Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

1352 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 63, NO.

3, MARCH 2016

An Analytical Model to Estimate FinFET’s VT


Distribution Due to Fin-Edge Roughness
S. Mittal, A. S. Shekhawat, and U. Ganguly, Member, IEEE

Abstract— Line-edge roughness induced fin-edge rough- for FER variability evaluation at device level is attractive.
ness (FER) is the primary source of VT variation in FinFETs. To this purpose, there are few analytical models reported
Conventionally, stochastic simulations are performed to predict in literature [2], [4], [5]. Cheng et al. [4] assume fin edges to
the device variability due to FER for a technology, which are
computationally expensive. An analytical formulation to predict be sinusoids with wavelength as multiples of correlation length
variability due to FER enables understanding of the effect of to model FER, which is a gross simplification and unable
input parameters as well as provides quantitative results at frac- to provide stochastic VT distribution. The formulation in [5]
tional computational costs. In this paper, we develop and present is applicable to planar MOSFETs, not FinFETs. Patel [2]
an analytical model to estimate saturation VT (VT −sat ) variability assumes a linearity between VT and Wfin , which does not hold
due to FER. The model is capable of capturing the VT variability
dependence on device parameters (L G and Wfin ) and variability true for FinFETs affected by quantum confinement (QC).
parameters (correlation length  and standard deviation ) In our recent model [6], we show that suppose a FinFET
accurately. The entire VT −sat distribution obtained by the model with FER has a minimum Wfin (Wmin ) occurring somewhere
is also presented and compared against the VT −sat distribution between source and drain, then its VT −sat is essentially the
of stochastic simulations to show that the model captures the same as VT −sat of a FinFET with uniform Wfin of Wmin .
distribution effectively. We show that not only σ VT but even
μVT is affected by variability parameters. Hence, such modeling This is because maximum QC occurs at the position of
is critical to defining nominal FinFET structure (L G and Wfin ), Wmin in a FinFET, which provides the maximum barrier and,
which is affected by variability ( and ) especially for scaled hence, defines VT −sat of FinFET. Thus, to obtain the VT −sat
FinFETs, where quantum-confinement effects are enhanced. distribution, first, the Wmin distribution is obtained from the
Index Terms— Fin-edge roughness (FER), FinFET, line-edge Langevin equation-based FER simulations [3]. Second, Wmin
roughness (LER), modeling, VT variability. distribution is used to estimate VT −sat distribution based on
the dependence of VT −sat on uniform Wfin FinFETs obtained
I. I NTRODUCTION either from simulations [6] or compact models of FinFETs [7].
The implicit assumption in this model is a strong local con-
L INE-EDGE roughness (LER) induced fin-edge rough-
ness (FER) is a major contributor to the variability
in advanced FinFETs in sub-20-nm node [1]. Various fin
finement at Wmin which may not be accurate for approximately
uniform fins or where QC is weak.
In this paper, we present a model to calculate VT −sat
patterning technologies, for example, self-aligned dual pat-
distribution of FinFETs due to FER variability, which is valid
terning, extreme ultraviolet, nanoimprint lithography, and so
both in weak and strong QC regimes. The model is validated
on [2], are characterized by their LER at a given critical
for a wide range of L G s (12–20 nm) [which is equivalent
dimension. The impact of the obtained LER-dependent FER
to a technology nodes (7–14 nm [8])], Wfin s (3.5–8 nm),
variation at device level is essential in order to benchmark
and correlation length  (5–50 nm). The entire saturation
and qualify a technology. The conventional procedure to
VT (VT −sat ) distribution obtained by the model is compared
study the device level impact of FER variation is to perform
against VT −sat distribution of stochastic simulations to show
a large number of stochastic simulations [1], [3]. Stochas-
that the model captures the distribution effectively.
tic simulations provide accurate analysis of various device
This paper is arranged as follows. Section II provides
parameters but are prohibitively computationally expensive.
the details of the Sentaurus [9] simulation deck used in
In addition, analytical models provide superior insights of the
simulations. Section III presents the methodology of the
significance of individual parameters compared with stochastic
model development. In this section, we describe in detail the
simulations. Therefore, the development of an analytical model
formulation of phenomenological model for the distribution of
Manuscript received November 2, 2015; revised December 27, 2015; minimum fin width (Wmin ). This is followed by VT estimation
accepted January 20, 2016. Date of publication February 8, 2016; date of based on percolation model [10] to generate Q–Q plots of
current version February 23, 2016. This work was supported in part by the
Department of Science and Technology, India, and in part by the Department VT distribution. Results using this formulation are presented
of Electronics and Information Technology within the Ministry of Communi- in Section IV. Benchmarking and discussion are presented in
cations and Information Technology. The review of this paper was arranged by Section V. Finally, the conclusion is drawn in Section VI.
Editor R. M. Todi.
The authors are with the Department of Electrical Engineering, IIT Bombay,
Mumbai 400076, India (e-mail: smittal@iitb.ac.in; shekhawat@ee.iitb.ac.in;
udayan@ee.iitb.ac.in). II. TCAD S IMULATION D ECK
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. The Sentaurus TCAD test bench was used to perform
Digital Object Identifier 10.1109/TED.2016.2520954 FinFET variability studies. For reasonable estimations,
0018-9383 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
MITTAL et al.: ANALYTICAL MODEL TO ESTIMATE FinFET’s VT DISTRIBUTION DUE TO FER 1353

