Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Variability and Reliability Aware Design for Emerging System-on-Chips
2015. 09. 01
Yongchan Ban, Ph.D.
Principal Engineer, [email protected]
System IC R&D, LG Electronics
0
제16회 한국테스트학회 Tutorial
Outline
SoC and Embedded Systems
Technology Scaling: Variability, Reliability, Energy
Variability-aware Design
Reliability-aware Design
Power-aware Design
1
System-on Chips
2
Today’s IT (Focus on Person)
Real-time Model of the World
Global Intelligence
Future Technology
Cyber-physical System
The Internet of Things
Research Agenda
Tomorrow’s IT (Focus on Objects/World)
Public/Social Issues
Applications
Research Devices (System on Chips)
[ slide source: Prof. Jaeyong Jung ]
Technology Scaling in Every 2 Years
32 nm
2009
90 nm
2003
45 nm
2007
65 nm
2005
22 nm
2011
[ picture source: Intel ]
14 nm
2013
3
1st Gen FinFET 2nd Gen FinFET HKMG/Gate Last 2nd Gen HKMG Strain/Stress
To Go Out ~ 10 Years
4
[ picture source: Intel ]
30 Years of Scaling
Rice – single grain
Intel Atom™ - dual-core
[ source: Intel, Synopsys Forum’11 ] 5
Goals of Technology Scaling
Design new devices to be:
– Faster?
– Smaller?
– Lower power?
– Add new features?
Growth of semiconductor industry has been fueled by
the “ever cheaper” transistor
Two major issues confront further scaling
– Energy consumption!!
– Variability and Reliability!!
6
Component of Variations
Total Variation
– Systematic Variation
– Random Variation
Variation Sources
– Due to lithography, mask alignment,
etching, CMP, deposition, implant,
doping profile, temperature, etc.
Systematic Variation
– Deterministic
– Depending on layout density,
orientation, and/or location
Random variation
– Due to random process variations
– Due to fundamental randomness
Lithography Etch Loading
CMP Dishing & Erosion
Random Dopant LER Grain Boundary
7
Lithography Systematic Variation
8
Lithography Lithography
[ Source: TI ]
Line-end shortening ∆CD Proximity Poly Corner Line-edge Roughness Active Corner
Systematic Variation-aware Design: PO/OD
Line-end
Shortening
Active
Corner Rounding
Poly Corner
Rounding
Process Variation
9
Impact of Gate Length Variation
ΔLgate is up to 10% @45nm node.
The small improvement of ΔLgate reduction can leads to significant
decrease of delay and leakage variations.
-10 -8 -6 -4 -2 0 2 4 6 8 10-30
-20
-10
0
10
20
30
NMOS
PMOS
De
lay V
aria
tio
n (
%)
Length Variation (%)
-10 -8 -6 -4 -2 0 2 4 6 8 10
10-1
100
101
102
NMOS Leakage
Leakage V
ariation (
%)
Length Variation (%)
Over 20%
40X
[Ban et al., ISPD 2010] 10
Total Sensitivity based Optimization
11
Non-Rectangular Gate-aware Cell Charac.
PMOSNMOS PMOSNMOS
PM
OS
NM
OS
PM
OS
NM
OS
[ Source: TI ]
Total Sensitivity = Circuit Criticality + Manufacturability [Ban et al., ISPD 2010]
Systematic Variation-aware Design: S/D Cnt
The S/D contact resistance is as much as 26% of the gate channel
resistance at 45nm node devices.
12
[ITRS 2007]
n+
Gate Source Drain
n+
Rco
Rchannel
Rco
contactchannel
dSat
ddon RR
I
VR 2
13
Impact of Contact Variation
S/D Contact Size (CD)
– ΔS/D contact area Δcontact resistance (Rco) Δcurrent (Ids).
– 32nm standard cell (nominal CD = 40nm).
– 10nm Contact ΔCD up to 5% ΔIds & over 100% ΔRco.
30 40 50 60-10.0
-7.5
-5.0
-2.5
0.0
2.5
5.0
I ds C
urr
ent
Va
ria
tio
n (
%)
S/D Contct CD (nm)
30 40 50 60-80
-40
0
40
80
120
160
200
Rco R
esis
tan
ce
Va
ria
tio
n (
%)
S/D Contct CD (nm)
PO
LY
PO
LY
[Ban et al., DAC 2010]
14
Impact of Contact Distance on Ids
S/D Contact Position
– The Ids degrades as contacts are placed closer to the gate.
