this methodology. 9.2.1 On-chip Measurements If chip area is available, it would be interesting to use the test chip methodology presented in this thesis together with on-chip current measurements, such as in [11] and [20]. This will simplify testing and can simplify the design of on-chip feedback for the ground line. 9.2.2 Study of Sub-threshold Variation It was found in our analysis that variation in the sub-threshold regime is of great impact. It was, however, difficult to examine at the low end of the weak inversion region because of the leakage of the off tiles. It would be interesting to repeat the experiment with lower leakage currents, using the methods suggested in Section 8.4 or others, so that variation in this area can be studied for lower gate voltages. 9.2.3 Charge Based CapacitanceMeasurementStructures The "blank" charge based capacitance structures (without any connected capacitor test loads) were found to exhibit non-zero capacitance measurements. As discussed in Section 6.1, these capacitance offsets were subtracted from all other CBCM test structures, which removes most of the chip to chip spatial trends in these structures. Future work should consider extensions to the CBCM approach to correct for these 146 offsets, preferably on an individual structure by structure basis. 9.2.4 Coupling CapacitancesStructures We were unable to decouple with our test structures the impact on variation due to metal thickness and spacing in coupling capacitances. Future work may consider an experiment with a larger array of different distances between the spacing or the width of the metal. 9.2.5 Directly Probed Transistors It would be useful to have some transistors directly connected to pads within our test chip, to prove the functionality of the test rows. Analog characteristics of these directly probed transistors can be compared to those of individual transistors within the scan chain, to estimate the efficiency and accuracy of the measurements of the transistors using the test chip architecture. We attempted to implement such directly probed transistors in our test chip, but they appear to have been damaged during the fabrication, packaging, or testing process, and we were unable to gather meaningful transistor current curves from them. We did not protect the pads or terminals with ESD structures, and it is possible that electrostatic discharge resulted in destructive breakdown of the gate dielectric or other damage to these devices. In particular, we observe that these devices do not turn off (the source current remains large even at low gate voltages), and large currents (in the gA range) are measured in the gate terminal. Additional caution must be used when laying out these directly probed transistors, so they have adequate protection to avoid damage during fabrication, packaging or testing. 