TABLE I
PARAMETERS U SED FOR FinFET S IMULATION

first, the test bench was calibrated against experimental


data [11] at L G = 25 nm. The details of calibration are
provided in [3]. As described in [3], the 2-D structure has
been used in simulations in consistence with FinFET structures
in [12].
This well-calibrated 2-D simulation deck was used to
simulate variability for FinFET at various technology nodes in
this paper. For all the technology nodes presented in this paper
[7 nm (L G = 12 nm), 10 nm (L G = 15 nm), and 14 nm (L G =
20 nm)], an Effective Oxide Thickness of 0.67 nm, L SPACER of Fig. 1. VT −sat versus uniform Wfin -based SC compared with VT −sat varia-
tion versus Wmin /Wmax /Wavg in SA for various Wfin and . (a) Wfin = 5 nm
5 nm, and VDD of 0.78 V [8] were used. A source/drain doping and  = 5 nm, (b) Wfin = 5 nm and  = 50 nm, and (c) Wfin = 3.5 nm and
of 4 × 1020 was used along with a channel doping of 1 × 1017 .  = 5 nm for typical 3 = 2 nm. Insets of (a)–(c): cartoon of sample FinFET
The parameters are shown in Table I. generated, which indicate that the increase in  produces a more uniform fin
and the decrease in Wfin causes enhanced QC. (b) shows for uniform fins,
To simulate the FER-based variability, method presented Wmin , Wmax , and Wavg correlate well with SC. (c) shows that at smaller 
in [3] has been adopted. Gaussian autocorrelation function [13] and Wfin , Wmin , Wavg , and Wmax split and Wmin correlates well with SC.
was used to generate stochastic lines with specified correlation (d) RMS error between SC and stochastic Wmin , Wavg , and Wmax
versus /L G . At smaller Wfin and /L G , Wmin has minimum error, while
length  and 3 variation. To study VT variability in this Wavg has least error at larger /L G and Wfin . Thus, there are different
paper, saturation VT (VT −sat ) was extracted from I D –VG regimes, where Wmin and Wavg define the distribution better.
curves at VDS = VDD by constant current method at
I = 10−2 mA/μm, as also described in [3].
To summarize, Fig. 1(d) shows the rms error of Wmin , Wmax ,
III. M ETHODOLOGY and Wavg -based SA results with SC for various Wfin s and
/L G s. Wmax does not show minimum rms error for any Wfin s
A. VT −sat Dependence on Wmin Versus Wavg
and s. At smaller Wfin = 3.5 nm and /L G < 0.8, Wmin
Previously, we have demonstrated that Wmin correlates with shows the minimum rms error. Wavg shows the minimum error
VT −sat of a FinFET, where local QC effects are strong [6]. elsewhere but the error increases sharply at lower /L G . Thus,
A large parameter space is explored in detail in this section. we see that Wmin better defines the VT −sat distribution for
VT −sat versus uniform Wfin sensitivity check (SC) curve is smaller Wfin s and s, while Wavg defines it better elsewhere.
compared with stochastic analysis (SA) results of 200 samples This indicates that the VT −sat distribution could be a weighted
at L G = 15 nm in Fig. 1. For each stochastically generated fin, function of both Wmin and Wavg for Wfin and  lying in
the average Wfin (Wavg ), minimum Wfin (Wmin ), and maximum between these two regimes. To obtain this function, we first
Wfin (Wmax ) are extracted along with VT −sat . Each VT −sat is develop an analytical model of Wmin and Wavg distribution
plotted against its respective Wmax , Wmin , and Wavg to evaluate depending on /L G and Wfin . Second, we use a percolation
its extent of correlation with SC. Fig. 1(a) shows the data for model to estimate VT −sat for a fin with varying Wfin . Finally,
a nominal Wfin = 5 nm and  = 5 nm, Fig. 1(b) shows we obtain a full VT −sat distribution and validate it with SA.
the data for nominal Wfin = 5 nm and  = 50 nm, while
Fig. 1(c) shows the data for nominal Wfin = 3.5 nm and B. Phenomenological Model for μWmin , μWmax , and μWavg
 = 5 nm. The schematic of one of the samples of randomly
Fig. 2(a) shows the μWmin , μWmax , and μWavg dependence
generated FinFET is also shown in inset. It simply shows
on /L G based on stochastic simulations [6]. μWmin decreases
that for larger  [Fig. 1(b)], fin looks uniform (closer to
with the decrease in /L G because of the presence of more
SC uniform fins case), thus all variants of SA, such as Wmin ,
local minima (undulations) in one channel length, while
Wmax , and Wavg , correlate well with SC. For smaller , the fin
μWmax increases because of the presence of more maxima.
has significant undulation (high spatial frequency), as shown in
μWavg remains constant. To model the distribution of Wmin ,
Fig. 1(a). Therefore, the distribution of Wmin , Wmax , and Wavg
we use the Brownian motion-based model of FER described
moves apart as  decreases, as shown in Fig. 1(a) compared
by Langevin equation [14]. Equation (1) shows the Langevin
with Fig. 1(b). The distributions move further apart as Wfin
equation
decreases, as shown in Fig. 1(c) compared with Fig. 1(a). For
smaller Wfin and  [Fig. 1(c)], Wmin correlates well with SC, dy y
= F(x) − . (1)
where QC effect is prominent. dx 
1354 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 63, NO. 3, MARCH 2016

