CADENCE CONFIDENTIAL 65nm Ombudsman Ted Vucurevich CTO Advanced Research and Development
CADENCE CONFIDENTIAL
65nm Ombudsman
Ted VucurevichCTOAdvanced Research and Development
CADENCE CONFIDENTIAL
IP based SOC Design:Complexity, Cost, and Risk
Digital IP
Product 2 Product 3Platform 1
Product 1 Product n
System, board, chip optimization
Software, hardware trade-offs
$50-70m@65nmCell libraries Software
Analog IP ProcessorsTesting
Packaging MemoryFoundry
CADENCE CONFIDENTIAL
Time to Market (Critical to success)
• Takes longer than expectedto complete >90nm designs, sometimes compromisingdesign objectives
• Time-to-market requirementspreclude the ability to do are-spin.
• Power minimization has become a key metric –especially for >90nm designs in the wireless communications, consumer, etc, segments.
• Manufacturing yields must be considered during design implementation at >90nm
Key Factors in Design (10 Being Most Important)
Feature Size 0.18µm 0.13µm 90nm
Complete On-Time - - 9Chip area 6 to 8 5 to 8 4 to 7 Performance 2 to 8 2 to 8 2 to 8 Power 3 to 8 5 to 8 6 to 8 Manufacturing yields 1 to 2 3 to 5 7 to 9 Ability to ensure 2 to 4 5 to 7 2 to 8
first-time success Source IBS
CADENCE CONFIDENTIAL
Less First Silicon Success and the Changing Rate of Failures
• Better– Functional Verification– Noise / SI– Clocking– IR Drops
• Worse– Analog Tuning– Mixed-Signal Interface– DFM (RET)
0%
4%
5%
10%
12%
14%
14%
16%
17%
18%
22%
28%
62%
3%
9%
13%
14%
8%
10%
23%
45%
0% 10% 20% 30% 40% 50% 60% 70%
RET
Other
Mixed-SignalInterface
Firmw are
Yield / Reliability
IR Drops
Analog Tuning
Slow Path
Fast Path
Clocking
Pow er Consumption
Noise / SI
Logic/Functional
2003
2001Collett International Research:
2000, 2002 Functional Verification Studies;2003 Design Closure Study, 01/04
• First silicon success rates declining
– First Silicon OK 48% in 2000 39% in 2002 34% in 2003
– Third Silicon OK>90% in 2000 >70% in 2002 >60% in 2003
Trends are Increasing
Tren
ds a
re D
ecre
asin
g
CADENCE CONFIDENTIAL
Design Hand-off Point Complexity
LibrariesRET Rules
Implementation
Sign-Off
IPFirmware App. LibsPlatform
Models
• Historic Design to Manufacturing Hand-Off– GDSII Sign-Off with
DRC/LVS
• Sub-Wavelength DFM Hand-Off– GDSII Sign-Off with
DRC/LVS
– Electrical Sign-Off– Signal and Physical
Integrity
– Design Robustness
– Yield
– Reliability
RTL
• RTL Handoff – Verilog / VHDL
– Vectors
• Platform– IP, SW
– SVP, IVP
System
Manufacturing
CADENCE CONFIDENTIAL
Addressing >65nm Design Challenges
Analysis
ModelingOptimization
Synthesis
What is Manufacturable?
Design Rule Complexity
01234
250 nm 180 nm 130 nm 90 nm 65 nm
Node
Normalized Pages
Increasing lithographic complexity is driving increasingly complex design rules.
Source: Mark Mason (TI)
Electrical ComplexityVariability: Process, Thermal
HotSpot
Runtime SW Dependent
Device
Wire
Source: “Models of process variations in device and
interconnect” by Duane Boning, MIT & Sani Nassif,
IBM ARL.
65nm Device Modeling challenges
DistantNeighbors
Near byNeighbors
• Identically drawn devices can be imaged quite differently due toeffect of neighbors
What we used to model
What we must model now
What we need to supportOne model: Many users
• One model for all transistor analysis
Parameter Extraction
Timing Simulation
Signal Integrity
IR Drop
Circuit Simulation
Static Timing Analysis
Electromigration
RF Simulation
Model
Device Modeling Standardization
The Compact Modeling Extensions to Verilog-A/MS
• Allows low level models to be defined in a standardized modeling language
– Device modeling
– Macromodeling
– Interconnect modeling
• Paramsets standardize SPICE .model files
– Vendor independent model files
– Directly supports modeling of variation
Mismatch Example
• Modeling variation across die vs. xy
rho11 rho01
paramset n180nm bsim6;...rho = rhostats.drho/sqrt($m∗$n) + (
(1 – $x)∗(1 – $y)∗rhostats.rho00 + $x∗(1 – $y)∗rhostats.rho10 +(1 – $x)∗$y∗rhostats.rho01 + $x∗$y∗rhostats.rho11
);...