147 9.3 Summary of Contributions This thesis has focused on the design, successful testing and analysis of a test chip capable of gathering analog measurements with a limited number of pads and input signals. Its implementation is more area effective than traditional methods of directly probing individual transistors. The large amount of data allows us to characterize the devices under test, both transistors and interconnects. Ideas have also been given for future implementations of the test chip. 148 References [1] K. M. Gonzilez-Valentin, "Extraction of variation sources due to layout practices," Master's thesis, Massachusetts Institute of Technology, Cambridge, MA, 2002. Available through MIT Libraries, Institute Archives Microforms, Noncirculating Collection 3, Cambridge, MA. [2] J. S. Panganiban, "A Ring-Oscillator-Based Variation Test Chip," Master's Thesis, Massachusetts Institute of Technology, Cambridge, MA, 2002. Available through MIT Libraries, Institute Archives Microforms, Noncirculating Collection 3, Cambridge, MA. [3] K. Bernstein, D. J. Frank, A. E. Gattiker, W. Haensch, B. L. Ji, S. R. Nassif, E. J. Nowak, D. J. Pearson, and N. J. Rohrer, "High-performance CMOS variability in the 65-nm regime and beyond," IBM Journal of Research and Development, Vol. 50, No. 4/5, pp 433-449, 2006. [4] R. Lefferts and C. Jakubiec, "An Integrated Test Chip for the Complete Characterization and Monitoring of a 0.25um CMOS Technology that fits into five scribe line structures 150um by 5,000um," International Conference on Microelectronic Test Structures, pp. 59-63, 2003. [5] S. Nassif, "Modeling and analysis of manufacturing variations," IEEE Conference on Custom-Integrated Circuits, pp. 223-228, 2001. [6] International Conference on Microelectronic Test Structures Tutorial Short Course, 2007. [7] T. M. Mak and S. Nassif, "Process Variation and Stochastic Design and Test," IEEE Design & Test of Computers, Vol.23, No. 6, pp. 436-437, 2006. [8] K. Agarwal, F. Liu, C. McDowell, S. Nassif, K. Nowka, M. Palmer, D. Acharyya, and J. Plusquellic, "A Test Structure for Characterizing Local Device Mismatches," Symposium on VLSI Circuits, Digest of Technical Papers, pp. 6768, 2006. [9] K. Agarwal, D. Sylvester, D. Blaauw, F. Liu, S. Nassif, and S. Vrudhula, "Variational Delay Metrics for Interconnect Timing Analysis," Design Automation Conference, pp. 381-384, 2004. [10] K. Agarwal, S. Nassif, F. Liu, J. Hayes, and K. Nowka, "Rapid Characterization of Threshold Voltage Fluctuation in MOS Devices," International Conference on Microelectronic Test Structures, pp. 74-77, 2007. [11] N. Drego, A. Chandrakasan, and D. Boning, "A Test-Structure to Efficiently Study Threshold-Voltage Variation in Large MOSFET Arrays," IEEE International Symposium on Quality Electronic Design, pp. 281-286, 2007. 149 [12] J. Cassard, "A Sensitivity Analysis of SPICE Parameters Using an Eleven-Stage Ring Oscillator," IEEE Journal of Solid-State Circuits, Vol. 19, No. 1, pp. 130135, 1984. [13] M. Bhushan, A. Gattiker, M. Ketchen, and K. K. Das, "Ring Oscillators for CMOS Process Tuning and Variability Control," IEEE Transactions on Semiconductor Manufacturing, Vol. 19, No. 1, pp. 10-18, 2006. [14] M. Bhushan, M. B. Ketchen, S. Polonksy and A.Gattiker, "Ring Oscillator Based Technique for Measuring Variability Statistics," Proceedings of the International Conference on Microelectronic Test Structures, pp. 87-92, 2006. [15] M. Bhushan, M. Ketchen, and D. Pearson, "High Speed Test Structures for InLine Process Monitoring and Model Calibration," International Conference on Microelectronic Test Structures, pp 33-38, 2005. [16] M. B. Ketchen, M. Bhushan, "Product-representative "At