effect of discrete dopant distribution on mosfets scaling...
TRANSCRIPT
Effect of Discrete Dopant Distribution on MOSFETs Scaling into the Future
Yoshio Ashizawa and Hideki Oka Fujitsu Laboratories Ltd.
50 Fuchigami, Akiruno, Tokyo, 197-0833, Japan
E-mail: [email protected]
Abstract - Device characteristics fluctuation due to
discrete dopant distribution is one of the serious
problems for future device scaling. We have evaluated
the potential field on the substrate by applying 2D FFT.
It is found that specific wave length range, 20 nm to 25
nm, is dominant to device characteristics. Moreover, if
device size scales less than this specific wave length, there
will be size effect caused by local potential fluctuation.
I. INTRODUCTION
Device characteristics fluctuation induced by random
dopant distribution has been investigated using combination
of Synopsys Taurus Process Atomistic simulator[1] and
SELETE HyDeLEOS device simulator[2]. 2D FFT is
applied to the doping concentration and potential fields at the
Si surface to characterize the fluctuations. It is found that the
larger fluctuation will be predicted among transistors when
device size becomes smaller than specific the wave length
determined by spatial potential fluctuations.
II. FLUCTUATION INDUCED BY DISCRETE DOPANT
DISTRIBTUTION
Fig. 1 shows the simulation procedure and device
structure. Initial boron is given at constant concentration in
the substrate. After annealing, boron field is converted to
discrete atomistic distribution by DADOS. As dopant
distribution in the substrate is dominant to device
characteristics fluctuation in our preliminary works, all
dopant profiles are the same except for substrate, in which
dopant distribution is atomistically simulated.
• Process Simulator
– TAURUS Process Atomistic
– Initial boron distribution• Uniform
– Anneal• Kinetic Monte Carlo
• DADOS integrated into TPA
– Discrete Boron distribution
• Device Simulator– SELETE HyDeLEOS
– Drift diffusion
– Long-range part of Coulomb potential• Sano model
90nm Node Structure
W=100nm100nm
50nm
B Na=5×1018cm-3
S
G
DTox=1.2nm
B
AsAnalytic
Atomistic
L=50nm
Atomistic
90nm Node Structure
W=100nm100nm
50nm
B Na=5×1018cm-3
S
G
DTox=1.2nm
B
AsAnalyticAs
Analytic
Atomistic
L=50nm
Atomistic
32
3
)(
)cos()()sin(
2)(
rk
rkrkrkkr
c
cccclong −=π
ρ
Fig. 1: Simulation procedure and device structure.
After process simulation, atomistic boron positions are
applied to drift diffusion device simulator, in which Sano
model[3] is used for the long-range part of Coulomb
potential of individual impurities.
Table 1 is the summary of target device parameters for
scaling nMOSFETs, where Na is the acceptor concentration
in the substrate.
Table 1: Target parameters for scaling nMOSFETs.
0
5
10
15
20
25
30
0 0.01 0.02 0.03 0.04 0.05
ITRS 2003Mizuno et. alToriyama et. al
σVth
(mV)
65nm
90nm
L=W=10nm
Tox=0.4nm
L=W=25nm
Tox=0.8nm
L=W=100nm
Tox=3nm
32nm Node
45nm
LWTox / ( )
Na=1018cm-3
0
5
10
15
20
25
30
0 0.01 0.02 0.03 0.04 0.05
ITRS 2003Mizuno et. alToriyama et. al
σVth
(mV)
65nm
90nm
L=W=10nm
Tox=0.4nm
L=W=25nm
Tox=0.8nm
L=W=100nm
Tox=3nm
32nm Node
45nm
LWTox / ( )
Na=1018cm-3
Fig. 2: σVth vs. dimensionless size factor with scaling.
Fig. 2 shows the relation between calculated and reported
[4][5] σVth and dimensionless size factor LWTox , which
reflects the degree of geometrical scaling. Because vertical
scaling is behind horizontal scaling, dimensionless size
factors are increased with scaling.