endparamsetrho00
drho
rho10
y
x
CADENCE CONFIDENTIAL
Integrated Extraction will be necessary
• Extractors will produce instance based models
– Not just parameters
– Tailored models are more powerful, compact
Parameter Extraction Simulation/AnalysisModels
• Examples
– Interconnect extraction, variability extraction, macromodeling
Mismatch Example
• Extractor passes location and orientation for instantiation
n180nm #( ..., .$x(–2), .$y(4), .$angle(0)) M1 ( ... )n180nm #( ..., .$x(4), .$y(2), .$angle(-90)) M2 ( ... )n180nm #( ..., .$x(2), .$y(–4), .$angle(180)) M3 ( ... )n180nm #( ..., .$x(–4), .$y(–2), .$angle(90)) M4 ( ... )
M1 M2
M3M4
Interconnect Modeling:Process Modeling Complexity
• Resistance
– Non-rectangular wires
– Wire edge effects– Actual wire widths vary depending
on spacing to neighboring wires on each side
– Dishing, slotting
• Capacitance
– Sidewall >> Substrate
– Complex 3d structure
Source: David Overhauser
Net to Extract
Optical Modeling ComplexityRET Corrections not complete
No RET
Lars Liebmann of IBM
RET
POE Modeling:Kernel based techniques
• Example:– φ : 7→ 7
– Shown, projectionsin 3
M1 M2
M4M3
L1L2
L3 L4
L5 L6
V+outV-
out
V-in
V+in
Vbias
• Set of configuration parameters:{W1, L1, Ibias, W3, L3, W/L|mirror, Cload}
• Set of performance parameters:{gain, noise, power, HD2, HD3, Q, Cload}
• Mapping from configuration to performance through simulation
Source: Fernando De Bernardinis
Modeling Variability:Machine Learning
Cross-section view
Aggressor
Net to Extract
Source: David Overhauser
Capacitance models (27 independent parameters)Models exist for 3D structuresX architecture SUPPORT
ML and Experimental Design:NX Architecture and Capacitance Application
Seed +Experimentaldesign
Fieldsolver
Modelgeneration
Training data /test data
Profileconfigurations
Prob. distributionon models
model #
prob
Modelapplication
Profileconfigurations
PredictedcapacitancesProfile
decompositionChip
NX-C EngineExtractiontech. file
RCGen
CADENCE CONFIDENTIAL
Addressing >65nm Design Challenges
Analysis
ModelingOptimization
Synthesis
CADENCE CONFIDENTIAL
Electrical Verification: Silicon predictability at nanometer geometries
IR-Drop Analysis
case3
case4
case5
clock driver
BUF4
BUF4 x 4
25 points
tpd
500 um
200
um
Delay Calculation
0 200 400 600 800Victim delay pushout (ps)
0
200
400
600
800
Rece
iver
outp
ut dela
y push
out (p
s) Consistent pessimism reduction over hundreds of nets
More pessimism reduction for noisier nets
Crosstalk Analysis
Timing Engine
Statistical Timing Analysis
STA
N worst Case Paths
5 interconnectcorners
Propagation
Symbolic
Monte Carlo
Slow, Nom, FastLibraries
PDF’sCDF’s
Source: Sreeja Raj
Robust Clock Design:Improving design margin
Q DD2
D1D0
D3
-11,+16
-12, +14-8, +16-9, +12-14,+18-10, +20
Variance
3525353Metal5
22165297 Total
5962471Metal45953089Metal35807360Metal2826299Metal1
LengthLayer
Fatal Timing Problems when process variation induces:
Hold Violation: Fast D1, Fast D2, Slow D3Setup Violation: Slow D1, Slow D2, Fast D3
Addressing >65nm Design Challenges
Analysis
ModelingOptimization
Synthesis
Manufacturing Robustness: Via Optimization
Before Optimization After Optimization
Via Optimization ResultsVia Optimization ResultsLayer Before AfterVIA1 67.2% 89.1%VIA2 64.7% 89.3%VIA3 70.3% 96.2%VIA4 73.9% 98.6%Total 67.3% 91.7%
Power Via Optimization ResultsPower Via Optimization Results6,393 Areas for improvement78,262 vias missing and added
CADENCE CONFIDENTIAL
Functional Yield Improvement:Design Robustness (variability analysis)
Statistical, Electrically AwareDefect Density
Driven Optimization
Yield = (1 + λ/α)- α
defect densitycritical areafailure rate
λ = ?x()[CA(x)][D(x)]dxA defect causing a short
Yield = (1 + λ/α)- α
defect densitycritical areafailure rate
λ = ?x()[CA(x)][D(x)]dx
defect densitycritical areafailure rate
λ = ?x()[CA(x)][D(x)]dxA defect causing a shortA defect causing a short
Yield Improvement for ARM 9 core:Wire Spreading and Tuning Example
Significant Critical Area Reduction
Incremental Extraction and Optimization:Wire Spreading
GeometryEngine
TimingCore RouterExtractor
DelayCalc.