of Wfin . The expectation of the minima and maxima will


provide estimates for μWmin and μWmax , respectively. There
exist two cases for the position of Wmin , as shown in Fig. 2(c).
For the mth upper line, the minimum will lie at x = 0 for 1st
to mth lower line [Fig. 2(c) (Case 1)] and at x = L G for rest
of the lower lines [Fig. 2(c) (Case 2)]. A weighted summation
to obtain μWmin is done as follows:

μWmin
 n     
W W
= pm p1 Am + − A1 −
2 2
m=1
   
W W
+ p2 Am + − A2 −
2 2
   
W W
+ · · · + pm Am + − Am −
2 2
   
LG W LG W
+ pm+1 Am e−  + − Am+1 e−  −
2 2
    
LG W L W
Am e−  + − An e−  −
G
+ · · · + pn
2 2
(2)

where A1 , A2 etc. are the initial amplitudes, i.e., y(x = 0),


of different lines, p1 , p2 etc. are the probabilities of A1 , A2
etc., respectively, drawn from a normal probability distrib-
ution, W is the nominal Wfin , L G is the channel length,
Fig. 2. (a) μWmin , μWmax , and μWavg with /L G at nominal
Wfin = 5 nm at FER 3 = 2 nm, extracted from randomly generated lines.
 is the correlation length, and n is the total number of
μWavg does not change (=nominal Wfin ) with /L G , while μWmin decreases random lines considered. By simplifying the above equation,
and μWmax increases as /L G decreases. (b) Each random line is governed we get
by two phenomenon (1): i) random force causing variation and ii) decay to
nominal position. Average of few random lines starting from a fixed y is thus

n 
m LG

a decay line and it matches the exponential decay. (c) Spectrum generated
around upper and lower edge of FinFET with the mean decay lines. For the
μWmin = W + pm A m pi 1 − e − 
mth upper fin line and the ith lower fin line, depending on i, two cases exist. m=1 i=1
Case 1: for i < m, Wmin lies at x = 0. Case 2: for i > m, Wmin lies at 
n 
m LG

x = L G . For i = m, Wfin is constant (=W ) from x = 0 to L G . − pm pi A i 1 − e −  (3)