speed" test structures for CMOS characterization," IBM Journal of Research and Development, Vol. 50, No. 4/5, 2006, pp 451-468. [17] L.T. Pang, B. Nikolic, "Impact of Layout on 90nm CMOS Process Parameter Fluctuations," Symposium on VLSI Circuits, Digest of Technical Papers, pp. 6970, 2006. [18] M. Quarantelli, S. Saxena, N. Dragone, J. A. Babcock, C. Hess, S. Minehane, S. Winters, J. Chen, H. Karbasi, and C. Guardiani, "Characterization and Modeling of MOSFET Mismatch Of a Deep Submicron Technology," International Conference on Microelectronic Test Structures, pp. 238-243, 2003. [19] S. Saxena, S. Minehane, J. Cheng, M. Sengupta, C. Hess, M. Quarantelli, G. M. Kramer, and M. Redford, "Test Structures and Analysis Techniques for Estimation of the Impact of Layout on MOSFET Performance and Variability," International Conference on Microelectronic Test Structures, pp. 263-266, 2004. [20] V. Wang and K. Shepard, private communication, April 2006. [21] International technology roadmap for semiconductors, http://public.itrs.net. [22] J. C. Chen, D. Sylvester, C. Hu, H. Aoki, S. Nakagawa, and S.-Y. Oh, "An OnChip, Interconnect Capacitance Characterization Method with Sub-Femto-Farad Resolution," International Conference on Microelectronic Test Structures, pp. 7780, 1997. [23] D. Sylvester, J. Chen, and C. Hu, "Investigation of Interconnect Capacitance Characterization Using Charge-Based Capacitive Measurement (CBCM) Technique and Three-Dimensional Simulation," IEEE Journal of Solid-State Circuits, Vol. 33, No. 3, pp. 449-453, 1998. 150 [24] A. Bogliolo, L. Vendrame, L. Bortesi, and E. Barachetti, "Charge-based on-chip measurement technique for the selective extraction of cross-coupling capacitances," IEEE Workshop on Signal Propagation on Interconnects, pp. 7577, 2002. [25] L. Dobrescu, M. Petrov, D. Dobrescu, and C. Ravariu, "Threshold Voltage Extraction Methods for MOS Transistors," CAS 2000 Proceedings of the International Semiconductor Conference, pp. 371-374, 2000. [26] D. Johns, K. Martin, Analog Integrated Circuit Design, Wiley, 1997. [27] R. Lefferts, "Transistor MUX Design," private communication, October 2003. [28] B. Froment, F. Paillardet, M. Bely, J. Cluzel, E. Granger, M. Haond, and L. Dugoujon, "Ultra low capacitance measurements in multilevel metallisation CMOS by using a built-in electron-meter," International Electron Devices Meeting, pp. 897-900, 1999. [29] D. Drain, Statistical Methods for Industrial Process Control, Chapman & Hall, NewYork, NY, 1997. [30] W.Mendenhall, R.L.Scheaffer, and D.D.Wackerly, Mathematical Statistics with Applications, Duxbury Press, Second Edition, 1981. [31] The Mathworks, Inc., Matlab (The Language of Technical Computing), Statistics Toolbox Documentation, 2007. [32] T. Mizuno, N. Sugiyama, H. Satake, S. Takagi, "Advanced SOI-MOSFETs with Strained-Si Channel for High Speed CMOS - Electron/Hole Mobility Enhancement -," 2000 Symposium on VLSI Technology Digest of Technical Papers, pp. 210-211. [33] S.E. Thompson, G. Sun, K. Wu, J. Lim, and T. Nishida, "Key Differences For Process-induced Uniaxial vs. Substrate-induced Biaxial Stressed Si and Ge Channel MOSFETs," IEEE International Electron Device Meeting Technical Digest, pp. 221-224, 2004. [34] Y. Tsividis, Operation and Modeling of the MOS Transistor, Oxford University Press, Second Edition, 1999. [35] M. J. M. Pelgrom, A. C. J. Duinmaijer, and A. P. G. Welbers, "Matching properties of MOS transistors," IEEE Journal of Solid-State Circuits, Vol. 24, No. 5, pp. 1433-1440, October 1989. [36] A. Chandrakasan, W. J. Bowhill, and F. Fox, Design of High-Performance Microprocessor Circuits, Piscataway, NJ: IEEE Press, 2001. 