As for the slope between the plot and the origin, it is a
concentration factor, which represents the contribution of
concentration Na to σVth. Generally, the slope is considered
to get steeper with increasing Na. But all our results for these
roadmap parameters are on the same line even though each
impurity concentration is different. However, we got the
same results with others when Na is decreased to 1018 cm-3 as
is not shown in this figure. This means that the concentration
effect becomes saturated over 5 x 1018 cm-3.
We have extended scaling parameters to make sure if the
concentration factor actually saturates with other parameters.
Fig. 3 shows the response surface for the slope, in which
F(Na) is calculated from σVth divided by the dimensionless
size factor for each technology node and acceptor
concentration Na.
2x10191.4x10198x10185x1018Na(cm -3)
0.91.01.11.2Vdd(V)
0.60.70.91.2Tox(nm)
12182636L(nm)
32nm Node45nm Node65nm Node90nm Node
2x10191.4x10198x10185x1018Na(cm -3)
0.91.01.11.2Vdd(V)
0.60.70.91.2Tox(nm)
12182636L(nm)
32nm Node45nm Node65nm Node90nm Node
31
2-4
The surface is mildly curved at lower Na but getting evener
at higher Na not only along the roadmap parameters but also
with other parameters.
Fig. 3: Response surface of the concentration factor with
various target parameters.
III. POWER SPECTRUM ANALYSIS
It is greatly helpful to characterize dopant distribution or
potential field by the spatial frequency space because spatial
frequency space is closely related with real space (Table 2).
There is only 2D random noise observed for phase angle in
spatial frequency space because discrete dopant is randomly
and disorderly distributed. Its effect is considered to be the
same level as baseline noise for all calculations. Therefore,
only power is discussed in this work although phase angle is
also important.
Table 2: Correspondence between real space and spatial
frequency space.
Phase AnglePosition
In teraction
Power SpectrumConcentra tion
Potentia l
Spatial Frequency SpaceReal Space
Phase AnglePosition
In teraction
Power SpectrumConcentra tion
Potentia l
Spatial Frequency SpaceReal Space
2
i
2 ),(),(),( vuFvuFvuP r +=
),(
),(tan 1
vuF
vuF
r
i−=ξ
2D FFT1
0
1
0
)(2
2),(
1),(
N
x
N
y
vyuxN
j
eyxfN
vuFπ π
θ
θ2
0
),()( fPfQ
1D Trans.
Real Space Spatial Frequency Space
x
y
22 ),(),(
),(
vuFvuF
vuP
ir
v),( vuF
1D Power Spectrum
)1log(
log)(log
α
α
fk
fkfQ
log f
log Q
Slope= -α
2D Complex F(x,y)
0
),( yxf
100nm x 100nm
2D Field
Integral
Contribution
Na(x,y)f(x,y)=
ϕ(x,y)
22 vuf
P(f,θ)
u0
2D FFT1
0
1
0
)(2
2),(
1),(
N
x
N
y
vyuxN
j
eyxfN
vuFπ π
θ
θ2
0
),()( fPfQ
1D Trans.
Real Space Spatial Frequency Space
x
y
22 ),(),(
),(
vuFvuF
vuP
ir
v),( vuF
1D Power Spectrum
)1log(
log)(log
α
α
fk
fkfQ
log f
log Q
Slope= -α
2D Complex F(x,y)
0
),( yxf
100nm x 100nm
0
),( yxf
100nm x 100nm
2D Field
Integral
Contribution
Na(x,y)f(x,y)=
ϕ(x,y)
Na(x,y)f(x,y)=
ϕ(x,y)
22 vuf
P(f,θ)
u0
Fig. 4: Application of 2D FFT to f(x,y) and transformation to
1D power spectrum.
Fig. 4 shows the schematic procedure to apply FFT.
