Impact of Statistical Variability and Charge Trapping on 14 nm SOI ...
Impact of Statistical Variability and Charge Trapping on 14 nm SOI FinFET SRAM Cell Stability ! X. Wang1, B. Cheng1, A.R. Brown2, C. Millar2, J. B. Kuang3, S. Nassif3, A. Asenov1,2! !
1 Device Modelling Group,! University of Glasgow, UK
! 2 Gold Standard Simulations Ltd, UK
3 IBM Research – Austin, USA
! 1! ESSDERC, 16-20 September 2013, Bucharest Romania! !
Outline! q Introduction! q 14 nm node DG SOI FinFETs! q Simulation of Random Charge Trapping and
Statistical Variability Sources! q Compact Modelling Methodology! q Charge Trapping Impact on SRAM SNM! q Charge Trapping Impact on SRAM WNM! q Summary!
2!
Introduction • Why this study?! • Novel 3-D architecture FinFET will be widely adopted at 14 nm technology, with reduced variability on SOI substrate due to tolerance to low channel doping.!
• However, (1) statistical aspect of reliability due to random individual trapping becomes an increasingly important issue. (2) In addition, charge trapping impact is affected by statistical variability sources.!
• Accurately modeling reliability of nanoscale transistors in circuit level should take care of above mentioned properties, therefore requires a “statistical” method, rather than describing average reliability behavior.!
• SRAM stability is susceptible to variability, therefore statistical study is needed.! 3!
FinFET
IMEC! Intel! 22nm! !
Veloso et. IEDM, 2009!
Fin Edge Roughness:! width, height, slope!
TSMC! Chang et., IEDM, 2009!
Bulk substrate!
IBM! Chang et., VLSI tech., 2011!
SOI substrate! 4!
Simulation Design of 14nm SOI FinFETs (GU-IBM collaboration) TE GA
MC calibrated @ 85°C!
W
fin
SOURCE
Double-Gate ! SOI FinFET!
LG
to
x
! E
XID DO
IE UR
B
TE RA
T BS
SU
Monte Carlo!
-1
2.5 2.0 source
7
fin
1.5
Velocity [x10 cm s ]
H
DRAIN
HM
Calibrated DD
drain
20!
EOT (nm)!
0.8!
WF (nm)!
10!
HF (nm)!
25!
NSD (cm-3 )!
3.0E20!
NCH (cm-3 )!
1.0E15!
VDD (V)!
0.9!
IOFF (nA/μm) !
10!
IDSAT (mA/μm) ! 0.9/0.8!
1.0
DIBL (mV/V)!
Default DD
0.5 0.0
Lg (nm)!
VG 0.420
30 40 Position [nm]
-0.8!
56/65!
Ref.: ITRS 2010 update!
50
5!
Intrinsic Parameter Fluctuations Statistical Variability Sources
TiN!
potential!
Random dopants!
Polysilicon/Metal Gate! Granularity!
Line Edge Roughness! 6!
Statistical variability simulation GER: ΔLG , ΔSCE! FER: ΔWFIN , Δconfinement! ΔRSD! RDD: ΔRSD , ΔNA! MGG: ΔΦM , Δψsurf !
• Each variability source has different impact on the device parameters and performance.! Wang, et al, IEDM 2011, pp103-106! 7!
Interaction: Charge trapping vs Statistical variability sources Sensitive regions!
• FER: local shortenings!
• MGG: metal grains with high currents underneath!
• RDD: current percolation paths! Wang et al, SISPAD 2012, pp.296-299! 8!
Vt RTS Distribution and Reliability are affected by Statistical Variability
1-CDF
1
Single Trapping!
0.1
Single Trapping 0.01 Uniform Device ’Atomistic’ Devices
0.001 0
Uniform device!
2
4
6
6VT (mV)
8
10
Wang et al., SNW 2012, pp.77-78!
Atomistic device!
• In the presence of SV, the RTS distribution tail is increased! Multi-trapping!
RTS: random telegraph signal!
9!
Random charge trapping effect on VT
Normal Quantile
4
2
-2
Trapping Density (cm ) 0 1E11 5E11 1E12
0
-2
-4 0.1
0.15
0.2
0.25
VT (V)
0.3
0.35
0.4
Poisson distribution ! of trapping charge ! number is assumed!
• First, the average VT shift increases with degradation heuristically;! • Most important, the statistical variability increases with degradation.! 10!
Statistical Compact Modelling Method • A small set of BSIM-CMG compact model parameters is used to extract statistical samples at fresh stage, also applied to degradation.! • In circuits random fresh samples are assigned, responding stressed samples are put for stressed transistors.! • Assume trapping effect is dynamically recoverable.! • e.g., M2 is biased with high VG and low VD, subject to PBTI!
retention!