– Neighboring contact holes locally relax the strain in channel.
– Generate Look-up table for Distance weighting factor (wD)
PO
LY
50 100 150
-4
-3
-2
-1
0
1
2
3
I ds C
urr
ent
Va
ria
tio
n (
%)
Gate to Contact Space (nm)
[Eiji Morifuji, IEEE TED’09]
[Ban et al., DAC 2010]
15
Impact of Contact Shape on Ids
S/D Contact Shape
– As the contact length is larger, the Ids is increased.
– Less current crowding from the S/D electric field.
– Generate Look-up table for Shape weighting factor (wS)
40 60 80-1.0
-0.5
0.0
0.5
1.0
I ds C
urr
ent V
aria
tion
(%
)
Contact height (nm)
ACTIVE PO
LY
ACTIVE PO
LY
[Ban et al., DAC 2010]
16
Equivalent S/D Resistance Model
There are two look-up tables for a new contact model
– Distance weighting factor (wD) due to stress
– Shape weighting factor (wS) due to the current crowding
– Given ith slice of a contact,
the resistance is
where ρ is resistivity and A is
the area of a slice
– The driving current is
– The lower ω, the higher R and the smaller Ids.
ii,Si,D
iAww
R
i
iiSiD
i icodsch
ds
tot
ds AwwRRRR
V
R
V,,ds
111I
Dw
POLY SwContact
[Ban et al., DAC 2010]
Compact S/D Resistance Model
17
Contact Area Contact Position Contact Shape
0 50 100 150 200
-4
-3
-2
-1
0
1
2
3
TCAD
Proposed Model
Conventional
I ds C
urr
en
t V
ari
atio
n (
%)
Gate to Contact Space (nm)
20 40 60 80-1.0
-0.5
0.0
0.5
1.0
TCAD
Proposed Model
Conventional
I ds C
urr
en
t V
ari
atio
n (
%)
Contact Length along Gate (nm)
20 30 40 50-20
-15
-10
-5
0
5
TCAD
Proposed Model
Conventional
I ds C
urr
ent V
ari
atio
n (
%)
Contct CD (nm)
[Ban et al., DAC 2010]
S/D Contact Layout Optimization
Variability-driven Design – The main goal is to minimize the contact CD variation between the fastest
and slowest process corners.
Performance-driven Design – The contact CD increases as the pitch decreases.
– We can make vertically long contacts by reducing the space between contacts.
18
POLY METAL
ACTIV
E
POLY METAL
ACTIV
E
POLY METAL
ACTIV
E
Conventional Variability-driven Performance-driven
[Ban et al., DAC 2010]
Line-Edge Roughness (LER)
What is LER (Line Edge Roughness)?
– Random variation of MOS gate length along the gate width.
– LER is formed in the acid generation, the acid diffusion and development process.
Why LER becomes so critical
– LER does not scale accordingly and becomes an increasingly larger fraction.
– Below 30 nm the LER takes over and becomes the dominant fluctuation source.
19
Gate Length
[Chandhok, SPIE’07]
Exposure Development
[Namatsu, JVST’98]
[Asenov, IEEE TED’03]
Random Variation-aware Design on Std. Cell
Line-edge Roughness and Contact-edge Roughness
Gate Length
5.0222
2
21
2)(
c
cLER
Lf
LfS
Power Spectral Density (PSD) LER Modeling
Line-edge Roughness (LER)
[Ban et al., SPIE 2010]
[Ban et al., JETCAS 2011] [Ban et al., DAC 2011] [Ban et al., JM3 2010]
• Distance weighting factor (wD) due to stress
• Shape weighting factor (wS) due to the current crowding
Contact-edge Roughness Modeling
ii,Si,D
iAww
R
Contact-edge Roughness (CER)
20
LER-aware Poly Pitch Optimization
21
60 80 100 120 140 1603
4
5
6
7
8
9
3 L
ER
[n
m]
Poly-to-Poly Space [nm]
design
violation
process
violation
Data
Polynomial Fit
Linear Fit
5.0222
2
21
)(2)(
c
c
Lf
LxfS
3
22
11
0
)(
,...,0,)(
x
x
x
NixaxN
i
i
i
where,
or,
if
if
if
𝑥 ≤ 𝑥𝑝1
𝑥 ≥ 𝑥𝑝1
𝑥 ≥ 𝑥𝑝2
𝑃 𝑥𝑝1 𝐷 𝑥𝑝2 𝑥 > 𝑃, where 𝑃 is process violation
𝑥 < 𝐷, where 𝐷 is design violation
[Ban et al., DAC 2011]
LER-aware Poly Optimization
22
• N: a positive integer
• dj: distances from the edge
• Ψ: device criticality
[Ban et al., DAC 2011]
Cu erosion and dishing change R & C variation timing/power
Topographic variation translates to focus variation for imaging of
subsequent layers depth-of focus manufacturability
CMP impacts both IC parametric and manufacturability
CMP Induced Process Variation
[figures: Prof. Boning,MIT] Cupper Interconnect Problem
Multilevel Process
[Ban et al., KCS 2015] 23
CMP variation is highly dependent upon metal density
and pitch (line width & line space).