m=1 i=1
n
The first term in the RHS is the Gaussian zero-mean random where we have used relation that m=1 pm = 1
n
force, which is the cause of randomness in the line. The second and p
m=1 m mA = 0. To obtain the distribution of
term is the decay function, which brings the line back to its A [y(x = 0)], the following procedure is used. A line
nominal position after it is deviated from the nominal because is expressed as y = Ae−(x/). The mean of this line is
of the random force term. If the initial point, i.e., y(x = 0), estimated as
LG
 LG 
is deviated from the mean, then in the absence of the random
1
term, the line will decay to its mean position. Now, even in the My = yd x = A 1−e−  . (4)
presence of the random force, the mean trajectory of y(x) is LG 0 LG
an exponential decay function as the random force will, on the Thus, we can write
average, get cancelled out as it is a zero-mean process. As an
example, Fig. 2(b) shows that the mean line of many random LG 1
A = My . (5)
lines from the same starting point would approximately be a  1 − e− LG
decay function even though the individual lines are subjected
to the random force.
We assume that M y has a Gaussian probability distribution
Next, we generate a set of exponentially decaying mean
with standard deviation σ () of the line with FER. Thus,
lines with different starting deviations from the nominal edge,
substituting (5) in (3), we get μWmin as
for both the upper and lower edges of a FinFET [Fig. 2(c)].
The initial amplitude of the decay function, i.e., y(x = 0),  n 
LG    
m n m
is assumed to have a normal probability distribution. From μWmin = W − pm pi M yi − pm M ym pi
this basis of lines, each combination of one upper and one 
m=1 i=1 m=1 i=1
lower edge mean line will provide a minima and maxima (6)
MITTAL et al.: ANALYTICAL MODEL TO ESTIMATE FinFET’s VT DISTRIBUTION DUE TO FER 1355

Fig. 4. (a) Schematic showing the methodology to estimate μVT due to FER
variability. Thickness of line shows the probability of occurrence of that line.
Wi (x) is the function of μWmin , μWmax , and Wmin −i. (b) cartoon of FinFET
with Wfin = W i (x). (c) Resistive network model of FinFET showing the
maximum resistance at Wmin position and least resistance at Wmax position.

Fig. 3. Comparison of μWmin and μWmax model with the data generated
using Gaussian autocorrelation function-based random lines versus /L G Thus, n fins may be created, as shown in Fig. 4(a).
at L G = 15 nm and Wfin = 5 nm for different 3 deviations. Greater The i th fin has a linearly changing W (x) given by
than 92% R 2 match is observed for all ranges of 3 deviations. A set  
of 50 exponentially decaying lines [n = 50 in (6) and (7)] were used in
i Wmax (i ) − Wmin (i )
these results. W (x) = x + Wmin (i ) (8)
LG
 
μWmax − μWmin
where M yi varies from −3σ to 3σ as i goes from 1 to n. W i (x) = x + Wmin (i ). (9)
LG
Similarly, μWmax is given as
μWmax As Wmax (i )−Wmin (i ) = μWmax −μWmin , the slope of W (x)
  is fixed. Such a tapered fin, as shown in Fig. 4(b), is amenable
LG 
n 
m 
n 
m
=W+ pm pi M yi − pm M ym pi . to the application of percolation theory [10]. The position of
 minimum fin width would offer maximum resistance, while
m=1 i=1 m=1 i=1
(7) the position of maximum fin width would offer least resistance
[Fig. 4(c)]. Next, we see how these resistances are added in
Fig. 3 compares the results of this model with the data subthreshold regime in order to find effective threshold voltage
generated using Gaussian autocorrelation function-based ran- for this FinFET with variable Wfin .
dom lines for different  variations at L G = 15 nm and Ohm’s law current equation is given as
Wfin = 5 nm. n = 50 [(6) and (7)] has been used in these
results. Excellent match (R 2 > 92%) is obtained between the I = Aqμn E (10)
presented model and the stochastic data. Thus, this simple
model captures the dependence of μWmin and μWmax on , , where A is the area, μ is the mobility, n is the number of
Wfin , and L G . The results become insensitive to n for n > 30 charge carriers, and E is the electric field. At a position x
(not shown). [Fig. 4(b)], for a fin of length dx, current would be given as
We have shown earlier that the standard deviation
of Wmin and Wmax is constant and independent of /L G [6]. dV x
I = Aqμn(x) (11)
Wfin is the difference between the upper and lower edges dx
of a fin. As both upper and lower edges are defined by
where n(x) is the number of charge carriers at position x
independent Gaussian random process, as holds true√for resist
and dVx is the potential drop across fin of length dx. In sub-
defined fins [2], the standard deviation of Wfin is 2 times
threshold regime, the charge carriers are given by the following
standard
√ deviation of one edge. This theoretical value of
relation:
2 is experimentally verified for Wmin , Wavg , and Wmax
distributions [6]. Based on the above model for average and q(V gs−VT (x))