F(u,v) is calculated by 2D FFT against 2D function f(x,y),
which corresponds to concentration or potential field.
Power P(u,v) is defined as the square norm of the complex
function of F(u,v), where Fr(u,v) and Fi(u,v) represent the
real and imaginary part of F(u,v) respectively.
Because P(u,v) is a 2D function in spatial frequency space,
it is still difficult to compare or characterize it directly.
Therefore, P(u,v) is transformed to 1D power spectrum Q(f)
by integrating P(f,θ) along all circumferences. 1D power
Q(f) represents the contribution of each spatial frequency.
Moreover, the slope characterizes the degree of the
fluctuation.
0.0001
0.01
1
100
104
1 10 100
1e18(cm-3)
5e18(cm-3)
8e18(cm-3)
1.4e19(cm-3)
2e19(cm-3)
Power Spectrum (arb. unit)
Spatial Frequency (1/100nm-1)
Na
S DG
(a) Acceptor Conc. Field
90nm Node
L=36nm
W=100nm
Up
Slope= - 2
0.0001
0.01
1
100
104
1 10 100
1e18(cm-3)
5e18(cm-3)
8e18(cm-3)
1.4e19(cm-3)
2e19(cm-3)
Power Spectrum (arb. unit)
Spatial Frequency (1/100nm-1)
Na
S DGS DG
(a) Acceptor Conc. Field
90nm Node
L=36nm
W=100nm
Up
Slope= - 2
(b) Potential Field
1000
104
105
106
107
108
109
1010
1 10 100
1e18(cm-3)5e18(cm-3)8e18(cm-3)1.4e19(cm-3)2e19(cm-3)
Power Spectrum (arb. unit)
Spatial Frequency (1/100nm-1)
Na
S DG
90nm Node
L=36nm
W=100nm
Slope= - 2
(b) Potential Field
1000
104
105
106
107
108
109
1010
1 10 100
1e18(cm-3)5e18(cm-3)8e18(cm-3)1.4e19(cm-3)2e19(cm-3)
Power Spectrum (arb. unit)
Spatial Frequency (1/100nm-1)
Na
S DGS DG
90nm Node
L=36nm
W=100nm
Slope= - 2
Fig. 5: 1D power spectrum for acceptor concentration field
(a) and potential field (b) at the Si surface.
Fig. 5 (a) shows power spectrum for acceptor
concentration field. Power is also observed to be saturated at
higher acceptor concentration, which is similar in Fig. 2 and
Fig. 3. Fig. 5 (b) shows the case of the potential field.
Both fields obey the 1/f2-rule generally because these fields
are calculated from the second order of PDE for diffusion or
potential. However, there is less noise in potential field than
in concentration field. In potential field, high spatial
frequency wave is eliminated from concentration field by
solving the Poisson equation because the radius of the
long-range part of Coulomb potential by an acceptor has
LWT
VNANF
ox
thxaa
σ)(
LWT
VNANF
ox
thxaa
σ)(
32
wider distribution than that of concentration, as is discussed
later.
There is almost no concentration dependence among
potential fields except for the specific frequency range of
5/100 nm-1 to 4/100 nm-1. This spatial frequency range
corresponds to the wave length of 20 nm to 25 nm.
IV. POTENTIAL FIELD WITH SPECIFIC WAVE
LENGTH RANGE
There are two reasons why FFT is adopted to characterize
fluctuation. One reason is to find what wave length range
characterizes the fluctuation and the other reason is to
reconstruct the original field with the specific wave length
range by applying FFT using a window function.
Fig 6 (a) shows the original surface potential field with
random dopant distribution. Fig. 6 (b) shows the
reconstructed potential field by inverse FFT with waves
above 20 nm wave length only.