Dummy metal fill (conventional)
Pattern Dependent Effect
[figure: Prof. Boning,MIT]
Active Interconnect
Dummy Metal
[Ban et al., KCS 2015] 24
But, The CMP Variation is Still Huge
The dummy metal fill approach just cares for metal
layout density.
– Layout density variation is still huge.
– Layout density is not an only parameter controlling CMP.
– CMP variation mismatch and impact (e.g., 10%↑ ΔTcu @28nm)
28nm M4 Layout Density 28nm M4 Cu Thickness due to CMP
[Ban et al., KCS 2015] 25
Systematic Variation-aware Design on IP/Chip
LDE (Layout Dependent Effect)
Length of Diffusion
Well Proximity Effect Metal Boundary Effect Dual Stress Liner
[Ban et al., SPIE 2014] 22% Delay and 31% Leakage Spreading in 16nm Library
26
Variability Analysis Summary
LEA variability analysis was used to:
– Quantifies the delay and
leakage variability spread
– Verify the characterization
context positioning
– Help improve design
methodology
MUX2D0
Cell Under Test
F
I
L
L
2
F
I
L
L
2
I
N
V
D
0
SDF0
Latch
SDF0
MUX2D0
Cell Under Test
TAP
CELL INVD2
AOI22
MUX2D8 DFFD2
Fast Context Slow Context
Up to 21ps or 22% delay spread Up to 111pW or 31% leakage spread
Characterized
delay
Characterized
leakage
27 [Ban et al., SPIE 2014]
Variability-aware Implementation
Reducing Variability of Cells/IPs
– Quantify LDEs related variability in Std. Cells and Ips
– Pin point root causes of excess variability and optimize cells
Variability-Aware Placement
– Context library litho/stress simulation and analysis with LEA
– A variability score for cell combinations (CCI)
– Use cells with low variability score
Timing Margin Tuning
28 [Cadence, CDNLive’09]
Layout Guidelines to Mitigate LDEs
Avoid functional cells on boundaries
– Functional cells should be avoided on the block periphery
– Block periphery should be padded with the non-functional cells (e.g., fillers, decaps,
endcaps, and so on.).
Fix context of clock cells
– Surround all instances of clock cells with an identical context and characterize the
clock cells with the context.
– Minimizes impact of the un-modeled LDEs on the clock cells.
Surround double-height cells with non-functional single-height cells
– Double-height cells can cause high LDEs around them.
– Placing non-functional single-height cells around double-height cells reduces LDEs
on nearby functional cells.
Avoid very small drive strength cells on critical paths and clocks
– Small drive strength cells are more susceptible to LDEs.
[Freescale, SPIE 2012]
29
Routing: a key stage for printability optimization
– Last major design stage
– Wire embedding step
Lithography Aware Routing
– OPC Friendly Detailed Routing [Huang+ DAC’04]
– Multi-level Routing with OPC [Chen+ ASPDAC’05]
– RADAR: Litho-Aware Routing [Mitra+ DAC’05]
Layout Chip
Pass
Many Rerouting
ORC (check)
Systematic Variation-aware Design on Routing
Fail!!