standard deviation, the distribution of Wmax , Wmin , and Wavg n(x) = λe ηkT (12)
is determined. Next, we present how to use these distributions
where λ is the proportionality constant, Vgs is the applied
in order to estimate VT variability due to FER.
gate bias, VT (x) is the threshold voltage at position x, which
depends on W (x), and η is the ideality factor which is essen-
C. Tapered Fin Percolation Model for a Non-Uniform Fin tially the effectiveness of source-channel barrier modulation by
The Wmin , Wavg , and Wmax distributions are Gaussians applied gate bias. As a first-order approximation, the variation
with the same σ with μWmax and μWmin symmetrically in η with fin width was ignored. We assumed a constant
shifted about μWavg , which is the same as nominal Wfin η = 1.16 as the electrostatics is excellent in FinFETs with
[Fig. 2(a)]. If each distribution is divided into n fin widths (W ) subthreshold slope 70 mV/decade. VDS dependence in n(x)
from −3σ to +3σ , then the i th W of each distribution, is indirectly captured in (12) as VT (x) depends on VDS . Here,
i.e., Wmin (i ), Wavg (i ), and Wmax (i ), would have the same we also ignore the dependence of mobility on fin width as the
probability. We assume that they belong to the i th fin. carrier concentration is an exponential function and, therefore,
1356 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 63, NO. 3, MARCH 2016

would be dominant. Thus, (11) can be rewritten as


⎧ ⎫
⎨ Aqμλ e qV ηkT ⎬
gs
dV x
I = dV x =
⎩ dx qVT (x)
⎭ R(x)
e ηkT
⎧ ⎫
T (x) ⎬
⎨ d x e qVηkT dV x
R(x) = = (13)
⎩ qV gs
⎭ I
Aqμλe ηkT
where R(x) is the resistance at position x. Summing R(x)
from 0 to L G would give net effective resistance (Reff )


