1
EE241 Spring 2010EE241 - Spring 2010Advanced Digital Integrated Circuits
Lecture 7: Variability
AnnouncementsHomework 1 posted, due Feb 18Project proposals due today
TitleHalf a pageFive (or more) references
2
2
OutlineLast lecture
Gate delaysStatic timing
This lectureIntroduction to variability
3
Design VariabilityDesign Variability
Sources and Impact on Design
3
Roadmap Acknowledges VariabilityInternational Technology Roadmap for Semiconductors2005 data
Node year 2007 2010 2013 2016 2019
DRAM ½ pitch [nm] 65 45 32 22 16
Total gate CD 3 [nm] 2.6 1.9 1.4 0.9 0.6
Lith h 3 [ ] 2 1 4 1 0 7 0 5
5
Lithography 3 [nm] 2 1.4 1 0.7 0.5
LER 3 [nm] 2 1.4 1 0.7 0.5
http://www.itrs.net/Common/2005ITRS/Home2005.htm
VariabilityNature of process variability
Within-die (WID), Die-to-die (D2D), Wafer-to-wafer (W2W), Lot-to-lot (L2L)Systematic vs randomSystematic vs. randomCorrelated vs. non-correlated
Spatial variability/correlationDevice parameters (CD, tox, …)Supply voltage, temperature
Temporal variability/correlation
6
Within-node scaling, Electromigration, Hot-electron effect, NBTI, self-heating, temperature, SOI history effect, supply voltage, crosstalk [Bernstein, IBM J. R&D, July/Sept 2006]
Known vs. unknown
4
Systematic and Random Device VariationsParameter Random Systematic
Channel Dopant Concentration Nch
Affects ϭVT [1] Non uniformity in the process ofdopant implantation, dosage, diffusion
G O S /S O & S O / S f fGate Oxide Thickness Tox
Si/SiO2 & SiO2/Poly-Si interface roughness[2]
Non uniformity in the process of oxide growth
Threshold Voltage VT(non Nch related)
Random anneal temperature and strain effects
Non-uniform annealing temperature[5](metal coverage over gate)Biaxial strain
Mobility μ Random strain distributions Systematic variation of strain in the Si due to STI, S/D area, contacts, gate density, etc
7
Gate Length L Line edge roughness (LER)[3] Lithography and etching:Proximity effects, orientation[4]
[1] D. Frank et al, VLSI Symposium, Jun. 1999 .[2] A. Asenov et al, IEEE Trans on Electron Devices, Jan. 2002.[3] P. Oldiges et al, SISPAD 2000, Sept. 2000.[4] M. Orshansky et al, IEEE Trans on CAD, May 2002.[5] Tuinhout et al, IEDM, Dec 1996
Sources of VariabilityTechnology
Front-end (Devices)Systematic and random variations in Ion, Ioff, C, …
Back-end (Interconnect)Systematic and random variations in R, C
EnvironmentSupply (IR drop, noise)
8
Temperature
5
Spatial Variability
Fab to fab Temperature
Global Local
Deployed environment
Lot to lot
Across wafer
Across reticle
Metal polishing
Transistor Ion, IoffLine-edge roughness
Dopant fluctuation
9
106 103 100 10-3 10-6 10-9
Across chipAcross block
After RohrerISSCC’06 tutorial
Film thickness
Spatial range [m]
Temporal Variability
Tech. node scaling Temperature
Technology Environment
Within-node scaling
Electromigration
NBTI
Hot carrier effect
Data stream
SOI history effectSelf heating
Supply noise
10
1012 103 100 10-3 10-6 10-12
Tooling changesLot-to-lot
After RohrerISSCC’06 tutorial
Coupling
Temporal range [s]10-9109 106
Charge
6
Systematic vs. Random VariationsSystematic
A systematic pattern can be traced down to lot-to-lot, wafer-to-f ithi ti l ithi di f l t t l twafer, within reticle, within die, from layout to layout,…
Within-die: usually spatially correlatedRandom
Random mismatch (dopant fluctuations, line edge roughness,…)Things that are systematic, but e.g. change with a very short time
t t (f t d thi b t it) O d ’t d t d it
11
constant (for us to do anything about it). Or we don’t unedrstand it well enough to model it as systematic. Or we don’t know it in advance (“How random is a coin toss?”).
Unknown
Corners
Within wafer
Within die
12
TypicalSlow Fast
7
Dealing with Systematic Variations
13Lin, DAC’06 tutorial
Systematic (?) Temporal VariabilityMetal 3 resistance over 3 months
14P. Habitz, DAC’06 tutorial
8
Chip Yield Depends on Inter-Gate Correlation
Variation remains constant with correlated gates, = 1
20%
d1 d2 dn
n stages
D D
1 / sqrt(n)0%
5%
10%
15%
0 2 4 6 8 10
/m
ean
of to
tal d
elay Variation is reduced with
non-correlated gates, = 0
15
Yield = Pr (sum of n delays < clock period) = 0 gives highest yield through averaging
0 2 4 6 8 10
Number of stages (n)
Non-correlated gates in a path reduce impact of variation
Chip Yield Depends on Inter-Path CorrelationMean delay increases as K increases for uncorrelated paths
D D
K u
ncor
rela
ted
path
s
Normalized Critical Path Delay
Nor
mal
ized
PD
F
0.8 0.9 1 1.1 1.20
K =1K =2K =10000
aP bP cPD D
a1 b2 c1D D
16Bowman et al, JSSC, Feb 2002 .