30
Systematic Variation-aware Design on Routing
Lithography-friendly Detailed Routing
D D D
D=2 D=3
Fre
quency
Total EPE
OPC & Simulation
N-th set of Weak Pattern OPC & Simulation
D=1 D=2
Fre
quency
Total EPE
[Lin, Ban, Pan, and Li, ICCAD 2011] [Cho, Yuan, Ban, and Pan, TCAD 2009] [Cho, Yuan, Ban, and Pan, DAC 2008]
31
In-design DFM w/ Pattern Matching
[courtesy of Synopsys]
32
LGE In-Design DFM Approaches
(1) Conventional Litho. Check at the signoff stage
(2) In-design Pattern Matching from an early stage of routing
(3) Correct-by-construction Routing
Final GDS Litho. Check Hot Spots Routing Tool
Fixing DFM Clean
Routing DB Pattern
Matching Hot Spots Routing Tool
Fixing Litho. Check
Pattern Lib. Routing w/
Concur. Fix Litho. Check
Flow Runtime (sec) # Hot spot
remaining Litho. sim Pattern Match Fixing Overall
(1)Conv. DFM sign-off 57,120 487
(24 violations) 1903 59,510 0
(2)In-design DFM - 499
(49 violations) 2402 2,901 0
TAT reduction - 0.97 0.79 X 20.5
< 16FF CA57 CPU >
[Shin, Ban et al., SPIE 2014] 33
The Evolution of Pattern Matching
// Sample DRC Rules File AND poly diff gate NOT diff gate srcdrn WIDTH diff LT 5 OUT w5001 30 3 EXT[H] diff LT 5.5 OUT x5002 31 4 WIDTH poly LT 4.5 OUT w5003 32 EXT poly LT 4.0 OUT x5004 33 EXT poly srcdrn LT 2.0 OUT x5013 41 EXT[T] poly diff LT .001 OUT x5013 42 WIDTH srcdrn LT 5.5 OUT w5013 42 WIDTH gate LT 5.0 OUT w5016 45 WIDTH epi LT 7.0 OUT w5020 52 4 ENC[T] diff poly LT 4.0 OUT e5301 57 7 ; PREPARE FOR NODAL SPACING CHECK ; FLATTEN POLY LAYER FOR CONNECTION FLATTEN poly fpol ; FLATTEN SRCDRN LAYER ALSO FLATTEN srcdrn fsd ; FLATTEN SRCDRN LAYER ALSO FLATTEN epi fepi }
Classic DRC Code
Classic Pattern Matching
is a specific physical
representation
New representation is
Topological “Squish” Pattern
<rule>
<rulename>Hook.with.notch.M2</rulename>
<report merged_overlay_layer="15.102"/>
<pattern_def x_deltas="1 50 50-150 50 1"
y_deltas="1 0-100 50-100 0-500"
signature="L1
0 0 0 0 1/
0 0 1 0 1/
0 0 1 0 0/
1 0 1 0 0>
<pattern sel="15.0" />
<overlay_scanline_range>x1 x4 y1 y4</overlay_scanline_range>
</pattern_def>
</rule>
[Ban et al., SPIE 2014] 34
Topological “Squish” Pattern
Very simple Pattern Abstraction
More like DRM than DRC Rules
Easy to maintain
Natively supports all orientations (mirror, rotations)
Support multi-level patterns
High performance search engine
[Ban et al., SPIE 2014] 35
In-design DFM Optimization
36 [Source: Cadence]
Lithography Challenges in sub 45nm
k1: lithography difficulty NA: numerical aperture λ: wavelength of source HP: min. printable half pitch NA
kR
1
[S. Borkar, 2014]
37
Why Double Patterning?
Too hard with the current infrastructure!
– Increase NA (3rd generation fluid >1.4, larger lens)
– Use shorter wavelengths (13.5nm-EUV)
Have no choice but to bear with the current state, but
increase HP to print 32/22nm designs
– Double patterning!
– Reuse the current infrastructure
NAkR
1
Ex) λ = 193nm NA = 1.4 K1 = 0.26 R = 35.8 Around pitch 70nm ↓ is
a lithography limit.