LG Vds
dV x Vds
R(x) = = = Reff
I I
x=0 V =0
⎧ qVT (x)
⎫ ⎧ ⎫
LG ⎨ ηkT ⎬ ⎨
qVT −eff
ηkT ⎬
dxe LGe Fig. 5. Q–Q plots of VT −sat distribution obtained using the above-
Reff = = (14) mentioned formalism for different Wfin s and s at L G = 15 nm and
⎩ qV gs
⎭ ⎩ qV gs
⎭ 3 = 2 nm [1], [15]. Deviation from the dashed lines represents the
x=0 Aqμλe ηkT Aqμλe ηkT
non-Gaussian nature of FER-induced VT variability.
where VT −eff is the effective threshold voltage. The summation
from 0 to L G is valid both for low and high VDS . For low
VDS case, channel extends from source to drain. Thus, all the
values of W (x) (for all x) contribute. So equally summing
over from 0 to L G over all the W ’s is correct. For high
VDS , channel extends from source to pinchoff point, unlike
low VDS case. Therefore, only the W (x) in the channel (from
source to pinchoff point) should be considered in the resistance
summation. However, different sections of the fin, i.e., W (x),
can be arranged in many equally likely ways, while ensuring
Fig. 6. (a) Contour plot of coefficient of determination (R 2 ) between
Wmin , Wmax , and Wavg of the fin are preserved. All such Q–Q plots generated using the model and stochastic simulations at
fins belong to the same probability quantile. Thus, to find L G = 15 nm for various Wfin s and /L G s. (b) R 2 contour plot at
effective VT at this particular probability quantile, summing Wfin = 5 nm for various L G s and s.
over channel (from source to pinchoff) in a large sample space
of fins with different W (x) arrangements may be replaced by where pi is the probability of occurrence of Wmin (i ), drawn
a simple, position-independent summation of all W (x) from a Gaussian probability distribution. The standard devia-
(from 0 to L G ) of a linearly varying fin, as presented in (14). tion of threshold voltage (σ VT ) is obtained similarly. Both
Thus, the final expression of VT −eff is given by VT −sat and VT −lin distribution can be obtained using this
⎧ ⎫ model; however, for brevity, the results of only VT −sat are
ηkT ⎨ ⎬
LG presented in the discussion to follow.
T (x)
d x qVηkT
VT −eff = ln e (15)
q ⎩ LG ⎭
x=0 IV. R ESULTS AND D ISCUSSION
where VT (x) is VT at x, which depends upon the local fin A. VT −sat Distribution’s Q–Q Plot Comparison
width W (x). Using this fin width, VT (x) is obtained from Using the proposed model, the results of VT −sat distribu-
SC curve at VDS = VDD for VT −sat (Fig. 1) or using an tion’s Q–Q plot are shown and compared against stochastic
analytical expression for VT versus uniform fin width [7]. simulations (200 samples) in Fig. 5 for various fin widths
VT −eff for the i th fin depends upon W i (x) which in turn and s, at a fixed channel length of 15 nm. As can be seen,
varies uniquely with Wmin (i ) as the model is able to predict the VT −sat distribution very well
for a wide range of Wfin s and s. As a metric of compar-
VT −eff (i ) = VT (Wmin (i )). (16) ison between VT −sat distributions obtained using model and
stochastic simulations, the coefficient of determination (R 2 )
As Wmin has a Gaussian distribution √ with mean was calculated and is shown in Fig. 6. Fig. 6(a) shows the
as μWmin (6) and standard deviation as 2 [6], from the R 2 value for the Q–Q plots shown in Fig. 5, i.e., at fixed
distribution of Wmin , the VT distribution due to FER can be L G = 15 nm, while Fig. 6(b) shows it for a fixed Wfin = 5 nm
obtained using (16). (Q–Q plots not shown). R 2 values with μ of 93% and
In order to find the mean of threshold voltage (μVT ), the σ of 13% are obtained for Fig. 6(a). Similarly, for Fig. 6(b),
following relation is used: R 2 values with μ = 94% and σ = 10% are obtained. In the

n region of interest, ( ranging from 15 to 50 nm [13]), the
μVT = pi VT (Wmin (i )) (17) R 2 value improves to μ = 98% and σ = 1.5% for Fig. 6(a)
i=1 and μ = 96% and σ = 5% for Fig. 6(b). Thus, the model
MITTAL et al.: ANALYTICAL MODEL TO ESTIMATE FinFET’s VT DISTRIBUTION DUE TO FER 1357

Fig. 7. (a) μVT −sat obtained using the model compared against stochastic
simulations data for various fin widths at L G = 15 nm. A very close match is Fig. 9. (a) μVT −sat and (b) σ VT −sat versus /L G for various estimation
obtained. (b) μVT −sat obtained using the model compared against stochastic schemes compared against stochastic simulations data at L G = 15 nm and
simulations for various L G s at Wfin = 5 nm. Wfin = 3.5 nm. Wavg -based estimation [2] matches well at higher /L G
ratios, while Wmin -based estimation [6] is able to capture the trend. The model
proposed in this paper provides a good match to stochastic data at all /L G
along with capturing the trend. VT −sat at L G = 15 nm and Wfin = 3.5 nm
is also shown in (a).

