Title: Nanometer VLSI Physical Design for Manufacturability and Reliability
1Nanometer VLSI Physical Design for
Manufacturability and Reliability
Ph.D. Proposal May 3rd, 2007
- Minsik Cho
- Dept. of Electrical and Computer Engineering
- The University of Texas at Austin
- thyeros_at_cerc.utexas.edu
2Outline
- Introduction
- Introduction VLSI Routing and DFM
- Manufacturability-driven routing
- BoxRouter for global routing framework
- Wire density-driven global routing for CMP
variation and timing - TROY Track routing with yield-driven wire
planning - Litho-driven detailed routing
- Conclusion
3Modern VLSI Design
- VLSI Physical Design
- Last design step before manufacturing
- Big impact on performance, power, noise, and
manufacturability - Different designs lead to different
manufacturability. - Better yield and low cost from better
manufacturability
4Routing in VLSI Physical Design
a
c
a
Placement
b
Routing
Routing
a
d
d
b
a
c
c
590/65/45nm DFM Issues
2007
2005
2006
45 nm
65 nm
90 nm
130nm
Courtesy Cadence, 2007
- DFM(Y) Design for Manufacturability (Yield)
- Chemical-mechanical polishing (CMP), particle
defects (CAA), and lithography are key
contributors to DFM. - - Dr. Scheffer (Cadence), Dr.
Domic (VP, Synopsys), Dr. Puri (IBM)
6Sub-wavelength Lithography
- Minimum feature size is scaling down faster than
wavelength. - Nanolithography (e.g., Nanoimprint, E-BEAM, EUVL)
are not mature/economical for mass production
yet. - It looks like optical lithography with help of
immersion and RET will be a work horse for next
510 years.
7RET Optical Proximity Correction (OPC)
Original mask
Courtesy Intel, 2006
- Insufficient critical dimension causes not only
circuit failure but also timing degradation. - Additional polygons are added to enhance
resolution. - Need to run computationally expensive OPC tools.
- Mask data volume is increased due to more
features.
8Chemical-mechanical Polishing (CMP)
Topography variation translated into focus
variation for lines which results in width
variation
- CMP (Chemical-Mechanical Polishing)
- Key multilevel metallization technique for copper
interconnect - CMP (topography) variation
- Performance degradation due to increased wire
resistance - Depth of focus issues due to non-uniform surface
- Systematic variation due to non-uniform metal
density distribution - Dummy fill increases mask volume and coupling
capacitance.
9Random Defects
open
short
- Two kinds of random defects
- Open defects due to missing material
- Short defects due to additional material
- Technology scaling makes a design more vulnerable
to random defects. - More (smaller) defects become smaller than
minimum feature size.
10Yield Loss Projection Scary
Critical Area
CMP Lithography
Courtesy IBS
Need to do something for CMP, Critical Area, and
Lithography
11Outline
- Introduction
- Introduction to DFM and VLSI Routing
- Manufacturability-driven routing
- BoxRouter for global routing framework
- Wire density-driven global routing for CMP
variation and timing - TROY Track routing with yield-driven wire
planning - Litho-driven detailed routing
- Conclusion
12Rule-based Approach
Cong, ICCAD06
- Conventional approach to separate design from
manufacturing RULES - Rules are starting running out of steam from 65nm
- Exploding number of rules
- VERY complicated rules (Rule book for 90nm has
600 pages) - Not accurate any more
13Model-based Approach
Lithography enhancement
CMP variation optimization
Critical area minimization
- The most effective VLSI design step for DFM
- Clearer picture about the final design layout
than any other steps - Those DFM issues are tightly coupled with wires.
- Significant design flexibility for wiring
- CMP variation optimization
- Minimum effective window (20x20µm2) and global
effect - Global routing plans approximate routing path.
- Critical area minimization
- Adjacent parallel wires contribute majority of
critical area. - Track router has decent flexibility in
optimization with wire adjacency information. - Lithography enhancement
- Effective windows (1µm2 ) and local effect
- Detailed router performs localized connection
between pins/wires.
