Next Time on Parametric Yield. What Drives Worst Case Analysis?

Transcription

1 Next Time on Parametric Yield Current and Advanced DFM techniques Worst Case Files The role of process simulation (TCAD) Complete Process Characterization Statistical Design The economics of DFM 1 What Drives Worst Case Analysis? Basically, optimize an inverter design then, since most designs are just combinations of inverter-like gates it follows that the entire design would be OK even at the extreme points of process variation! p n Process domain: Vt Vt n n,, L L n n,, W n n,, kp kp n n,, Vt Vt p p,, L L p p,, W p p,, kp kp pp Tox, T, T, Vdd Performance domain: τ, τ, P Yield Body: P < x µw, t t < y nsec. 2

2 Simple Digital Worst Case I dsat sets power and speed in digital logic, so the extremes of I dsn and I dsp set the performance spread. Method Identify typical, fast and slow N, P. Extract the respective process parameters. Name the cases TT, SS, FF, SF, FS, respectively. Make sure performance meets specs at extremes. 3 Parametric Transistor Measurements 4

3 Simple Digital Worst Case (cont) Ids(p) SF FF SS FS Ids(n) Problems It implies that the yield body is convex. The Box might be unnecessarily big (over-design). Idsatextremes do not always map to performance extremes. Local variability is ignored (under-design). 5 Table Driven Digital Worst Case N and P device behavior is highly correlated! Process Ids(n) Ids(p) Related? Cox DW DL LatDiff msurface Yes Yes Yes no no Method Use table to relate process variability to Idsat deviations. Identify extremes through measurements. Extract the process parameters for each. Call them TT, SS, FF, SF, FS, respectively. Make sure performance goals are met by all. 6

4 Table Driven Worst Case (cont.) Ids(p) SF FF SS FS Problems It still implies that the Yield Body is convex. Idsat extremes do not always map to the actual performance spread. Local variability is ignored (under-design). Ids(n) 7 Working with Complete Parameter Sets Do not just focus on Idsat. Assume that all process parameters are varying. Independence cannot be assumed - use Principal Component Analysis (PCA) to transform the Problem. Method Extract process parameters from a population of devices. Find the correlation matrix of the process parameters. Apply PCA to transform to an independent parameter space. Use Constrained optimization to find the performance extremes (space is nicely convex). 8

5 Complete Parameter Sets (cont.) Problems Multiple extractions of correlated parameters. Performances analyzed at device - not circuit level. Local variability is ignored. 9 Working with Reduced Parameter Sets Assume that only few process parameters (V fb, T ox, L, W) and operational parameters (V dd, T) are varying independently. (No PCA necessary!) Use circuit simulation to define extremes. Method Develop special statistical device model. Perform n+1 circuit simulations at extremes. Establish linear approximation of yield body. Integrate Yield numerically. (Find Yield gradient, optimize yield) (Non-linear modeling and complex experimental designs also possible) 10

6 Reduced Parameter Sets (cont.) Problems Special device modeling considerations. Linear approximations good for high yield only. Local variability is ignored (no analog designs!) 11 Modern Worst Case Files... We now fully realize that the device level worst case files are just an intermediate abstraction to capture the process variability. We also know that process parameters are highly correlated. So, state-of-the art techniques attempt to approximate the ellipse that typically describes the distribution of device parameters. 12

7 Modern Worst Case files 13 Carefully designed arrays of simulations can reproduce statistics of process 14

8 Time out! What do all these Techniques need? Process Electrical Accurate process characterization! Deep understanding of how variability propagates from one level of abstraction to the next! Circuit Performance 15 An Example of Using TCAD in Process Characterization Accurate Statistical Process Variation Analysis for m CMOS with Advanced TCAD Methodology, Hisako Sato, Hisaaki Kunitomo, Katsumi Tsuneno, Kazutaka Mori, and Hiroo Masuda, Senior Member, IEEE, IEE TSM, Vol 11, No 4, November