Yield = Pr (max delay of K paths < clock period) K = 1 results in highest yield
yMax delay of P paths
Correlated paths reduce impact of variation
9
Technology VariabilityLithographyDopantsLine edge roughnessFilm thicknessesNBTI
17
Optical Lithography: Variability Causes
i193 (immer.)=193nm
18
10
Lithography: Density EffectsIsolated
Dense Masks
Resist exposure thresholdLiso
Ldense
19
Denser features: More accurate line width and less variation. Dense lines are wider than isolated lines.
Brunner, ICP’2003
Lithography: OpticsDefocusLens aberrations:
Spherical aberrationsAstigmatismComa
20
Spherical aberrations - affect the reticle-level featuresStepper dependent
Coma affects individual featuresChip-location dependent
11
Lens Aberrations – Coma EffectComa effect: optical aberration due to lens imperfection.Causes mirrored structures to display non equivalent properties S t ti hift b t th 2 l t
Image of a circular dot shows a tail
Systematic shift between the 2 layouts
21
prints differently from
Step-and-Scan LithographySlit of light
Mask moved to the rightMask
Light sourceSlit direction:
Lens aberration in the slitCD’s more correlated to the right
Wafer moved to the left
slit
scan
Wafer
Optics
Scan direction:Dosage, scan speed and other fluctuationsCD’s less correlated
22
to the left
12
Lithography: FlareLight scattering and reflectionsMore stray light under dark features in the mask
Local flare depends on the density of chrome in the maskSurface
scattering
Resist exposure threshold
CD
Intensity
23
LensReflections With flare
No flare
Processing: Line-Edge Roughness
•Sources of line-edge roughness:• Fluctuations in the total dose due to quantization
24
• Fluctuations in the total dose due to quantization• Resist composition• Absorption positionsEffect:• Variation (random) in leakage and power
13
Random Dopant FluctuationsNumber of dopants is finite
25
Frank, IBM J R&D 2002
Random Dopant Fluctuations
Lg = 17nm, VDS = 0.7V Lg = 11nm, VDS = 0.7V
26VT = 23mV VT = 52mV
14
Oxide ThicknessSystematic variations +Roughness in the Si./SiO2 interfaceS ll ff t th RDFSmaller effect than RDF
27
Asenov, TED’2002
Negative Bias Temperature InstabilityPFET VTh’s shift in time, at high negative bias and elevated temperaturestemperaturesThe mechanism is thought to be the breaking of hydrogen-silicon bonds at the Si/SiO2 interface, creating surface traps and injecting positive hydrogen-related species into the oxide.
28
oxide.Also other charge trapping and hot-carrier defect generationSystematic + random shifts
Tsujikawa, IRPS’2003
15
Some Measurements: Gate Length (CD)Exhaustive ELM poly-CD measurements (280/field) :
†
Combine voltage drop measurements with sheet resistance measurements f i hb i V d P
29
J. Cain and C. Spanos, SPIE’03.
from neighboring Van der Pauw structures to get CD
Decomposition of Spatial CD Variation
= += +
Average Wafer Scaled Mask Errors Across-Field Variation
+ + +
30
Across-Wafer Variation
+ + +
Die-to-Die Variation “Random” Variation
Friedberg and Spanos, SPIE’05
16
Poly Density 90nm: RO Frequencies
31
• Max F between layouts > 10%• Within-die 3/ ~ 3.5%, weak dependency on density
L.T. Pang, VLSI’06
Poly Density 45nm: RO Frequencies
32
• Weak effect on performance. ΔF ~ 2%• Small shifts in NMOS leakage and bigger shifts in PMOS leakage
17
Longer Source/Drain Diffusion (45nm)
33
• Stronger effect observed, ΔF ~ 5%• No significant shift in leakage currents• Likely due to contact etch stop layer (CESL) induced stress
Spatial Correlation (), RO Frequency (90nm)
horizontal
Dotted lines = 99% confidence bounds
vertical Stronger for vertical gate, Vertical gate horizontal gate
1 3 5 7 9-0.10
0.10.20.30.4
1 2 3 4 5-0.10
0.10.20.30.4
ghor. col. spacing
34
• 3/ within-die variation ~ 3.5%• depends on gate orientation and spacing• No spatial correlation observed for ILEAK
1 3 5 7 9hor. column spacing (62.5m) ver. row spacing (100m)
1 2 3 4 5RO RO
RO
RORO
RO
18
Field Errors in the Mask (90nm)
5 10 15
2
4
6
8
105 10 15
2
4
6
8
10
RO Frequency ILEAK
Surface plots averaged over normalized data of 36 chips
5 10 15
2
4
6
8
105 10 15
2
4
6
8
10
RO Frequency
row
column
35
• RO with vertical gates shows ~1% frequency shift between 4 and 5 rows
• Rotated gates do not show this trend• Leakage current shows weak trend• Possible cause: mask stitching errors, speed changes
Mask/reticle
Ebeam field
Conclusions: What to Do?
Layout Extraction Consider proximity effects
Circuit Macros D2D > 15%. On-chip compensation circuits
L2L > 10%. Regular layouts, OPCCircuit Layout
36
Circuit Simulation
Floor Planning WID < 4%. Exploit spatial correlation in different directions
Incorporate systematic vs. random; correlation
19
What to do?
Layout Extraction VDD
Circuit Macros
Circuit Layout
InOut
37
Circuit Simulation
Floor Planning GND
What to do?
Layout Extraction
From J. Hartmann, ISSCC’2007 keynote
Circuit Macros
Circuit Layout Flip-flop:Regularlayout
38
Circuit Simulation
Floor Planning
20
Practical Variability-Aware DesignWill be covered in next several lectures
SRAM designStatistical timing analysis
39
Next LectureSRAM
40