14Outline
- Introduction
- Introduction to DFM and VLSI Routing
- Manufacturability-driven routing
- BoxRouter for global routing framework
- Wire density-driven global routing for CMP
variation and timing - TROY Track routing with yield-driven wire
planning - Litho-driven detailed routing
- Conclusion
15BoxRouter Overview
- Incremental Box Expansion
- From the most congested region
- Progressive ILP
- adaptive maze routing
- PostRouting
- Negotiation-based approach
- 2nd place in ISPD07 routing contest
Global routing cell (G-cell)
16Overall Flow of BoxRouter
Minimum Steiner Tree Net Decomposition (Node
shifting)
PreRouting Initial Box
BoxRouting
2D Global Routing
Robust Negotiation-based ReRouting (Topology-aware
wire ripup)
Greedy Wirelength/Via Minimization
Via/Blockage-aware Layer Assignment (Progressive
via/blockage-aware ILP)
Greedy Wirelength/Via Minimization
17PreRouting
- Preroute as many flat wires as possible
- 60 of wirelength can be prerouted
- Benefits
- Overall congestion captured
- Runtime improved
- Starting Box of BoxRouting
18BoxRouting in Action
ILP returns the optimal (high quality) solution.
However, in general, it is believed to be
unacceptably slow. Is runtime OK?
routing cap. constraint
YES!
solution constraint
19Progressive ILP with Box Expansion
- Progressive ILP
- Incorporate the solution from inner boxes
- Solve it between two consecutive boxes
20Proposed ILP vs. Conventional ILP
max
s.t
- Conventional ILP formulation
- Minimize max overflow
- Overflow should be fixed by maze router
- Proposed ILP formulation
- Maximize routed nets (min. total overflow)
- Unrouted wire should handled by maze router
Yang ICCAD01
21Proposed ILP vs. Conventional ILP
GNU Linear Programming Kit 4.10, Industrial
circuit (7mm x 7mm)
22Via/Blockage-Aware Layer Assignment
- Layer assignment follows 2D global routing.
- Most of advanced technology have multiple layers.
- Via provides electrical connection between wires
in different layers. - Via has higher probability of failure (e.g.,
redundant via)
23Layer Assignment Example
- The number of vias is decided by layer assignment
and blockages
24Via/Blockage-aware Layer Assignment
for every steiner point s of each net i
min
assigned only once
compute layer
get top/bottom layer of steiner point
pins on M1
routing cap. constraint
- Unassigned wires will be picked up by simple maze
routing which only shuttles between layers.
25Routability Improvement
0 overflow
- Only BoxRouter completes ISPD98 IBM benchmark.
26Wirelength Improvement
- BoxRouter shows the minimum wirelength.
27Outline
- Introduction
- Introduction to DFM and VLSI Routing
- Manufacturability-driven routing
- BoxRouter for global routing framework
- Wire density-driven global routing for CMP
variation and timing - TROY Track routing with yield-driven wire
planning - Litho-driven detailed routing
- Conclusion
28Predictive CMP Model For Global Routing
- Global routing is the first routing planning
phase - Predictive and fast CMP model is required
29Predictive CMP Model (Cu Thickness)
- Cu Thickness is systematically dependent on metal
density
30Predictive CMP Model (Metal Density)
- Metal density Wire density Dummy fill density
- Look-up table for fast computation
31Global Routing Flow With Predictive CMP Model
Dummy Fill Density From Lookup Table
Wire Density
Metal Density Wire Density Dummy Fill Density
Cu Thickness
Cu Thickness
Predictive CMP Model for Global Routing
- Predictive CMP Model guides global router
- More uniform wire distribution
- Consider metal blockages (power/ground rail, IPs)
32CMP Variation Improvement
- On average 7.5 reduction
- Up to 10.1 reduction
33Outline
- Introduction
- Introduction to DFM and VLSI Routing
- Manufacturability-driven routing
- BoxRouter for global routing framework
- Wire density-driven global routing for CMP
variation and timing - TROY Track routing with yield-driven wire
planning - Litho-driven detailed routing
- Conclusion
34Critical Area For Yield
C
A
B
- Random defect can cause open/short defect
- Wire planning for critical area reduction
- Defect size distribution
- Chance of getting larger defect decreases rapidly
TCAD85 - Concurrent optimization for open/short defect
- Larger wire width for open, but larger spacing
for short defect - Limited chip area
35TROY Overview
- TROY is a track router for yield enhancement
- Find initial solution through interval packing
- Wire ordering to minimize overlapped wirelength
between neighbors - Wire sizing/spacing to minimize critical areas
- Contribution of TROY
- The impact detailed wirelength on critical area
is considered - SOCP provides globally optimal wire size/spacing
for entire layer for the given objective in near
linear time
36Math for Critical Area
Defect size distribution r 3
Critical are due to open defect
Critical are due to short defect
Probability of failure (POF) for open/short
defects
Approximated POF
- POF is convex function.