151 [37] T. Sakurai and A. R. Newton, "A Simple MOSFET Model for Circuit Analysis," IEEE Transactions on Electron Devices, Vol. 38, No. 4, pp. 887-894, Apr. 1991. [38] Y. Taur and T. H. Ning, Fundamentals of Modem VLSI Devices, Cambridge University Press, 1998. [39] S. Narendra, "Effect of MOSFET Threshold Voltage Variation on HighPerformance Circuits," PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, 2002. Available through MIT Libraries, Institute Archives Microforms, Cambridge, MA. [40] K. A. Bowman, B. L. Austin, J. C. Eble, X. Tang, and J. D. Meindl, "A Physical Alpha-Power Law MOSFET Model," IEEE Journal of Solid-State Circuits Vol. 34, No. 10, pp. 1410-1414, 1999. [41] M. Pelgrom, H. Tuinhout, and M. Vertregt, "Transistor Matching in Analog CMOS Applications," International Electron Devices Meeting, pp. 915-918, 1998. [42] S.C. Wong, T.G.Y. Lee, D.J. Ma, and C.J. Chao, "An empirical three-dimensional crossover capacitance model for multilevel interconnect VLSI circuits," IEEE Transactions of Semiconductor Manufacturing, Vol.13, No.2, pp. 219-227, May 2000. [43] L. Pileggi, H. Schmit, A. J. Strojwas, P. Gopalakrishnan,V. Kheterpal, A. Koorapaty, C. Patel, V. Rovner, and K. Y. Tong, "Exploring Regular Fabrics to Optimize the Performance-Cost Trade-Off," Proceedings of Design Automation Conference, pp. 782-787, June 2003. [44] J. Cohen, Statistical Power Analysis for The Behavioral Sciences, Lawrence Erlbaum Associates, Second Edition, 1988. 152 Appendix I Test Structures Nomenclature and Row Location FEOL Cells There are 90 different FEOL cells. These are the FEOL included in the design: SIN (N type structures of different sizes. All gate lengths are 0.18u, and the widths are from 0.50-3u in 0.25u increments) S2N (N type structures of different sizes. All gate lengths are 0.36u, and the widths are from 0.50-3u in 0.25u increments) S3N (N type structures of different sizes. All gate lengths are 0.54u, and the widths are from 0.50-3u in 0.25u increments) tilelp18wp50 tilelpI8wp75 tilelpl8wIpO tilelp18w1p 2 5 tilelpl8wlp50 tilelpl8wlp75 tilelpi 8w2p0 tilelpI8w2p25 tilelp I 8w2p50 tilelp 1 8w2p75 tilelp 1 8w3p0 tilelp36wp5O tilelp36wp75 tilelp36wlpO tilelp36wlp25 tilelp36wlp5O tilelp36wlp75 tilelp36w2pO tilelp36w2p25 tilelp36w2p5O tilelp36w2p75 tilelp36w3pO tilelp54wp5O tilelp54wp75 tilelp54wlpO tilelp54wlp25 tilelp54wlp5O tilelp54wlp75 tilelp54w2pO tilelp54w2p25 tilelp54w2p5O tilelp54w2p75 tilelp54w3pO 153 I tilePlpl8wp5O SIP (P type structures of different sizes. All gate lengths are 0.18u, and the widths are from 0.50-3u in 0.25u increments) S2P (P type structures of different sizes. All gate lengths are 0.36u, and the widths are from 0.50-3u in 0.25u increments) S3P (P type structures of different sizes. All gate lengths are 0.54u, and the widths are from 0.50-3u in 0.25u increments) tilePlp 18wp75 tilePlp18wlpO