9 The basic Macromodeling Idea 1. tune process/device simulator for good agreement with process. 2. Run the tuned simulator over a designed experiment. 3. Fit polynomial (macromodel) to simulator results. 4. Replace costly simulator with inexpensive (but rather local ) macromodel. 17 Good Statistical Match Can be Achieved... 18

10 How about Statistical Process Simulation? Low level, physical process parameters can be used as statistically independent disturbances. Method Identify process disturbances. Infer multi-level statistics of process disturbances. Use fats process and device simulation to generate device parameters. Global and local variations can be represented. 19 Statistical Process Simulation Compact Process / Device Models. Typically 1-D or compact 2-D models. The concept of process disturbances. Low level process parameters that are, almost by definition, independent from each other. 20

11 The Role of Statistical Process Simulation Diffusivity Oxide Growth Coef. Problems Process Characterization is difficult Modeling accuracy is an issue Monte Carlo simulation is still expensive Leff Tox 21 Possible to Model Statistics of Device Performance 22

12 For complex Analog Circuits... Matching matters. It is impossible to consider matching of individual devices It is practical to group devices, and consider matching across groups. 23 Why do we really loose Yield? Catastrophic defects due to contamination. One time departures from Statistical Control. 24

13 Problems with earlier Techniques Earlier Techniques assume: The entire fab line is under SPC. Critical steps can be characterized via measurements. VLSI Technologies, however, are often too short-lived and too complex to achieve SPC status! 25 A Binning Prediction Problem Create a model that relates yield to easy-to-measure, in-line and electrical parameters. Base model on process physics - so that is is valid even when the process is out of SPC. 26

14 The Binning Prediction (cont.) Use this model in conjunction with measurements to predict functional yield (speed binning) early on. This will give early feedback about the line and the product in it. Some of the costly performance testing might be avoided. Performance Prediction Model Binning Step Step Step P. P. Test 27 The Generic Process Model 28

15 The Specific IC Product Model 29 Simulated Statistical Distribution of Performances Energy vs. Read 0 Access time Read 0 vs. Read 1 Access Time 30

16 Yield Body Representation Using a linear model, each side of the yield body is a hyperplane of the form: Performance = A 0 +A 1 (N ld ) + A 2 (N tox )+... The adequacy of the model can be examined by plotting the distance of good and bad parts from the hyperplanes: d p =AX p Projecting constraints into two dimensional planes gives insight into how to improve the model. 31 Linear Approximation of the Yield Body 32

17 Test Patterns We need to characterize both interconnect and active areas of the chip. Test patterns include: Transistor arrays (three sizes of each type). Capacitors for three typical oxide thicknesses. Polysilicon structures for word line resistance. 33 In-Line and Electrical Measurements In order to use this model as a yield predictor, we have to feed to it easy to collect measurements. Standard test patterns are in use to collect Tox, Vt, kp, L, W, doping levels, poly resistivities. 34

18 Overall Binning Prediction Methodology 35 36

19 Issues in IC Manufacturability Importance and Causes of Yield Loss A Formulation of the DFM Problem Circuit Design for Manufacturability Yield Modeling Methods CIM and DFM 37 Projected Parametric Variation in Future Nodes 38

20 Projected Parametric Variation in Future Nodes Test Circuits used to facilitate Analysis Logic Interconnect 39 Projected Parametric Variation in Future Nodes % Sigma / Mean Performance Variation Trend - 8 Bit Mirror Adder % Variation caused by vth Delay and Energy variation caused by vth0 variation - 8 Bit Mirror Adder Technology Node (nm) Linear (% Sigma / Mean (Delay)) Linear (% Sigma / Mean (Energy)) Technology Node (nm) Linear (% Delay variation caused by vth0 variation ) Linear (% Energy variation caused by vth0 variation) % Variation caused by lint Delay and Energy variation caused by lint variation - 8 Bit Mirror Adder Technology Node (nm) Linear (% Delay variation caused by lint variation ) Linear (% Energy variation caused by lint variation ) 40