Inverse FFT
λ≧20nm
90nm Node L=36nm
W=100nm
(a) Surface Potential Field (Original)
(b) Reconstructed Field with Low
Spatial Frequency Component
S
D
D
S
Na=5 ×1018cm-3
Inverse FFT
λ≧20nm
90nm Node L=36nm
W=100nm
(a) Surface Potential Field (Original)
(b) Reconstructed Field with Low
Spatial Frequency Component
S
D
D
S
Na=5 ×1018cm-3Na=5 ×10
18cm-3
Fig. 6: Original surface potential field and reconstructed
field by inverse FFT with waves above 20 nm wave length.
Comparing Fig. 6 (a) with Fig. 6 (b), the original potential
field is almost reconstructed and the specific wave length
takes a dominant role in fluctuation of Vth. Moreover, the
Poisson equation behaves as a low-pass filter for
concentration field because potential filed doesn’t reflect the
long-range part of individual impurities. Fig. 7 (a) shows
relative acceptor concentrations for the long-range part of
Sano model by an acceptor at the origin. Fig. 7 (b) shows the
case of potential field. As Na is increased, the relative
acceptor concentration increases but the radius shrinks
gradually. However, as for potential field, the peak value at
the origin goes down but the radius stays almost constant in
contrast to the concentration field.
Because potential field is described by overlapping these
single waves, specific wave length is observed by Fourier
analysis, whose range corresponds to the wave length of 20
nm to 25 nm.
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 5 10 15 20 25
1e17(cm-3)
1e18(cm-3)
5e18(cm-3)
8e18(cm-3)
1.4e19(cm-3)
2e19(cm-3)
Relative Acceptor Conc. ( )
Radius Distance (nm)
Na
(a) Relative Acceptor Concentration
Shrink
Up
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 5 10 15 20 25
1e17(cm-3)
1e18(cm-3)
5e18(cm-3)
8e18(cm-3)
1.4e19(cm-3)
2e19(cm-3)
Relative Acceptor Conc. ( )
Radius Distance (nm)
Na
(a) Relative Acceptor Concentration
Shrink
Up
-80
-70
-60
-50
-40
-30
-20
-10
0
0 5 10 15 20 25
1e17(cm-3)
1e18(cm-3)
5e18(cm-3)
8e18(cm-3)
1.4e19(cm-3)
2e19(cm-3)
Potential (mV)
Radius Distance (nm)
Na
(b) Potential Distribution
Down
Half of Specific
Wave Length
-80
-70
-60
-50
-40
-30
-20
-10
0
0 5 10 15 20 25
1e17(cm-3)
1e18(cm-3)
5e18(cm-3)
8e18(cm-3)
1.4e19(cm-3)
2e19(cm-3)
Potential (mV)
Radius Distance (nm)
Na
(b) Potential Distribution
Down
Half of Specific
Wave Length
Fig. 7: Relative acceptor concentration (a) and its potential
field (b) by an acceptor at the origin.
V. DISCUSSION
Assuming that the specific wave length in the potential
field, there will be a new fluctuation caused by size effect
between wave length and device size (Fig. 8).The potential
field is fluctuated within a device for large scale devices.
Scaling the device size, if Na in the substrate is increased
accordingly, potential is still averaged within a device
(Center). But if the device size becomes smaller than
specific wave length of 20 nm to 25 nm, the fluctuation is
determined by the local potential field directly (Right). In
such case, the frequency of Vth distribution will obey
32
3
)(
)cos()()sin(
2)(
rk
rkrkrkkr
c
cccclong −=
πρ
33
Poisson distribution rather than symmetric normal
distribution. Moreover, this will cause a fundamental
problem to engineers in designing circuits with conventional
3σ criteria because Poisson distribution has broader
distribution even if σ is the same level with that of normal
distribution.
Geometry scaling is aggressive, while electrical properties,
such as unit charge, dielectric, long-range part of Coulomb
potential and so on, are kept almost constant. This unbalance
for scaling causes the size effect.
ρε
φeDopant
Normal dist.