tilePIp18wlp25 tilePlp 18w lp50 tilePIp18wlp75 tilePlp 1 8w2p0 tilePlp 1 8w2p25 tilePlp 1 8w2p50 tilePlp 1 8w2p75 tilePlp 1 8w3p0 tilePlp36wp5O tilePlp36wp75 tilePlp36wipO tilePlp36wlp25 tilePlp36wlp5O tilePlp36wlp75 tilePlp36w2pO tilePlp36w2p25 tilePlp36w2p5O tilePlp36w2p75 tilePlp36w3pO tilePlp54wp50 tilePlp54wp75 tilePlp54wipO tilePlp54wlp25 tilePlp54wlp5O tilePlp54wlp75 tilePlp54w2pO tilePlp54w2p25 tilePlp54w2p5O tilePlp54w2p75 tilePlp54w3pO 154 OR NMOS and PMOS oriented vertically and horizontally. All are 1/.18u PS Three finger structures with different spacing between the fingers (minimum, 1.2x minimum, 1.5x minimum, 3x minimum) , NMOS and PMOS. All are 0.5u/. 18u per finger PF Structures with equivalent gate length but different number of fingers. All w=0.5u. Lengths are 3x minimum (1 finger, 2 finger and 3 finger), and 4x (1 finger, 2 fingers and 4 fingers) NMOS and PMOS tileNV tileNH tilePV tilePH tileNS Ix tileNSlp2x tileNSlp5x tileNS3x tilePSIx tilePSIp2x tilePS 1 p5x tilePS3x tileND3xif tileND3x2f tileND3x3f tilePD3xlf tilePD3x2f tilePD3x3f tileND4xlf tileND4x2f tileND4x4f tilePD4xif tilePD4xIf tilePD4x2f tilePD4x4f 155 BEOL Cells There are 23 unique BEOL test structures, which are: Combs Comb structures with minimum width and minimum space, I.2x minimum width and minimum space, and minimum width and 1.2x minimum space, in MI, M2, M3, M4, M5 Grids A combination of Combs and Planes, these are a horizontal metal layer under a vertical metal layer of higher order, all at minimum width and spacing Planes A large plane of metal-x over another one of metal x-1. The x- 1 plane is larger in case they misalign during the mask process CbcmTileCombM1 CbcmTileCombMlS CbcmTileCombMlW CbcmTileCombM2 CbcmTileCombM2S CbcmTileCombM2W CbcmTileCombM3 CbcmTileCombM3S CbcmTileCombM3W CbcmTileCombM4 CbcmTileCombM4S CbcmTileCombM4W CbcmTileCombM5 CbcmTileCombM5S CbcmTileCombM5W CbcmTileGridM2M1 CbcmTileGridM3M2 CbcmTileGridM4M3 CbcmTileGridM5M4 CbcmTilePlaneM2M1 CbcmTilePlaneM3M2 CbcmTilePlaneM4M3 CbcmTilePlaneM5M4 Rows The FEOL and BEOL rows are composed of the tiles described above. There are two FEOL rows in the chip. There are 728 FEOL tiles in a row, including eight blank tiles to prove the functionality of the scan chain. The row is subdivided in eight sub-rows. Each sub-row contains one of each of the unique FEOL structures. All FEOL structures are repeated eight times per row (the three finger, minimum spacing is repeated 16 times, as "tileNS Ix" and "tileND3x If'). The order in which the tiles are placed per sub-row varies, for a broader spatial analysis. This is the order of the FEOL sub-rows. 156 ROWA ROWB ROWC ROWD ROWE ROWF ROWG ROWH lp18wp50 Plp18wp5O NV NS1x lpl8wp50 Plp18wp5O NV NS1x lp18wp75 Plp18wp75 NH NSlp2x lpl8wp75 Plpl8wp75 NH NSlp2x lp18wlp0 Plp18wlp0 PV NSlp5x lp18wlp0 Plp18wlp0 PV NSlp5x p18w1p25 Plpl8wlp25 PH NS3x lp18w1p25 Plpl8wlp25 PH NS3x lpl8wlp50 Plpl8wlp5O NS1x PSix lp18wlp5O Plpl8wlp5O NS1x PSix lpI8wlp75 Plpl8wlp75 NSlp2x PSIp2x lp18w1p75 Plp18wlp75 NSlp2x PSlp2x lpl8w2pO Plpl8w2pO NSlp5x PSlp5x lpI8w2pO Plpl8w2p0 NS1p5x PSlp5x lp18w2p25 lp18w2p50 lp18w2p75 Plp18w2p25 NS3x PS3x NS3x PS3x PSIx ND3xlf Plpl8w2p50 PSIx ND3xlf Plpl8w2p75 PS p2x ND3x2f lp18w2p25 lp18w2p50 lp18w2p75 Plpl8w2p25 