PU! PG!
PG!
PD!
6-Transistor SRAM cell! PU: pull-up transistor, p-FinFET;! PG: pass-gate transistor, n-FinFET;! PD: pull-down transistor, n-FinFET;! 11!
What happens to SRAM SNM after stress? A
VR (V)
0.8
SNM(A)
0.6 0.4 0.2 0 0
SNM(B)
Fresh Stress (state A) Stress (state B)
0.2
0.4
B 0.6
VL (V)
0.8
SNM: static noise margin, the SRAM stability for read mode! State A: left 0, right 1; State B: left 1, right 0!
• Generally, stress induced trapping leads to less static noise margin ! • Heavier N/PBTI, more threshold shift, less stability! ! 12!
SNM Distribution
Two types of SRAM cells! with fin-number ratio! of PU:PG:PD,! 111 SRAM and 112 SRAM! are examined!
• First of all, the distribution is non-Gaussian.! • Compared with 111-fin SRAM cells, 112-fin cells increase SNM.! • With charge trapping induced degradations, the SNM is reduced.! 13!
Charge trapping effects on SNM 20
SNM (mV)
250
200
150
18
mSNM (mV)
SNM(A), 1:1:1 SNM, 1:1:1 SNM(A), 1:1:2 SNM, 1:1:2
16
SNM(A), 1:1:1 SNM, 1:1:1 SNM(A), 1:1:2 SNM, 1:1:2
14 12 10 8
100 0
2e+11
4e+11
6e+11
8e+11 -2
Trapping Density (cm )
1e+12
6 0
2e+11
4e+11
6e+11
8e+11 -2
1e+12
Trapping Density (cm )
• The average SNM is reduced by up to 30 mV, with charge trapping induced degradations.! • The statistical variation of SNM increases by 30-40% with degradation.! • 112-fin SRAM cells show better stability. Compared with 111-fin cells, 112-fin cells increases SNM by ~45% in average.! 14!
What happens to SRAM WNM after stress? 0.8
VL (V)
0.6 Fresh Stressed
0.4
0.2 WNM
0 0
!
0.2
0.4
0.6
0.8
VR (V) WNM: write noise margin, SRAM stability for write mode!
• In contrary to SNM, WNM increases a bit due to charge trapping.! • The WNM distribution is non-Gaussian.! 15!
Charge trapping effects on WNM
WNM (mV)
400
16
WNM (state A), 1:1:1 WNM, 1:1:1 WNM (state A), 1:1:2 WNM, 1:1:2
mWNM (mV)
420
380 360 340 320 0
2e+11
4e+11
6e+11
8e+11 -2
Trapping Density (cm )
1e+12
14
WNM (state A), 1:1:1 WNM, 1:1:1 WNM (state A), 1:1:2 WNM, 1:1:2
12
10
8 0
2e+11
4e+11
6e+11
8e+11 -2
1e+12
Trapping Density (cm )
• The average WNM increases after stress, which is contrary to read SNM.! • The standard deviation of WNM increases after stress, which is similar to read SNM.!
16!
SNM vs WNM with SV and random charge trapping Correlation Coefficient
0 Correlation between SNM and WNM
-0.2
-0.4
(VL=’0’,VR=’1’), 1:1:1 Minimum, 1:1:1 (VL=’0’,VR=’1’), 1:1:2 Minimum, 1:1:2
-0.6
-0.8 0
2e+11
4e+11
6e+11
8e+11 -2
1e+12
Trapping Density (cm )
• Anti-correlation between SNM and WNM exists for one storing node.! • Minimum defined SNM and WNM show decorrelations, due to statistically independent transistors responding to two storing states.! 17!
Impact on Six-sigma yield stress induced degradations
µ - 6m (mV)
400
300 SNM, 1:1:1 SNM, 1:1:2 WNM, 1:1:1 WNM, 1:1:2
200
100
0 0
2e+11
4e+11
6e+11
8e+11 -2
1e+12
Trapping Density (cm )
• 6-sigma of read SNM is greatly affected by stress induced charge trapping, not only due to average SNM reduction, but also by boosted statistical variability. ! • 112-fin SRAM cells show much better stability than high-density fin cells.! 18!
Summary
• The random charge trapping effect can be accurately captured using the similar statistical compact modelling practice with statistical variability.! • SRAM cell read stability is degraded by stress induced charge trapping; The statistical variation of SNM and WNM increased with degradations.! • 112 FinFET SRAM shows much better stability compared with high-density SRAM cells.! • With the more random trapping, the read SNM six-sigma yield is reduced dramatically due to enhanced variation. ! 19!