21 Projected Parametric Variation in Future Nodes Performance Variation Trend- Interconnect 140 Delay and Energy variation caused by width variation % Sigma / Mean Technology Node (nm) Linear (% Sigma / Mean (Delay)) Linear (% Sigma / Mean (Energy)) % Performance variation Technology (nm) Linear (% Delay variation induced by width variation) Linear (% Energy variation caused by width variation) 80 Delay and Energy variation caused by spacing variation Delay and Energy variation caused by thickness variation 80 % Performance variation % Performance variation Technology Node (nm) Technology Node (nm) Linear (% Delay variation caused by spacing variation) Linear (% Energy variation caused by spacing variation) Linear (% Delay variation caused by thickness variation) Linear (% Energy variation caused by thickness variation) 41 Projected Parametric Variation in Future Nodes % Performance variation Delay and Energy Variation caused by height variation Technology Node (nm) Linear (% Delay variation caused by height variation) Linear (% Energy variation caused by height variation) % Performance variation Delay and Energy variation caused by dielectric variation Technology Node (nm) Linear (% Delay variation caused by dielectric variation) Linear (% Energy variation caused by dielectric variation) % Performance variation Delay and Energy variation caused by interconnect variation Technology Node (nm) Linear (% Delay variation caused by interconnect variation) Linear (% Energy variation caused by interconnect variation) % Performance variation Delay and Energy variation caused by device variation Technology Node (nm) Linear (% Delay variation caused by device variation) Linear (% Energy variation caused by device variation) 42

22 wafer fab Let s Summarize... in-line tests real-time measurements Equipment Utilization step 1 step 2 step n e-test back-end wafer yield functional test die yield (functional) packaging binning/ parametric test field installation field data die yield (parametric) 43 IC production suffers from routine and assignable variability. Human errors, equipment failures Processing instabilities Material non-uniformities Substrate inhomogeneites Lithography spots Variability causes deformations Geometrical Electrical Lateral Global Vertical Local Spot defects Deformations have deterministic and random components, are global and/or local, can be independent or can interact. 44

23 Performance vs Yield 45 How does reduced variability help? 600k APC investment, recovered in two days... 46

24 Design Yield Prediction First we define functional (timing and other functionality constraints) and parametric (speed, power etc.) measures: Q F = {q F 1,q F 2,..., q F nf } T Q P = {q P 1,q P 2,..., q P np } T Then we define the space that contains both types of measures: F P S Q = S Q S Q The set of acceptable IC performances is given as: A Q = {Q S Q each q j acceptable j=1,2,..., n Q } n Q = n F + n P 47 Design Yield Prediction (cont.) All performances are determined from the state variables of the process (geometric and electrical parameters after wafer fabrication): A W = {X w S W Q (X w ) A Q } This is the acceptability region defined in the state variable space. The Yield is defined as: Y = g(x W ) f W (x W )dx W 1, x W A W f W (x W ) is the jpdf of X W and g(x W ) = { 0, otherwise Y = A W f W (x W )dx W 48

25 Design Yield Prediction (cont.) Let us now assume that Xwreally depends on the parameter sets defined as C (Controls), L (Layout), and D (disturbances). Given fixed values for C and L, then: Y = f D (δ)dδ A D (C 0,L 0 ) Let us further assume that C, L, D are separable (i.e. some affect performance and some functionality, but none affects both). Then: Y ' FUN = f D '(δ')dδ' A F D (C 0,L 0 ) and Y' PAR = f D ''(δ'')dδ'' A P D (C 0,L 0 ) 49 Manufacturing Yield vs. Design Yield Y M = N F N = N W N N F N W N P N F = Y W Y PT Y FT Because of test coverage limitations: Y M Y Because of imperfect separation: Y Y PAR ' Y FUN ' Finally, since we do not probe the actual circuit, we have: Y PRO Y PAR 50