Plpl8w2p50 Plp18w2p75 PS p2x ND3x2f lpI8w3pO Plpl8w3pO PSlp5x ND3x3f lpl8w3pO Plpl8w3p0 PSlp5x ND3x3f lp36wp50 Plp36wp5O PS3x PD3xlf lp36wp5O Plp36wp50 PS3x PD3xlf lp36wp75 Plp36wp75 Blank PD3x2f Ip36wp75 Plp36wp75 Blank PD3x2f lp36wlpO lp36w1p25 1p36w1p50 Plp36wlpO ND3xlf PD3x3f lp36wlpO Plp36wlpO ND3xlf PD3x3f Plp36w1p25 ND3x2f ND4xlf Plp36w1p25 ND3x2f ND4xlf Plp36wlp50 ND3x3f ND4x2f Plp36wlp5O ND3x3f ND4x2f 1p36w1p75 Plp36wlp75 PD3xlf ND4x4f 1p36w1p25 lp36w1p50 1p36w1p75 Plp36w1p75 PD3xlf ND4x4f lp36w2p0 1p36w2p25 lp36w2p50 lp36w2p75 lp36w3p0 lp54wp50 lp54wp75 Plp36w2pO PD3x2f PD4xlf 1p36w2p0 Plp36w2pO PD3x2f PD4xlf PIp36w2p25 PD3x3f PD4x2f lp36w2p25 Plp36w2p25 PD3x3f PD4x2f Plp36w2p50 ND4xlf PD4x4f Plp36w2p50 ND4xlf PD4x4f Plp36w2p75 ND4x2f lpl8wp50 lp36w2p50 1p36w2p75 Plp36w2p75 ND4x2f Plp36w3pO ND4x4f lpl8wp75 Ip36w3p0 Plp36w3pO ND4x4f lpl8wp50 lpl8wp75 Plp54wp5O PD4xlf lp18wlp0 lp54wp5O Plp54wp5O PD4xlf lpl8wlpO Plp54wp75 PD4x2f lpl8wlp25 1p54wp75 Plp54wp75 PD4x2f lp54wlpO Plp54wlpO PD4x4f lpl8wlp50 lp54wlpO Plp54wlpO PD4x4f lp54w1p25 lp54w1p50 lp54wlp75 lp54w2p0 lp54w2p25 lp54w2p50 lp54w2p75 lp54w3p0 Plp54w1p25 lpl8wp50 lpl8wlp75 1p54w1p25 PIp54w1p25 lp18wp50 1p18w1p25 Ipl8wlp50 lp18w1p75 Plp54wlp5O lpl8wp75 lpI8w2pO lpl8wp75 lpI8w2pO lpl8wlpO Plp54w1p75 lpl8wlpO lp18w2p25 Plp54w2pO 1p18w1p25 lp54w2p0 Plp54w2pO lp18w1p25 lpl8w2p5O Plp54w2p25 lpl8wlp5O lp18w2p25 1p18w2p50 lp18w2p75 lp54w1p50 1p54w1p75 Plp54wlp5O Plp54w1p75 PIp54w2p25 lpl8wlp5O lp18w2p75 Plp54w2p50 lp18w1p75 Ip18w3p0 Plp54w2p50 1p18w1p75 lpl8w3p0 Plp54w2p75 lpl8w2pO lp36wp5O Plp54w2p75 lpl8w2pO lp36wp5O Plp54w3p0 lp36wp75 lp54w2p25 lp54w2p50 1p54w2p75 1p54w3p0 Plp54w3pO lp36wp75 lp36wlpO Plpl8wp50 Blank Plpl8wp75 NV 1p18w2p25 lp18w2p50 lp18w2p75 Plpl8wlpO NH lpl8w3pO Blank NV lp18w2p25 lp18w2p50 Plpl8wp50 NH 1p18w2p75 Plpl8wp75 PV lpl8w3pO Plp18wlpO PH lp36wp5O lp36w1p25 lp36w1p50 lp36wlp75 Plp18wlp25 PV lp36wp5O lp36wlpO lp36w1p25 lp36w1p50 lp36w1p75 Plpl8wlp25 NS1x lp36wp75 lp36w2pO Plpl8wlp50 PH lp36wp75 lp36w2pO PIpl8wlp50 NSlp2x lp36wlpO lp36w2p25 Plpl8wlp75 NS1x lp36wlp0 Plpl8wlp75 NSlp5x 1p36w1p25 1p36w2p50 PlpI8w2pO NSlp2x 1p36wlp25 Plpl8w2pO NS3x lp36wlp5O lp36w2p75 Plp18w2p25 NSlp5x lp36wlp5O lp36w2p25 lp36w2p50 1p36w2p75 Plpl8w2p25 PSIx lp36w1p75 1p36w3p0 Plpl8w2p50 NS3x lp36w1p75 lp36w3p0 Plpl8w2p5O PSlp2x lp36w2pO lp54wp5O Plp18w2p75 PSlx 1p36w2p0 lp54wp50 PIp18w2p75 PSlp5x lp36w2p25 lp54wp75 PlpI8w3pO PSlp2x lp36w2p25 lp54wp75 Plpl8w3pO PS3x lp36w2p50 lp54wlpO Plp36wp5O PSlp5x lp36w2p50 lp54wlpO Plp36wp50 ND3xlf lp36w2p75 lp54w1p25 Plp36wp75 PS3x lp36w2p75 lp54w1p25 Plp36wp75 ND3x2f lp36w3pO lp54wlp5O Plp36w0pO ND3xlf lp36w3p0 lp54wlp5O 157 Plp36wlpO ND3x3f 1p54wp50 lp54wlp75 Plp36wlp25 ND3x2f lp54wV5O 1p54wlp75 Plp36wlp25 PD3xlf 1p54wp75 1p)54w2p0 PIp36wlp5O ND3x3f Ip54wp75 1p54w2P0 Plp36wlp5O PD3x2f lp54wlpO 1p54w2p25 Plp36wlp75 PD3xlf lp54wlpO 1p54w2p25 Plp36wlp75 PD3x3f 1p54wlp25 lp54w2p50 Plp36w2pO PD3x2f lp54wlp25 1p54w2p50 Plp36w2pO ND4xlf 1p54wlp50 1p54w2p75 Plp36w2p25 PD3x3f 1p54wlp50 1p54w2p75 Plp36w2p25 ND4x2f 1p54wlp75 1p54w3p0 Plp36w2p50 ND4xlf lp54wlp75 l1p54w3p0 Plp36w2p50 ND4x4f 1p54w2p0 Plpl18wp50 Plp36w2p75 ND4x2f Ip54w2 0 Blank Plp36w2p75 PD4xlf 1p54w2p25 Plpl8wp75 Plp36w3pO ND4x4f lp54w2p25 Plpl8wp5O Plp36w3pO