Litho and Design: Moore Close Than Ever Vivek Singh Intel Fellow Director, Computational Lithography Technology Manufacturing Group Intel Corporation
ISPD , March 2011
Gasp!
V. Singh, ISPD , March 2011
2
Outline The frenetic pace of Moore’s Law, and why we put up with it! How the coco-optimization of design and lithography have enabled Moore How computational lithography is squeezing more juice out of steppers
V. Singh, ISPD , March 2011
3
The Economics of Moore’s Law 1010
10
As the number of transistors goes UP
100
The $ Benefit:
10-1
108
10-2
107
Price per transistor goes DOWN
10-3 10-4 10-5
106 105
10-6
104
10-7
103 ’70
The Cost:
109
’80
’90
FAB PILOT LINE R&D PROCESS TEAM
Source: WSTS/Gartner/Intel
’00
$4 B $1-2 B $0.5-1 B
Source: Intel
Investment & co-optimized execution required to maximize the benefit V. Singh, ISPD , March 2011
4
Intel’s Silicon R&D Pipeline
15 nm
22 nm
Research
Pathfinding
32 nm
45 nm
Development Manufacturing
Continuous flow of new technologies from research to development to manufacturing
V. Singh, ISPD , March 2011
5
On--time 2 Year Cycles On 90 nm 2003
65 nm 2005
45 nm 2007
32 nm 2009
22 nm 2011
The image cannot be display ed. Your computer may not hav e enough memory to open the image, or the image may hav e been corrupted. Restart y our computer, and then open the file again. If the red x still appears, y ou may hav e to delete the image and then insert it again.
In development
V. Singh, ISPD , March 2011
6
Defect Density Trends 90nm
65nm
45nm
32nm
Defect Density (Log Scale)
2001
Higher Yield
2002
2003
2004
2005
2006
2007
2008
2009
Continuing to improve wafer defect density is crucial to achieving historical yield trends
V. Singh, ISPD , March 2011
7
Breakthroughs in Technology Information Turns 32nm 45nm
Information Turns
65nm 90nm
2001
2002
2003
2004
2005
2006
2007
2008
1.7 X improvement in Information turns Faster information back to process development and design teams V. Singh, ISPD , March 2011
8
SRAM Cell Size Scaling
Cell Area (um2)
10
65 nm, 0.570 um2
1
45 nm, 0.346 um2 0.1
0.5x every 2 years
32 nm, 0.171 um2 0.01 1995
2000
2005
2010
2015
22 nm, 0.092 um2
All Patterned with 193nm!! V. Singh, ISPD , March 2011
9
2 YEARS
TICK
2 YEARS
TICK TOCK
NEHALEM
2 YEARS
Tick--tock: an example from Intel of process Tick and design marching in cadence
TICK
WESTMERE
Pentium® D, Xeon™, Core™ processor
65nm
TOCK
Core 2 processor, Xeon processor
PENRYN Family 45nm
32nm
TOCK
SANDY BRIDGE
All product information and dates are preliminary and subject to change without notice
V. Singh, ISPD , March 2011
10
Design Rule Definition Process: How Litho and Design are connected 1 D Pitch target set by density goals Learning from previous process extrapolated OPC model Illumination techniques
OPC/litho test masks DR Modeling And Process evaluation
Modify older test chip
DOF, MEEF, etc
Enhancement techniques
Layout studies
Photo resist evaluation
Design rules
X test chip
Product Evaluation
CAD tools
C. Webb, SPIE 2007
V. Singh, ISPD , March 2011
11
Rule Stability 120.00% 100.00%
X Chip Tape out
Growth 80.00%
in Number of
60.00% 40.00%
Rules
First Design
Debug
Ramp
20.00%
Process Development 0.00%
-2
-1
+2
+1
0
Relative Year
C. Webb, SPIE 2007
V. Singh, ISPD , March 2011
12
130nm
SRAM Bit Variable poly pitch and line widths X and Y transistor orientation Design rule definition was primarily done through simple scaling of rules from previous generation Limited modeling of layout and design rules C. Webb, SPIE 2007
V. Singh, ISPD , March 2011
13
90nm 47% increase in number of poly rules
SRAM Bit
All Devices in one orientation except SRAM Complex rules did not impact transistor density – Modeling and layout study effort increased from 130nm generation C. Webb, SPIE 2007
V. Singh, ISPD , March 2011
14
65nm 65% increase in number of rules
All Devices in one orientation including SRAM Rules to enable phase shift masks Different rules for minimum pitch, larger poly pitch and poly routing Layout more difficult but no significant density impact
SRAM
C. Webb, SPIE 2007
V. Singh, ISPD , March 2011
15
Problems with 65nm poly layout
Two direction routing
Poly corners affecting devices
Multiple poly widths
Variable poly pitch
Could all logic poly layout be simple like the SRAM bit?