26 Yield and Production Cost The objective is not to maximize yield but to minimize cost. Wafer Cost: C 1 = (N - N W )(c P *+ c W ) Probe Cost: C 2 = (N W - N P )(c PT + c P + c W ) Final Test Cost: C 3 = N P (c FT + c A + c PT +c P + c W ) Total Cost per chip sold: C = C 1 + C 2 + C 3 C N F = N P N F (c FT + c A + c PT +c P + c W ) + N W N F (1 - N P N W )(c PT + c P + c W ) + N N F (1 - N W N F )(c P *+ c W ) 51 Yield and Production Cost (cont) Finally, total cost per chip sold: C = (c FT + c A + c PT +c P + c W ) + Y P(1 - Y FT )(c FT + c A + c PT +c P + c W ) + N F Y M Y W (1 - Y PT )(c PT + c P + c W ) + 1 (1- Y W )(c P *+ c W ) Y M Y M This means that the yield maximization problem and the cost minimization problem are not equivalent! VLSI Design for Manufacturing: Yield Enhancement Director, Maly and Strojwas Kluwer Academic Publishers

27 In Conclusion Yield is good and Variability is bad. metrology statistical process control run-to-run and real-time control Must manipulate process steps to accommodate circuits and designs. modeling design of experiments Must keep cost of manufacturing low. Automation of product flow Automation of information management 53 Calculating IC Cost vs Defect Density An example... 54

28 The 1997 Roadmap (see transistor cost) Year Feature nm nm Area mm Density cm cm M 6.2M 10M 18M 39M 84M 180M Cost mc/tr technology ? 157? wafer size Function/milicent Overall Production Efficiency up by ~20X (!) from 1997 to Function/milicent 55 Negative Binomial (a widely accepted yield model) If f(d) follows a Gamma distribution, then: Y = 1 + A D α And if clustering becomes an issue, then: - α (α ~ 0.3-3) Y = Y o 1 + A D α - α where Yo is the gross cluster yield. 56

29 Typical Defect Size Distribution This means that defect density increases to to about 1/square~1/cube of of line width! 57 Line Yield, Memory Historical data data on on what what percentage of of WAFERS makes makes it it to to the the end. end. We We can can assume that that this this Yield Yield component will will be be a non-issue in in the the future... 58

30 Line Yield, CMOS Logic...even though CMOS does not have perfect wafer yield yet! 59 Memory Defect Density, µm A is is the the area. area. D is is the the defect defect density density (goes (goes up up to to 1/cube~1/square of of line line width width for for a given given clean clean room room class). class). Numbers Numbers here here are are for for class class See See next next slides slides for for data data points points from from older older technologies... Y = [ (1-e -AD )/AD ] 2 60

31 Logic Defect Density, µm CMOS Y = [ (1-e -AD )/AD ] 2 61 Logic Defect Density, µm CMOS Y = [ (1-e -AD )/AD ] 2 62

32 Bringing the Puzzle Pieces together... Year D CD Area GD/W Killer Def Wafer Size Density Yield Cent/Die 10^-6c/Tr Note how clean we must be! Note trade-offs between wafer size and defects! 20x improvement will not be easy Fall 2003 EE290H Tentative Weekly Schedule 1. Functional Yield of ICs and DFM. 2. Parametric Yield of ICs. 3. Yield Learning and Equipment Utilization. 4. Statistical Estimation and Hypothesis Testing. 5. Analysis of Variance. 6. Two-level factorials and Fractional factorial Experiments. 7. System Identification. 8. Parameter Estimation. 9. Statistical Process Control. Distribution of projects. (week 9) 10. Run-to-run control. 11. Real-time control. Quiz on Yield, Modeling and Control (week 12) 12. Off-line metrology -CD-SEM, Ellipsometry, Scatterometry 13. In-situ metrology - temperature, reflectometry, spectroscopy 14. The Computer-Integrated Manufacturing Infrastructure 15. Presentations of project results. IC Yield & Performance Process Modeling Process Control Metrology Manufacturing Enterprise 64