PD4x2f 1p54w2p50 Plpl8wlpO Plp54wp5O PD4xlf lp54w2p50 Ppw75 Plp54wp5O PD4x4f 1p54w2p75 Plpl8wlp25 Plp54wp75 PD4x2f 1p54w2p75 Plpl8wlpO Plp54wp75 Blank 1p54w3p0 Plpl8wlp5O Plp54wlpO PD4x4f 1p54w3p0 Plpl8wlp25 Plp54wlpO lpl8wp5O Plpl8wp5O PIpl8wlp75 Plp54wlp25 lpl8wp5O Plpl8wp50 Plpl8wlp5O Plp54wlp25 lpl8wp75 Plpl8wp75 PlplI8w2pO Plp54wlp5O lpl8wip75 PlplI8wp7 5 Plpl8wlp75 PIp54wlp5O lpl8wlpO Plpl8wlp0 Plp)18w2p25 Plp54wlp75 lpl8wlpO PIpl8wlpo Plpl8w2pO Plp54wlp75 1pl8wlp25 Plpl8wlp25 PlplI8w2p5O Plp54w2pO 1pt8wlp25 Plpl8wlp25 Plpl8w2p25 Plp54w2pO lP18wlp5O Plpl8wlp5O Plpl8w2p75 Plp54w2p25 lpl8wlp50 Plpl8wlp5O Plpl8w2p5O Plp54w2p25 lp18wlp75 Plpl8wlp75 Plpl8w3pO Plp54w2p50 1p18wlp75 Plpl8wlp75 PIp18w2p75 Plp54w2p50 lpl8w2pO Plpl8w2pO Pltp36wp5O Pl5wp5 lplI8w2pO Plpl18 2pO Plpl8w3pO Plp54w2p75 1p)18w2p25 Plp 18w2p25 Plp36wp75 Plp)54w3pO I18w2p)25 Ppwp25 Plp36wp5O Plp54w3pO 1p18w2p50 Plpl8w2p5O Plp36wlpO NV 1p1I8w2p50 Plpl8w2p5O Pl3w75 NV 1p 18w2p75 Plp18w2p75 Plp36wlp25 NH 1p18w2p75 Plpl8w2p75 Plp36wlpO NH lpl~w3pO Plpl8w3pO Plp36wlp5O PV lpl8w3pO PlplI8w3pO Plp36wlp25 PV 1p36wp50 Plp36wp)5O Plp36wlp75 PH lp36wp50 Plp36wp5O P1 36wp 50 PH lp36wp75 Plp36wp75 Plp36w2pO Blank lp36wp75 PIV36wp75 Plp36wlp75 NSlx lp36wlpO Plp36wlpO Pp6p25 NSlx lp36wlpO Plp36wlpO Plp36w pO NSlp2x 1p36wlp25 Plp36wIp25 Plp)36w2p50 NSlp2x lp36wlp25 Plp36wlp25 Plp36w2p25 NS lp~x 1p36wlp50 Plp36wlp50 Plp)36w2p75 NSlp5x 1p36wlp50 Plp36w ip5O Plp36w2p50 NS3x 1p36wlp75 Plp36wlp75 Plp36w3pO NS3x 1p36wlp75 Plp36wl7 P~3 PSix 1p36w2p0 Plp36w2pO Plp54wp5O PSix 1p36w2p0 Plp36w2pO Plp36w3pO PSlp2x 1p36w2p25 Plp36w2p25 P~5w75 PS lp2x 1p36w2p25 Plp36w p25 Plp54wp5O PSlp5x 1p36w2p50 Plp36w2p50 Plip54wlpO PSlp5x 1p36w2p50 Plp36w2p50 Plp)54wp75 PS3x lp36w2p75 Plp36w2p75 Plp54wlp25 PS3x 1p36w2p75 Plp36w2p75 Plp54wlpO ND3xlf lp36w3p0 Plp36w3pO PIp54wlp5O ND3xlf 1p36w3p0 Plp36w pO Plp)54wlp25 ND3x2f lp54wp5O PIP54wP5O Plp54wlp75 ND3x2f 15pPlp54wp5O Plp54wlp5O ND3x3f lp54wp75 Plp54wp75 Plp54w2pO ND3x3f lp54wp75 Plp54wp75 Plp54wlp75 PD3xlf lp54wlpO Plp54wlpO Plp)54w2p25 PD3xlf lp54wlpO Plp54wlpO Plp54w2pO PD3x2f lp54wlp25 Plp)54wlip25 Plp)54w2p50 PD3x2f 1p54wlp25 Plp54wlp25 Plp54w2p25 PD3x3f 1p54wlp50 Plp54wlp5O Plp54w2p75 PD3x3f 1p54wlp50 Plp54wl5 PIp54w2p50 ND4xlf 1p54wlp75_ Plp54wlp75 Plp54w3pO ND4xlf 1p54wlp75 Plp54wlp75 Plp54w2p75 ND4x2f 1p54w2p0 Plp54w2pO Blank ND4x2f 1p54w2p0 Plp54w2pO Plp54w3pO ND4x4f 1p54w2p25 Plp54w2p25 NV ND4x4f 1p54w2p25 Plp54w2p25 NV PI)4xlf 1p54w2p50 Plp54w p50 NH PD4xlf lp54w2p50 Plp54w2p50 NI-I PD4x2f lp54w2p75 Plp)54w2p75 PV PD4x2f 1p54w2p75 Plp54w2p75 PV PD4x4f 1p54w3p0 Plp54w3pO PH PD4x4f 1p54w3p0 Plp54w3pO PH p75 158 There are 2 BEOL rows on the chip, with 8-subrows, like in the FEOL case. These sub-rows have 42 tiles each. This is the order of the BEOL sub-rows: ROWA ROWB ROWC ROWD ROWE ROWF ROWG ROWH CombMl GridM2Ml PlaneM2Ml CombMl GridM2M1 PlaneM2M1 CombMl GridM2Ml CombMlS GridM3M2 PlaneM3M2 CombMlS GridM3M2 PlaneM3M2 CombMlS GridM3M2 CombMlW GridM4M3 PlaneM4M3 CombMlW GridM4M3 PlaneM4M3 CombMlW GridM4M3 CombM2 GridM5M4 PlaneM5M4 CombM2 GridM5M4 PlaneM5M4 CombM2 GridM5M4 CombM2S PlaneM2Ml CombMl CombM2S PlaneM2Ml CombMl CombM2S PlaneM2ML CombM2W PlaneM3M2 CombMlS CombM2W PlaneM3M2 CombMlS CombM2W PlaneM3M2 CombM3 PlaneM4M3 CombM2 