SI
OA-UDM
NetAreaChip
EstimatedMSTGeometry
3D corridor
Reduced network
Timing annotations
ARM Core Spreading DetailBefore Wire Spreading
Example segment at min space of 0.14um
Negative slack for path
After Wire Spreading
Example segment spread to 0.7um
Total Cap for net decreased by 6%
Coupling Cap for net decreased by 9%
Positive slack for path (increase of 44ps)
ARM Core Example: Impact on Parasitics
-11.8%
Less Coupling-Cap. => Total Cap. reduced by ~ 12%
Addressing >65nm Design Challenges
Analysis
ModelingOptimization
Synthesis
Serving the opportunityThe “Everything Closure” problem
• Closure issues extend across multiple design dimensions as well as throughout the design process
– Performance: Area, Speed, Power, Thermal
– Manufacturability: Yield, Reliability
– Test: Cost, Diagnostics
• Must re-architect the entire implementation tool chain to address the resulting consequences.
“Everything Closure”
Mask Handoff
≥ 0.25µ 0.18µ - 0.13µ ≤ 90nm
Yield Optimization
Design
Routing
Finishing
PhysicalVerification
Planning
Placement
SynthesisPBS
RET/OPC
ElectricalSign-Off
PhysicalSynthesis
ChipOptimization
Manf.Optimization
Logic (gates)
Placement
Routing (wires)
Manufacture
Next Generation Core ArchitectureControl and Data “Busses”
DriversComponent Plug-in Component Software Integration mechanism
Control and coordination of the EnginesM
odel
ers
Ana
lyze
rs
Synt
hesi
zers
Verif
iers
Transaction layerOpen Access Unified Persistent Data Model
Thread Safe Run-time Data transactions
Re-entrantHeterogeneous IncrementalHierarchical
Topologically driven, Shape aware Interconnect Synthesis
• Constraint Driven
• Incremental
• Re-entrant
• Threadsafe
• Gridless w/ optional GriddedMode
• Shape-based
Topology &Global Routing
Detail Routing
Corridor Routing
Routing Optimization & Refinement
Manufacturing Aware Interconnect: Model-based topology criteria
Preferred Routing rules such as segment widths, spacings can be further refined. “Printing Rules such as min-jog length and width can also be enforced
Very little rip-up-&-re-route is performed to complete the routing and route optimization – resulting in layout that finishes as the designer predicted in the previous phases
Diagonal “X” interconnectManufacturability Improvement
“X Architecture”Total Wire Length= 128.6Total vias= 17
Traditional:Total Wire length= 171.9Total vias= 62
65 Nanometer Delay TestChallenge & requirements
Resistive bridge defects
Signal required
Delaydefect zone
Proper operation
Logic failure
Transition fa
ilure
Logi
c 1
Logi
c 0
Based on: “Defect-Based Test: A Key Enabler for Successful Migration to Structural Test”, Sanjay Senguptaet al, Intel Technology Journal, 1st quarter 1999
2.5V
2.0V
1.5V Delay Test requires
• High transition fault coverage
• Compact test patterns
• Post-wire timing (True-Time)
1.0V
0.5V
20ns0.0V
2ns 2.2ns 2.4ns 2.6ns 2.8nsStuck-at test capture
Design For Test (DFT) Delay test, compression, & diagnostics
Delay test Diagnostics
• Meet manufacturing cost requirements• Minimum design impact• Diagnostics compatibility
Compression
MIS
R w
/ X m
aski
ng1 0 0 0 1 1 0 1 0 1 1 1 0 1 0 0 0 1
1 0 0 1 1 0 1 0 1 1 1 0 1 0 1 0 1 0
0 0 1 0 1 1 0 1 0 1 1 1 0 1 1 1 1 1
0 0 1 0 1 0 0 1 0 0 1 1 0 1 0 1 1 1
1 1 0 0 1 1 0 1 0 1 1 1 0 1 0 0 0 1
1 0 1 1 1 0 0 1 0 1 1 1 0 1 1 1 0 1
• High transition fault coverage• Compact test patterns• Post-wire timing
• High Callout Defect correlation • Logic and scan chain diagnostics• Diagnostic test generation
Logic 0
Logic 1
Delay defect zone
Successful Nanometer Design
Modeling
Analysis
Synthesis
Optimization