- Global optimal can be found, but POF is
complicated. - Approximated POF is also convex, but easy enough
to enable SOCP.
37Yield-Driven Track Routing
POF for short
POF for open
- Integer nonlinear programming
- Extremely hard to solve
- Integer variable for the wire order (which is
above/below which) - Key observation
- Objective is convex, which can be further written
in rotated conic form if the approximated POF is
adopted. - If all integer values are given, efficiently
solvable by interior point method.
38TROY Strategy
Solved by finding minimum Hamiltonian path
Integer nonlinear programming
Wire sizing/spacing
Second order conic programming
- A class of convex programming researched from
late 90s - Guaranteed to find the global optimum
- As efficiently solvable by primal-dual interior
point algorithm as LP
39Yield Improvement
- Monte-Carlo simulation with 10K defects
- On average 18 reduction in yield loss, up to
30 - 2.2 improvement loss with discrete wire width
40Outline
- Introduction
- Introduction to DFM and VLSI Routing
- Manufacturability-driven routing
- BoxRouter for global routing framework
- Wire density-driven global routing for CMP
variation and timing - TROY Track routing with yield-driven wire
planning - Litho-driven detailed routing
- Conclusion
41Litho-driven Detailed Routing
Our simulator
PROLITH
- Future work to round off manufacturability-driven
routing - Fast litho-simulator to guide detailed router
- Validated with PROLITH (orders of 106 faster)
- Should consider OPC into account
- Never considered in routing research so far
42Conclusion
- State-of-the-art router, BoxRouter is developed.
- Three key contributors to VLSI Manufacturability
are addressed in routing. - Global routing for CMP variation optimization
along with timing improvement - Track routing for critical area minimization
- Detailed routing for lithography enhancement
(future work)
43Publications
- 2005
- Minsik Cho, Suhail Ahmed and David Z. Pan ,
"TACO Temperature Aware Clock-tree
Optimization", Proc. ACM/IEEE Int'l Conference
on Computer-Aided Design (ICCAD), Nov, 2005 - 2006
- Minsik Cho, Hongjoong Shin and David Z. Pan ,
"Fast Susbstrate Noise-Aware Floorplanning with
Preference Directed Graph for Mixed-Signal
SOCs", Proc. Asian and South Pacific Design
Automation Conference (ASPDAC), January, 2006
(Nominated for Best Paper Award) - Minsik Cho and David Z. Pan , "PEAKASO
Peak-Temperature Aware Scan-Vector
Optimization", Proc. IEEE VLSI Test Symposium
(VTS), April, 2006 - Minsik Cho and David Z. Pan , "BoxRouter A New
Global Router Based on Box Expansion and
Progressive ILP", Proc. Design Automation
Conference (DAC), July, 2006 (Nominated for Best
Paper Award) - Minsik Cho, Hua Xiang, Ruchir Puri and David Z.
Pan , "Wire Density Driven Global Routing for
CMP Variation and Timing", Proc. ACM/IEEE Int'l
Conference on Computer-Aided Design (ICCAD),
November, 2006 - Minsik Cho, Hongjoong Shin and David Z. Pan ,
"Fast Susbstrate Noise-Aware Floorplanning with
Preference Directed Graph for Mixed-Signal
SOCs", IEEE Transactions on VLSI (TVLSI), under
review - Minsik Cho and David Z. Pan , "BoxRouter A New
Global Router Based on Box Expansion and
Progressive ILP", IEEE Transactions on
Computer-Aided Design of Integrated Circuits and
Systems (TCAD), accepted for publication - 2007
- Minsik Cho, Hua Xiang, Ruchir Puri and David Z.
Pan , "TROY Track Router with Yield-driven Wire
Planning", Proc. Design Automation Conference
(DAC), June, 2007 - Minsik Cho, Katrina Lu, Kun Yuan and David Z. Pan
, Architecture and Implementation of a Robust
Global Router for Ultimate Routability", Proc.
ACM/IEEE Int'l Conference on Computer-Aided
Design (ICCAD), submitted for review
44BACK UP SLIDES
45Sub-wavelength Lithography
Courtesy Intel, 2006
- Minimum feature size is scaling down faster than
wavelength. - Nanolithography (e.g., Nanoimprint, E-BEAM, EUVL)
are not mature/economical for mass production
yet. - It looks like optical lithography with help of
immersion and RET will be a work horse for next
510 years.