CombM3 PlaneM4M3 CombM2 CombM3 PlaneM4M3 CombM3S PlaneM5M4 CombM2S CombM3S PlaneM5M4 CombM2S CombM3S PlaneM5M4 CombM3W CombM1 CombM3 CombM3W CombMl CombM3 CombM3W CombMl CombM4 CombMlS CombM3S CombM4 CombMlS CombM3S CombM4 CombMlS CombM4S CombM2 CombM4 CombM4S CombM2 CombM4 CombM4S CombM2 CombM4W CombM2S CombM4S CombM4W CombM2S CombM4S CombM4W CombM2S CombM5 CombM3 CombM5 CombM5 CombM3 CombM5 CombM5 CombM3 CombM5S CombM3S CombM5S CombM5S CombM3S CombM5S CombM5S CombM3S CombM5W CombM4 GridM2Ml CombM5W CombM4 GridM2Ml CombM5W CombM4 GridM2Ml CombM4S GridM3M2 GridM2Ml CombM4S GridM3M2 GridM2Ml CombM4S GridM3M2 CombM5 GridM4M3 GridM3M2 CombM5 GridM4M3 GridM3M2 CombM5 GridM4M3 CombM5S GridM5M4 GridM4M3 CombM5S GridM5M4 GridM4M3 CombM5S GridM5M4 GridM2Ml Blank GridM5M4 GridM2Ml Blank GridM5M4 GridM2Ml PlaneM2Ml GridM3M2 PlaneM2Ml PlaneM2Ml GridM3M2 PlaneM2M1 PlaneM2Ml GridM3M2 PlaneM3M2 GridM4M3 PlaneM3M2 PlaneM3M2 GridM4M3 PlaneM3M2 PlaneM3M2 GridM4M3 PlaneM4M3 GridM5M4 PlaneM4M3 PlaneM4M3 GridM5M4 PlaneM4M3 PlaneM4M3 GridM5M4 PlaneM5M4 Blank PlaneM5M4 PlaneM5M4 Blank PlaneM5M4 PlaneM5M4 Blank Blank PlaneM2Ml CombMl Blank PlaneM2MI CombMl Blank PlaneM2M1 CombMl PlaneM3M2 CombMlS CombMl PlaneM3M2 CombMlS CombMl PlaneM3M2 CombMlS PlaneM4M3 CombMlW CombMlS PlaneM4M3 CombM1W CombMlS PlaneM4M3 CombM2 PlaneM5M4 CombM2 CombM2 PlaneM5M4 CombM2 CombM2 PlaneM5M4 CombM2S CombMl CombM2S CombM2S CombMl CombM2S CombM2S CombMl CombM3 CombMlS CombM2W CombM3 CombMlS CombM2W CombM3 CombMlS CombM3S CombMlW CombM3 ConbM3S CombMlW CombM3 CombM3S CombM1W CombM4 CombM2 CombM3S CombM4 CombM2 CombM3S CombM4 CombM2 CombM4S CombM2S CombM3W CombM4S CombM2S CombM3W CombM4S CombM2S CombM5 CombM2W CombM4 CombM5 CombM2W CombM4 CombM5 CombM2W CombM5S CombM3 CombM4S CombM5S CombM3 CombM4S CombM5S CombM3 GridM2Ml CombM3S CombM4W GridM2Ml CombM3S CombM4W GridM2Ml CombM3S GridM3M2 CombM3W CombM5 GridM3M2 CombM3W CombM5 GridM3M2 CombM3W GridM4M3 CombM4 CombM5S GridM4M3 CombM4 CombM5S GridM4M3 CombM4 GiridM5M4 CombM4S CombM5W GridM5M4 CombM4S CombM5W GridM5M4 CombM4S PlaneM2M1 CombM4W GridM2Ml PlaneM2Ml CombM4W GridM2Ml PlaneM2M1 CombM4W PlaneM3M2 CombM5 GridM3M2 PlaneM3M2 CombM5 GridM3M2 PlaneM3M2 CombM5 PlaneM4M3 CombM5S GridM4M3 PlaneM4M3 CombM5S GridM4M3 PlaneM4M3 CombM5S PlaneM5M4 CombM5W GridM5M4 PlaneM5M4 CombM5W GridM5M4 PlaneM5M4 CombM5W 159 160 Appendix II Calculation for Crossover Capacitance Parameters: h2=8e-7 wl=0.23e-6 w2=0.28e-6 si=0.23e-6 s2=0.28e-6 t0=5300e-10 EOX =3.85e-1 1 Equations (from Wong et al. [42]: c2/Eox= 3.73*w2 0 .6*(s 1 *s2) 02 *(tl/(tl+(0.035*h2)))1- 64*(tI/(t1 +(0.85 1*S1 ))) 0 12*exp(-h2/(0.7*(s 1+.4*h2))) c I/ cox =w l*w2/h2 Results: cl = 3.0993e-018 c2 = 4.5094e-018 (cl+c2)*25*76= 1.4457e-014 161 162 Appendix III Layout Screen Shots These are some of the structures included in the test chip: FEOL: PMOS transistors in different sizes NMOS transistors in different sizes 163 Transistors with different number of poly fingers and different spacing between poly fingers PMOS and NMOS transistors with different orientation 164 BEOL: Metal 1 Comb Structures, Different Spacing and Different Widths 165 -9100"OW-- ---owd Metal 2 Comb Structure in Tile 166 Grid Structure (M2-M1) 167 Plane Structure (M2-M1) 168 Rows: FEOL Row BEOL Row 169