Acknowledge
• It is in part supported by Scottish Funding Council through Knowledge Transfer Project “Statistical Design and Verification of Analogue Systems”.! ! ! !! !Thank you for your attention.! ! ! ! 20!
1 Device Modelling Group,! University of Glasgow, UK
! 2 Gold Standard Simulations Ltd, UK
3 IBM Research – Austin, USA
! 1! ESSDERC, 16-20 September 2013, Bucharest Romania! !
Outline! q Introduction! q 14 nm node DG SOI FinFETs! q Simulation of Random Charge Trapping and
Statistical Variability Sources! q Compact Modelling Methodology! q Charge Trapping Impact on SRAM SNM! q Charge Trapping Impact on SRAM WNM! q Summary!
2!
Introduction • Why this study?! • Novel 3-D architecture FinFET will be widely adopted at 14 nm technology, with reduced variability on SOI substrate due to tolerance to low channel doping.!
• However, (1) statistical aspect of reliability due to random individual trapping becomes an increasingly important issue. (2) In addition, charge trapping impact is affected by statistical variability sources.!
• Accurately modeling reliability of nanoscale transistors in circuit level should take care of above mentioned properties, therefore requires a “statistical” method, rather than describing average reliability behavior.!
• SRAM stability is susceptible to variability, therefore statistical study is needed.! 3!
FinFET
IMEC! Intel! 22nm! !
Veloso et. IEDM, 2009!
Fin Edge Roughness:! width, height, slope!
TSMC! Chang et., IEDM, 2009!
Bulk substrate!
IBM! Chang et., VLSI tech., 2011!
SOI substrate! 4!
Simulation Design of 14nm SOI FinFETs (GU-IBM collaboration) TE GA
MC calibrated @ 85°C!
W
fin
SOURCE
Double-Gate ! SOI FinFET!
LG
to
x
! E
XID DO
IE UR
B
TE RA
T BS
SU
Monte Carlo!
-1
2.5 2.0 source
7
fin
1.5
Velocity [x10 cm s ]
H
DRAIN
HM
Calibrated DD
drain
20!
EOT (nm)!
0.8!
WF (nm)!
10!
HF (nm)!
25!
NSD (cm-3 )!
3.0E20!
NCH (cm-3 )!
1.0E15!
VDD (V)!
0.9!
IOFF (nA/μm) !
10!
IDSAT (mA/μm) ! 0.9/0.8!
1.0
DIBL (mV/V)!
Default DD
0.5 0.0
Lg (nm)!
VG 0.420
30 40 Position [nm]
-0.8!
56/65!
Ref.: ITRS 2010 update!
50
5!
Intrinsic Parameter Fluctuations Statistical Variability Sources
TiN!
potential!
Random dopants!
Polysilicon/Metal Gate! Granularity!
Line Edge Roughness! 6!
Statistical variability simulation GER: ΔLG , ΔSCE! FER: ΔWFIN , Δconfinement! ΔRSD! RDD: ΔRSD , ΔNA! MGG: ΔΦM , Δψsurf !
• Each variability source has different impact on the device parameters and performance.! Wang, et al, IEDM 2011, pp103-106! 7!
Interaction: Charge trapping vs Statistical variability sources Sensitive regions!
• FER: local shortenings!
• MGG: metal grains with high currents underneath!
• RDD: current percolation paths! Wang et al, SISPAD 2012, pp.296-299! 8!
Vt RTS Distribution and Reliability are affected by Statistical Variability
1-CDF
1
Single Trapping!
0.1
Single Trapping 0.01 Uniform Device ’Atomistic’ Devices
0.001 0
Uniform device!
2
4
6
6VT (mV)
8
10
Wang et al., SNW 2012, pp.77-78!
Atomistic device!
• In the presence of SV, the RTS distribution tail is increased! Multi-trapping!
RTS: random telegraph signal!
9!
Random charge trapping effect on VT
Normal Quantile
4
2
-2
Trapping Density (cm ) 0 1E11 5E11 1E12
0
-2
-4 0.1
0.15
0.2
0.25
VT (V)
0.3
0.35
0.4
Poisson distribution ! of trapping charge ! number is assumed!
• First, the average VT shift increases with degradation heuristically;! • Most important, the statistical variability increases with degradation.! 10!
Statistical Compact Modelling Method • A small set of BSIM-CMG compact model parameters is used to extract statistical samples at fresh stage, also applied to degradation.! • In circuits random fresh samples are assigned, responding stressed samples are put for stressed transistors.! • Assume trapping effect is dynamically recoverable.! • e.g., M2 is biased with high VG and low VD, subject to PBTI!
retention!
PU! PG!
PG!
PD!