Pitch
R
38
Double Patterning Tech (DPT)
Conventional LELE (litho-etch-litho-etch) DPT
stitch
stitch
(1) target (2) core mask (3) sidewall spacer (4) trim mask
Spacer Type SA (self-aligned) DPT
[Ban et al., DAC 2011] 39
Layout Decomposition vs. 2-coloring
Layout decomposition for DPT is more complex than
phase assignment or 2-coloring
A
B C
D
E
A A
B
C D
E
A B
C D
E
Uncolorable
A
B C
D
E
A
B C
D
E
Stitch
[Cho, Ban, ICCAD’08]
A B
C D E
40
LELE DPT Aware Routing
a) # of stitches b) wirelength, via, and so on
[Cho, Ban, ICCAD’08] 41
SADP Aware Routing
SADP Layout Decomposition
SADP-awra Routing
[C. Kodama+, ASPDAC’13]
[Ban et al., DAC 2011]
42
Reliability Problems in 3D ICs
Mainly due to CTE (coefficients of thermal expansion) mismatch
between Si and Cu TSV
[ Prof. S. Lim, Gatech ]
Better Performance
Massive Bandwidth
Reduced Interconnect Delays
Power Reduction (Less IO driver)
Higher Functionality/Space
Heterogeneous Integration
43
TSV-TSV & TSV-Transistor Interactions
[ Source: IMEC ]
44
Keep-Out Zone (KOZ)
KOZ tradeoff: area vs delay vs reliability
[ Prof. S. Lim, Gatech ]
45
TSV Stress Impact on Mobility
Hole Mobility Variation Electron Mobility Variation
Thermal Mechanical Stress
[ D.Pan, ASPDAC’12 ]
46
TSV Stress-aware Placement
[ K.Athikulwongse, ICCAD’10 ]
Hole Mobility Variation Electron Mobility Variation
Hole Critical Cell
Electron Critical Cell
Regula
r TSV
Irre
gula
r TSV
47
Reliability-aware Design (EM)
Signal Electro-migration (EM)
– Mean-to-time-failure (DC Current)
– Mean-to-time-failure (AC)
– Mean-to-time-failure (AC+ΔProcess)
– Joule Heating, Signal interconnects
– Prone to process variation (lithography, etch and CMP)
– Higher driving strength in FinFET and higher circuit frequency
void
hillock
/kT)(E
nACae
JJ
A=MTTF
/kT)(E
nprocess)AC(ae
JJJJ
AA=MTTF
[Ban et al., SPIE 2014]
Failures due to Electromigration
[Ban et al., KTC 2014] 48
/kT)(E
nae
J
A=MTTF
Signal-EM in 16nm Design
(a) When the impact of the
geometry is higher than
one of the current
density, MTTF
increases.
(b) While when the impact
of the current density is
higher than one of the
geometry change, MTTF
decreases.
/kT)(E
nprocess)AC(ae
JJJJ
AA=MTTF
[Ban et al., SPIE 2014] 49
Signal-EM with Input Voltage
Front-end: Typical process corner
Back-end: 110 ℃ and RCbest parasitic corner
The higher deriving current from FinFET causes more EM
violations.
1 6
13
31
54
94
0
20
40
60
80
100
0.72 V 0.8 V 0.88 V
# o
f EM
Vio
lati
on
s
Input Voltages
# of Signal-EM w/ Input Voltage
@Irms @Iavg
101% 113%
124%
189% 211%
232%
0%
50%
100%
150%
200%
250%
300%
0.72 V 0.8 V 0.88 V
Max
. EM
Vio
lati
on
s
Input Voltages
Max. Signal-EM w/ Input Voltage
@Irms @Iavg
[Ban et al., KTC 2014] 50
Signal-EM with Ambient Temperature
12 12 13 15
86 86 94 97
0
20
40
60
80
100
120
-40 ℃ 0 ℃ 85 ℃ 125 ℃
# o
f EM
Vio
lati
on
s
Ambient Temperature
# of Signal-EM w/ Ambient Temp.
@Irms @Iavg
123% 123% 124% 125%
232% 232% 232% 232%
0%
50%
100%
150%
200%
250%
300%
-40 ℃ 0 ℃ 85 ℃ 125 ℃
Max
. EM
Vio
lati
on
s
Ambient Temperature
Max. Signal-EM w/ Ambient Temp.
@Irms @Iavg
Front-end: FF (Fast-Fast) process corner
Back-end: 110 ℃ and RCbest parasitic corner
The more FinFET current is induced as temperature increases
due to temperature inversion.