entire range of Wfin s shown in Fig. 8(a). This enhanced overall


error is due to mismatch at Wfin = 3.5 nm and  = 5 nm.
This mismatch is expected as discussed earlier. However, this
large error reduces to 2.4 ± 2.6 mV in the regions of interest
( varying from 15 to 50 nm [13]). At Wfin = 5 nm and
for various channel lengths of 12, 15, and 20 nm, σ VT −sat is
Fig. 8. (a) σ VT −sat obtained using the model compared against stochastic compared in Fig. 8(b). The model’s σ VT −sat matches closely
simulations data for various fin widths at L G = 15 nm. Model matches the with the stochastic simulations data. A low error of 4 ± 3.7 mV
trend of σ VT −sat with  along with closely matching the values. (b) σ VT −sat
obtained using the model compared against stochastic simulations data for is obtained for the entire surface. The error further reduces to
various L G s at Wfin = 5 nm. 2.9 ± 2.9 mV in the regions of interest. Furthermore, it might
appear that the mismatch at L G = 12 nm is large; however,
is capable of estimating the entire VT −sat distribution quite the percentage error at all L G s is not much different. It is just
accurately for a range of Wfin s, L G s, and s. The model’s because of larger σ VT −sat at L G = 12 nm that the mismatch
inaccuracy at  = 5 nm (/L G  1) is consistent with appears slightly larger. Thus, the proposed model is capable
the assumptions used in the derivation of Wmin and Wmax of capturing the standard deviation of threshold voltage quite
distributions. The assumption that the average distribution is accurately.
an exponential decay implies that the random perturbations are
infrequent which is valid for large correlation lengths. Hence, V. B ENCHMARKING AND D ISCUSSION
the model’s inaccuracy at small /L G is expected. Therefore,
A. Comparison With Average W f in (Wavg ) and
we observe slight mismatch at /L G  1, which is beyond
Minimum W f in (Wmin )-Based Estimations
our range of interest.
In order to benchmark the present model against models
B. μVT −sat Comparison present in [2] and [6], μVT −sat and σ VT −sat versus /L G at
From the Q–Q distribution, μVT −sat is obtained using (17) L G = 15 nm and Wfin = 3.5 nm is shown in Fig. 9(a) and (b),
and is compared against stochastic simulations (200 samples) respectively. Fig. 9(a) and (b) shows that Wavg -based
in Fig. 7(a) for various fin widths at a fixed channel length of estimation (where the average Wfin of an i th fin (9) is
15 nm. As can be seen, model is able to predict the μVT −sat estimated, and a uniform fin of Wfin = Wavg provides VT
dependence on /L G . The error over the entire surface shown for the i th FinFET; W i (x) = Wavg (i ) [2]) matches the
is 4.7 ± 7.5 mV. The error reduces to 3.4 ± 1.6 mV for the stochastic data well for larger values of /L G . However, it
practical ranges of  (in between 15 and 50 nm [13]). μVT −sat is insensitive to  and, hence, unable to capture the effect
obtained using the model at Wfin = 5 nm and various L G s is of , which becomes significant at lower s. On the other
compared in Fig. 7(b). Here, also match is appreciable with a hand, Wmin -based estimations (W i (x) = Wmin (i ) [6]) provide
low error of 2.5 ± 1.9 mV. The error reduces to 2.3 ± 0.9 mV a  dependence that captures the trend. However, it produces
for the practical ranges of  (in between 15 and 50 nm [13]). quantitative mismatch in μVT −sat and σ VT −sat estimations.
The present model provides sensitivity to  that enables
C. σ VT −sat Comparison improved prediction of trends. This model also provides an
improved quantitative matching to the stochastic simulations
From the complete VT −sat distribution obtained in Section data for all values of /L G compared with the other models.
IV–A., the σ value of entire distribution is computed. Fig. 8(a)
compares σ VT −sat for various Wfin s at L G = 15 nm. As can
be seen, the model’s estimate of σ VT −sat matches well with B. μVT −sat Dependence on Variability
stochastic data for a range of /L G . For lower Wfin and lower A general assumption is that the nominal device’s per-
/L G , σ VT −sat increases. 8.5±23-mV error is obtained for the formance (device with nominal dimensions, no variations)
1358 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 63, NO. 3, MARCH 2016