Radius by the
long-range part of
Coulomb potential
Potential
Normal dist. Poisson dist.
Potential
25nm
DopantPoisson Equation
Wave length
(Left)
Averaged
(Right)
Local difference(Center)
Still averaged
Device size
Device size
Scaling Scaling
ρε
φeDopant
Normal dist.
Radius by the
long-range part of
Coulomb potential
Potential
Normal dist. Poisson dist.
Potential
25nm
DopantPoisson Equation
Wave length
(Left)
Averaged
(Right)
Local difference(Center)
Still averaged
Device size
Device size
Scaling Scaling
Fig. 8: Size effect by relative relation between device size
and wave length of the potential fluctuation.
G
Narrow
ShrinkStay
almost
const.
0 50 100 150
90nm Node
65nm Node
45nm Node
32nm Node
32nm Node
32nm Node
⊿Potential (max) (mV)
W=22nm
W=10nm
W=36nm
Maximum Potential Difference in a Device
GG
NarrowNarrow
ShrinkShrinkStay
almost
const.
0 50 100 150
90nm Node
65nm Node
45nm Node
32nm Node
32nm Node
32nm Node
⊿Potential (max) (mV)
W=22nm
W=10nm
W=36nm
0 50 100 150
90nm Node
65nm Node
45nm Node
32nm Node
32nm Node
32nm Node
⊿Potential (max) (mV)
W=22nm
W=10nm
W=36nm
Maximum Potential Difference in a Device
Fig. 9: Scaling rules and maximum potential difference in a
device.
0
5
10
15
20
200 300 400 500
W=72nmW=36nmW=22nmW=10nm
Freq
uenc
y ( )
Vth (mV)
32nm NodeL=12nm
Na=2×1019(cm-3)
Fig. 10: Histogram of Vth Frequency distribution at 32 nm
node with narrowing device.
Fig. 9 shows scaling rules to see the size effect and
maximum potential difference in a device. There is a
remarkable jump at 32 nm node with narrowing width.
Fig. 10 is the frequency of Vth distribution at 32 nm node
for various device widths. Distribution is getting broader and
lower with narrowing device.
Fig. 11 shows “normal probability plot”, in which narrower
width’s curve deviates from others and it is the proof of
departing from symmetric normal distribution. This means
the increase of fluctuation should be notified because of the
broader distribution tail due to Poisson distribution.
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
200 250 300 350 400 450 500
W=72nmW=36nmW=22nmW=10nm
Nor
mal
Dev
iatio
n R
atio
to σ
Vth
( )
Vth (mV)
32nm NodeL=12nm
Na=2×1019(cm-3)
Poisson
Dist.
Fig. 11: Normal probability plot of σVth at 32 nm node with
narrowing device.
VI. CONCLUSION
We have evaluated the device characteristics fluctuation
against the technology roadmap. Absolute value of the
fluctuation will be increased with scaling but concentration
factor will be saturated for high acceptor concentration over
5 x 1018 cm-3. Fourier analysis shows that power of spatial
frequency generally obeys the 1/f2-rule and concentration
dependence is less sensitive in potential filed than in
concentration field.
This work suggests that it becomes hard to fabricate
hundreds million transistors with uniform characteristics in a
chip if device size is close to the specific wave length, 20 nm
to 25nm, of the potential fluctuation especially.
REFERENCES
[1] Taurus-Process Reference Manual, Version W-2004.09,
(Synopsys Inc., Mountain View, CA, 2004).
[2] Unpublished (http://www.selete.co.jp/).
[3] N. Sano et al., IEDM Tech. Dig., pp. 275-278, 2000.
[4] T. Mizuno et al., IEEE Trans. Electron Devices, vol. 41,
pp. 2216-2221, 1994.
[5] A. Asenov et al., IEEE Trans. Electron Devices, vol. 50, pp. 1837-1852, 2003.
34