6-Transistor SRAM cell! PU: pull-up transistor, p-FinFET;! PG: pass-gate transistor, n-FinFET;! PD: pull-down transistor, n-FinFET;! 11!
What happens to SRAM SNM after stress? A
VR (V)
0.8
SNM(A)
0.6 0.4 0.2 0 0
SNM(B)
Fresh Stress (state A) Stress (state B)
0.2
0.4
B 0.6
VL (V)
0.8
SNM: static noise margin, the SRAM stability for read mode! State A: left 0, right 1; State B: left 1, right 0!
• Generally, stress induced trapping leads to less static noise margin ! • Heavier N/PBTI, more threshold shift, less stability! ! 12!
SNM Distribution
Two types of SRAM cells! with fin-number ratio! of PU:PG:PD,! 111 SRAM and 112 SRAM! are examined!
• First of all, the distribution is non-Gaussian.! • Compared with 111-fin SRAM cells, 112-fin cells increase SNM.! • With charge trapping induced degradations, the SNM is reduced.! 13!
Charge trapping effects on SNM 20
SNM (mV)
250
200
150
18
mSNM (mV)
SNM(A), 1:1:1 SNM, 1:1:1 SNM(A), 1:1:2 SNM, 1:1:2
16
SNM(A), 1:1:1 SNM, 1:1:1 SNM(A), 1:1:2 SNM, 1:1:2
14 12 10 8
100 0
2e+11
4e+11
6e+11
8e+11 -2
Trapping Density (cm )
1e+12
6 0
2e+11
4e+11
6e+11
8e+11 -2
1e+12
Trapping Density (cm )
• The average SNM is reduced by up to 30 mV, with charge trapping induced degradations.! • The statistical variation of SNM increases by 30-40% with degradation.! • 112-fin SRAM cells show better stability. Compared with 111-fin cells, 112-fin cells increases SNM by ~45% in average.! 14!
What happens to SRAM WNM after stress? 0.8
VL (V)
0.6 Fresh Stressed
0.4
0.2 WNM
0 0
!
0.2
0.4
0.6
0.8
VR (V) WNM: write noise margin, SRAM stability for write mode!
• In contrary to SNM, WNM increases a bit due to charge trapping.! • The WNM distribution is non-Gaussian.! 15!
Charge trapping effects on WNM
WNM (mV)
400
16
WNM (state A), 1:1:1 WNM, 1:1:1 WNM (state A), 1:1:2 WNM, 1:1:2
mWNM (mV)
420
380 360 340 320 0
2e+11
4e+11
6e+11
8e+11 -2
Trapping Density (cm )
1e+12
14
WNM (state A), 1:1:1 WNM, 1:1:1 WNM (state A), 1:1:2 WNM, 1:1:2
12
10
8 0
2e+11
4e+11
6e+11
8e+11 -2
1e+12
Trapping Density (cm )
• The average WNM increases after stress, which is contrary to read SNM.! • The standard deviation of WNM increases after stress, which is similar to read SNM.!
16!
SNM vs WNM with SV and random charge trapping Correlation Coefficient
0 Correlation between SNM and WNM
-0.2
-0.4
(VL=’0’,VR=’1’), 1:1:1 Minimum, 1:1:1 (VL=’0’,VR=’1’), 1:1:2 Minimum, 1:1:2
-0.6
-0.8 0
2e+11
4e+11
6e+11
8e+11 -2
1e+12
Trapping Density (cm )
• Anti-correlation between SNM and WNM exists for one storing node.! • Minimum defined SNM and WNM show decorrelations, due to statistically independent transistors responding to two storing states.! 17!
Impact on Six-sigma yield stress induced degradations
µ - 6m (mV)
400
300 SNM, 1:1:1 SNM, 1:1:2 WNM, 1:1:1 WNM, 1:1:2
200
100
0 0
2e+11
4e+11
6e+11
8e+11 -2
1e+12
Trapping Density (cm )
• 6-sigma of read SNM is greatly affected by stress induced charge trapping, not only due to average SNM reduction, but also by boosted statistical variability. ! • 112-fin SRAM cells show much better stability than high-density fin cells.! 18!
Summary
• The random charge trapping effect can be accurately captured using the similar statistical compact modelling practice with statistical variability.! • SRAM cell read stability is degraded by stress induced charge trapping; The statistical variation of SNM and WNM increased with degradations.! • 112 FinFET SRAM shows much better stability compared with high-density SRAM cells.! • With the more random trapping, the read SNM six-sigma yield is reduced dramatically due to enhanced variation. ! 19!
Acknowledge
• It is in part supported by Scottish Funding Council through Knowledge Transfer Project “Statistical Design and Verification of Analogue Systems”.! ! ! !! !Thank you for your attention.! ! ! ! 20!