is the average performance. We show in Fig. 7(a) that this R EFERENCES


assumption holds true for the cases of high /L G (i.e., the [1] X. Wang, A. R. Brown, B. Cheng, and A. Asenov, “Statistical variability
fins are fairly uniform) and large Wfin [where QC is weak and reliability in nanoscale FinFETs,” in IEDM Tech. Dig., 2011,
enabling a fairly linear device response (such as VT and ION ) pp. 5.4.1–5.4.4.
[2] K. Patel, “Intrinsic and systematic variability in nanometer CMOS tech-
to device dimensions]. However, for the cases of small Wfin nologies,” Dept. Elect. Eng. Comput. Sci., Univ. California, Berkeley,
(where there is a QC-based large nonlinearity) and small Berkeley, CA, USA, Tech. Rep. UCB/EECS-2010-181, 2010.
/L G (which produces local confinement due to high spatial [3] S. Mittal et al., “Epitaxially defined FinFET: Variability resistant and
high-performance technology,” IEEE Trans. Electron Devices, vol. 61,
frequency-based undulations in fin edge), the mean of the no. 8, pp. 2711–2718, Aug. 2014.
distribution gets shifted from the nominal. Fig. 9(a) shows [4] Q. Cheng, J. You, and Y. Chen, “A generalized model to predict
that VT −sat of the nominal device (magenta line) can deviate fin-width roughness induced FinFET device variability using the bound-
significantly from the average performance (blue line) due to ary perturbation method,” Proc. SPIE, vol. 9053, p. 90530U, Mar. 2014.
[5] Y. Ban, S. Sundareswaran, R. Panda, and D. Z. Pan, “Electrical impact
FER at lower Wfin . Thus, for scaled Wfin s, the /L G effect on of line-edge roughness on sub-45 nm node standard cell,” Proc. SPIE,
μVT −sat gets enhanced. This effect may be further aggravated vol. 7275, pp. 727518-1–727518-10, Mar. 2009.
by transition to Ge FinFETs, where the effective mass is lower [6] S. N. Chinta, S. Mittal, P. Debashis, and U. Ganguly, “A FinFET LER
VT variability estimation scheme with 300× efficiency improvement,” in
causing stronger QC at larger Wfin s [16]. This motivates a Proc. Int. Conf. Simulation Semiconductor Process. Devices (SISPAD),
reduction of 3 in LER. Thus, this paper shows that even Sep. 2014, pp. 277–280.
μVT −sat of FinFET (not just σ VT −sat ) is affected by variability [7] T. Poiroux et al., “Multiple gate devices: Advantages and challenges,”
which will be critical to scaled FinFETs. Microelectron. Eng., vol. 80, pp. 378–385, Jun. 2005.
[8] (2011). International Technology Roadmap for Semiconductors (ITRS)
2011. [Online]. Available: http://www.itrs.net/reports.html
VI. C ONCLUSION [9] Sentaurus TCAD Design Suite. [Online]. Available: http://www.
synopsys.com
In this paper, we present an analytical model to estimate [10] M. A. Alam. Nanostructured Electronic Devices: Percolation and Reli-
the VT −sat variability due to FER in advanced FinFETs. The ability. [Online]. Available: https://nanohub.org/resources/7168
model consists of two formulations: 1) an analytical model for [11] C.-Y. Chang et al., “A 25-nm gate-length FinFet transistor module for
the estimation of μWmin and μWmax and 2) tapered fin perco- 32 nm node,” in IEDM Tech. Dig., 2009, pp. 1–4.
[12] T. Krishnamohan, “Band-engineering of novel channel materials for
lation model to estimate VT distribution due to FER variability. high performance nanoscale MOSFETs,” in Proc. Int. Conf. Simulation
The model can compute the entire VT −sat distribution quite Semiconductor Process. Devices (SISPAD), 2008, pp. 97–100.
accurately based on device parameters (e.g., L G and Wfin ) [13] A. Asenov, S. Kaya, and A. R. Brown, “Intrinsic parameter fluctuations
in decananometer MOSFETs introduced by gate line edge roughness,”
and FER parameters (e.g.,  and ). This paper demonstrates IEEE Trans. Electron Devices, vol. 50, no. 5, pp. 1254–1260, May 2003.
that even μVT −sat of FinFET (not just σ VT −sat ) is affected [14] L. Sjogren. Brownian Motion: Langevin Equation. [Online]. Available:
by variability and presents a model that accurately captures http://physics.gu.se/~frtbm/joomla/media/mydocs/LennartSjogren/kap6.pdf
this effect. Thus, this completely analytical formulation can [15] S. Mittal, A. S. Shekhawat, and U. Ganguly, “FinFET scaling rule
based on variability considerations,” in Proc. 73rd Annu. Device Res.
enable lithography engineers to compare the performance of Conf. (DRC), Jun. 2015, pp. 127–128.
conventional versus novel lithography techniques based on [16] S. Mittal et al., “Epitaxial rare earth oxide (EOx) FinFET: A variability-
VT variability metrics. Similarly, the model can be imple- resistant Ge FinFET architecture with multi VT ,” in Proc. 72nd Annu.
Device Res. Conf., Jun. 2014, pp. 97–98.
mented in SPICE to enable FER VT −sat variability-dependent
circuit level analysis of scaled FinFETs.

ACKNOWLEDGMENT
The authors would like to thank Prof. M. A. Alam from
Purdue University for the useful discussions. S. Mittal would
like to thank V. Sangwan for the helpful discussions. Authors’ photographs and biographies not available at the time of publication.

You might also like