46Optical Lithography System
47RET Phase Shift Mask
One wide feature will be printed
Two features will be printed as designed
48Routing in VLSI Physical Design
a
c
a
Placement
b
Routing
Routing
a
d
d
b
a
c
c
49Non-Trivial Synergistic Optimization Redundant
Via and Litho
- Contact Redundancy does
- not come for free
- Requires Synergistic
- Optimizations
L.Liebmann, SPIE06
50Rule-based Approach
End of line spacing rules
- Conventional approach to separate design from
manufacturing RULES - Rules are starting running out of steam from 65nm
- Exploding number of rules
- VERY complicated rules (Rule book for 90nm has
600 pages) - Not accurate any more
51Chemical-mechanical Polishing (CMP)
Courtesy TSMC
- CMP (Chemical-Mechanical Polishing)
- Key multilevel metallization technique for copper
interconnect
52CMP (Topography) Variation
Copper Material
High metal density
Low metal density
- CMP Variation
- Performance degradation due to increase
resistance - Depth of focus issues due to non-uniform surface
- Systematic variation due to non-uniform metal
density distribution - Dummy fill increases mask volume and coupling
capacitance.
53Random Defect and Critical Area
C
A
B
54Motivation for Box Expansion
c
a
c
b
d
a
b
d
- Different resource management for wires
in/outside of box - Box Expansion pushes/diffuses congestion outwards
progressively
55Negotiation-based ReRouting
vC
vD
x
x
- Negotiation-based approach
- Enhance routing path by rerouting wires
- The congestion history is maintained to avoid
previously congested regions. - Multiple iterations
56Via-aware Layer Assignment
min
s.t
57Via-aware Layer Assignment
for every steiner point s of each net i
min
s.t
assigned only once
compute layer
get top/bottom layer of steiner point
pins on M1
routing cap. constraint
- The proposed ILP formulation finds layer
assignment with minimum number of vias.
58Via/Blockage-aware Layer Assignment
min
- Unassigned wires will be picked up by simple maze
routing which only shuttles between layers.
59Wire Density and Timing
- Coupling capacitance becomes dominant factor in
timing. - 60 of total capacitance below 180nm node
- Rapidly increase due to higher aspect (W/L) ratio
- Less coupling capacitance
- With less number of wires within a given region
- Lower wire density will reduce coupling
capacitance. - In the first order, coupling capacitance is
linearly proportional to wire density.
60Wire Density Driven Global Routing Overview
- Modify the cost function of maze router in
ReRouting - Add wire density term as a cost factor
- Keep congestion term
- How to smartly guide the maze router?
- Reduce the wire density around timing critical
nets - Less coupling capacitance
- Timing sensitivity map
Timing
- Reduce the wire density of selected dense regions
- Predictive CMP model
CMP
61Timing Sensitivity Map
S
R1R2
C
D
T To (R1R2 ) C
- Lower downstream capacitance improves timing
- Higher wire density near the driver is better
than higher wire density near the critical sink
62Recap Proposed Global Routing
Timing
Timing Sensitivity Map
Discourage maze router from passing a sensitive
or dense G-cell
CMP
Predictive CMP Model
63Timing Improvement
- On average 7 reduction
- Up to 10 reduction
64Timing Sensitivity Map
- Each global routing cell has different impact on
timing - Whether it is closer to the timing critical sink
- Whether there are more timing critical nets inside
65Notations
- Wire spacing
- Adjacent wires
- Overlapped wirelength
- Adjacent and overlapped wirelength
- Preferred location for minimum detailed wirelength
66Second Order Conic Programming
- A class of convex programming
- Minimizing a linear objective over the
intersection of an affine linear manifold with
the Cartesian product of second-order cones - Guaranteed to find the global optimum
- Efficiently solvable by primal-dual interior
point algorithm - Constant Hessian
- Self-duality of quadratic cones
- Nesterov-Todd (1997) direction
- Mehrotra type predictor-corrector
Geometrical illustration of conic constraint in 3D
67TROY Strategy
Integer nonlinear programming
Wire sizing/spacing
Second order conic programming
68Wire Ordering
Initial solution from interval packing
Solved by finding minimum Hamiltonian path
69Wire Spacing/Sizing
- When wire order is fixed
- Two auxiliary variables are introduced
- Two rotated quadratic conic constraints are added
- SOCP is solved for one entire layer
- Provides globally optimal wire width and spacing