C. Webb, SPIE 2007
V. Singh, ISPD , March 2011
16
45nm Logic Layout 37% reduction in number of rules
One direction Trench contact local routing replaces orthogonal to gate poly routing
One Pitch with poly on grid No significant density impact Design had to adapt to limited channel length choices and new layout style C. Webb, SPIE 2007 V. Singh, ISPD , March 2011
17
Layout has adapted to litho constraints, without affecting Moore’s march 65 nm Layout Style
32 nm Layout Style
• Bi-directional features
• Uni-directional features
• Varied gate dimensions
• Uniform gate dimension
• Varied pitches
• Gridded layout M. Bohr, ISCC, 2009 V. Singh, ISPD , March 2011
18
What matters is Best Total Cost 100
10
Cell Area 2 (um )
SRAM Cell Size
SRAM - Functional silicon in Jan ‘06
Density reduction at Lowest cost and TTM 0.5x every 2 years 45nm Cell 0.346 um2
1
0.346 µm2 Cell 193 nm Dry Lithography
0.1 1992 1994 1996 1998 2000 2002 2004 2006 2008 Forecast
1.3
Critical Layer Lithography Cost Comparison 45nm Node assuming equal yield
1.2 1.1 1 0.9 193nm Dry
193nm Immersion V. Singh, ISPD , March 2011
19
Multi--mask patterning – a friend in need Multi As wafer dimensions reach optical limits, multimulti-mask patterning came to rescue Works well for simple repeated patterns Not so well with general patterns
Does not resolve
Overlay issues
Problem at corners
Need complex design rules to capture multi-mask conflicts Design tools need to be multi-mask aware to minimize impact
Modify design – does it impact performance?
V. Singh, ISPD , March 2011
20
One implication of Double-patterning Misalignment
Normalized IDSAT
Print 1
Print 2
Registration (nm)
K. Kuhn, ICVC, 2009
Misalignment between the 2 exposures is a crucial liability for this technique and can limit its usability Transistor parameters can be affected by asymmetry between the source and drain regions V. Singh, ISPD , March 2011
21
A B
A B
A B
B
SRAM Vccmin
Pitch doubling and gate CD matching
860 1262 1264 1266
C. Kenyon, TOK conf., Dec. 2008
Gate CD mismatch σ
Pitch doubling eliminates the close correlation which currently exists between the CDs of adjacent gates This has implications for memory cells and other circuits which depend upon this CD matching V. Singh, ISPD , March 2011
22
NGL solutions can help contain rising cost of double patterning
Source: ITRS 2009 V. Singh, ISPD , March 2011
23
Computational lithography – enabling Moore 1
1000 436 nm
365 nm
248 nm
Lithography Wavelength 193 nm
micron 0.1
100 nm Feature Size
0.01 1980
1990
65 nm 45 nm 32 nm 22 nm 15 nm
2000
2010
EUV 13 nm
10 2020
Computational lithography innovations and co-optimization with design necessary to bridge this gap V. Singh, ISPD , March 2011
24
2D Effects Become Important
Rule-based serif placement were first instance of on-mask pattern manipulation to deliver design intent S. Sivakumar, SPIE, 2011 V. Singh, ISPD , March 2011
25
One CL product: pixelated masks
Model black-box
Design Pattern
Pixelated mask
SEM image of wafer
Atomic Force Microscope Picture of the mask V. Singh, ISPD , March 2011
26
A FullFull-Chip View is worth a trillion pixels
Note: some cells are not visible due to graphics culling Top-cell pixels are not displayed Only 0-deg pixels are displayed
V. Singh, ISPD , March 2011
27
Convert pixelated mask to more manufacturable mask, while preserving performance pixelated mask
More manufacturable mask
V. Singh, ISPD , March 2011
28
Two pronged approach: simplify inverse masks improve mask shop capabilities
V. Singh, ISPD , March 2011
29
Optimizing Source and Mask Source Optimization
Source Mask Optimization
+
Computational Lithography is a key enabler for extracting maximum resolution from existing technology
V. Singh, ISPD , March 2011
30
Ultimately, the proof is in the pudding. Exotic techniques need to be viable for production. Atomic Force Microscope Picture of Pixelated Phase Mask
SEM Picture of 65nm MT1 resist Pattern from Pixelated Phase Mask
Defect free Pixelated Phase Masks produced, Leading 65nm node Microprocessor patterned at MT1 with Pixelated Phase Mask yielded close to manufacturing baseline
Computational lithography is helping extract more resolution from existing technology V. Singh, ISPD , March 2011
31
Other things being equal, you can measure CL innovation through compute cost
10000000
Relative cost
1000000
A new node increases compute cost due to increase in: number of OPC layers density required computational accuracy model accuracy needs convergence difficulty Number of OPC features
Cost savings from CL innovations
100000 ? 10000 1000 100 10
1 65
45
32
22
15
Technology node V. Singh, ISPD , March 2011
32
Future Technology Directions
CMOS
2000
More Moore
2010
2020
2030
New device technology will be needed by 2020 M. Bohr, SPIE, 2011 V. Singh, ISPD , March 2011
33
Future Technology Directions New Applications Digital + analog + optical Silicon + III-V Electrical + Mechanical
CMOS
2000
More Moore
2010
2020
2030
New device technology will be needed by 2020 M. Bohr, SPIE, 2011 V. Singh, ISPD , March 2011
34
Future Technology Directions New Applications Digital + analog + optical Silicon + III-V Electrical + Mechanical
CMOS
More Moore New Technologies Carbon Based? Spintronics? Magnetic Domains? Molecular Switches?
2000
2010
2020
2030
New device technology will be needed by 2020 M. Bohr, SPIE, 2011 V. Singh, ISPD , March 2011
35
What will be printed?
V. Singh, ISPD , March 2011
36
V. Singh, ISPD , March 2011
37
Conclusion
“No exponential is forever … but we can delay ‘forever’.”
Gordon Moore ISSCC 2003
V. Singh, ISPD , March 2011
38