[Ban et al., KTC 2014] 51
Signal-EM with Metal Temperature
1
13 13 13
94
1000
1
10
100
1000
85 ℃ 110 ℃ 125 ℃
# o
f EM
Vio
lati
on
s
Metal Temperature
# of Signal-EM w/ Metal Temp.
@Irms @Iavg
112% 124% 124% 124%
232%
647%
0%
100%
200%
300%
400%
500%
600%
700%
85 ℃ 110 ℃ 125 ℃
Max
. EM
Vio
lati
on
s
Metal Temperature
Max. Signal-EM w/ Metal Temp.
@Irms @Iavg
Front-end: Typical process corner
Back-end: 110 ℃ and RCbest parasitic corner
The signal-EM is so sensitive to the metal temperature.
110℃ is recommended for sub-20nm node design.
[Ban et al., KTC 2014] 52
Signal-EM with Metal RC Corners
6 12 11
4 4
77 86 86
77 70
0
20
40
60
80
100
Typical RCBest RCBest_D RCWorst RCWorst_D
# o
f EM
Vio
lati
on
s
Metal RC Corners
# of Signal-EM w/ Metal RC Corners
@Irms @Iavg
121% 124% 122% 120% 120%
230% 232% 232% 234% 234%
0%
50%
100%
150%
200%
250%
300%
Typical RCBest RCBest_D RCWorst RCWorst_D
Max
. EM
Vio
lati
on
s
Metal RC Corners
Max. Signal-EM w/ Metal RC Corners
@Irms @Iavg
Front-end: FF (Fast-Fast) process corner
RCBest corner gives the worst EM violation.
RC reduction on signal nets might induce more current.
The DPT corners minor changes EM violations.
**_D means DPT corner.
[Ban et al., KTC 2014] 53
Reliability-aware Design
[Ban et al., SPIE 2014]
16nm Cortex-A57 CPU
Redundant Via Metal Dummy Fill
1W
1S
2W
2S 2S
3W
Non-Default Rule
EM Fix Approaches [Ban et al., KTC 2014]
EM on Via
EM on Clock Net
54
Power Basics
The total power dissipated in a device consists of two components.
Static or leakage power is state, temperature, process and voltage
dependent.
Dynamic Power is comprised of internal power and switching power.
dynamicstatictotal PPP
leakstatic IVddP
switchingernaldynamic PPP int
clkeffdynamic fVCP 2
55
Overall Voltage drop is
In addition to IR drop, inductance for power ground network
also affects the voltage drop.
Dynamic Voltage Drop
dtLdiIRVdrop /
. . .
VDD
Time(ns)
IR-drop hot spot
[Voltage drop @ transistor level] [Hot-spot Area]
[Ban et al., ISOCC 2014] 56
Dynamic IR Drop w/ Process Corners
57 [Ban et al., ISOCC 2014]
RC corners # DVD
Max.
# DVD
Min.
Worst
Instance
Worst
Wire
RCBest_DPT 531 426 27.8% 27.0%
RCBest_dPV 712 433 29.1% 27.6%
RCBest 561 433 28.4% 27.6%
Typical 1537 513 32.9% 31.6%
RCWorst 4722 1199 40.5% 38.7%
RCWorst_dPV 2365 400 33.5% 30.5%
RCWorst_DPT 4543 1086 40.3% 38.5%
[DVD with Metal RC-corner]
Future Research Directions
Cyber-physical System
The Internet of Things
Reliable SoC/Embedded System
• Manufacturability
• Reliability
• Testability
• Energy Efficiency
Circuit/System-level Techniques for
New Processes/Devices
• Pi-gate, Wire, CNTs
• EUV, DSA, Nanowires
• Memristor-based Systems
3D SoC
• TSV (Through Silicon Via)
• Heterogeneous Integration
HW/SW Co-design
• HW Acceleration IPs
• System Software
VerificationSynthesis
Function Architecture
Hardware Software
Mapping
Re
fin
em
en
t Ab
stra
ctio
n
Abstract Co-design Process
New Devices based SoC
3D SoC HW/SW Co-design
Emerging Applications
58
Thank You!
제16회 한국테스트학회 Tutorial, LG전자, 반용찬 수석연구원