Академический Документы
Профессиональный Документы
Культура Документы
2, FEBRUARY 2011
Abstract—In this paper, a 3-D analytical model is proposed to circuit parameters that are important for the design of digi-
capture the threshold voltage, surface potential, and electric field tal and analog circuits [25]–[28]. For instance, the transistor
variations induced by random dopant fluctuations in the channel threshold voltage Vth affects the delay and the on/off cur-
region of metal–oxide–semiconductor field-effect transistors. The
3-D model treats the effect of each dopant separately and is rent of a digital circuit, and the vertical electric field affects
based on fundamental laws of physics. The proposed approach the mobility of the carriers inside the channel and, thus, the
enables determination of transistor threshold voltage variations transconductance.
with both very low computational cost and high accuracy. Using Variations of the threshold voltage have been ruminated upon
the developed model, we performed statistical analysis, simulating not only experimentally [24] but also numerically through the
more than 100 000 transistor samples. Interestingly, the results
showed that, although the distribution of the threshold voltage use of 3-D atomistic [1]–[3] or quantum-level simulators [30],
for large-channel transistors is Gaussian, for scaled transistors, [31]. The main disadvantage of these models is that they require
it is non-Gaussian. Furthermore, the proposed model predicts large computational effort and/or use complex techniques for
known formulas, which are proven for 1-D analysis and large threshold voltage determination [12]–[16]. On the other hand,
transistors, simply by setting the appropriate transistor size. As there exist analytical models [9], [10], but these models are not
a consequence, this model is a logical extension of the theory of
large transistors to nanoscaled devices. very accurate, because they fail to predict the non-Gaussian
Vth distribution for scaled technologies. Hence, average and
Index Terms—Electric field, random dopant placement (RDF), standard deviation is not enough statistical information to de-
surface potential, Vth variations, 3-D analytical model.
scribe the statistics of Vth . Nonzero skewness and kurtosis
throw more light on the real Vth distribution, and our proposed
I. I NTRODUCTION
model provides this information. In addition, in this framework,
ionized dopants away from the surface and close to the edge of
the depletion region is not as significant as compared to those
close to the oxide interface. 2) On the average, the dopants
that are ionized and outside the depletion region (as previously
defined) are canceled out by those that are neutral and inside
the depletion region. These two observations directly lead to the
approximation of the maximum depletion width, as expressed
by the following equation, that is shown for 1-D uniform
channel region [32]–[34]:
Wdm = 4εsi kT ln(NA /ni )/q 2 NA (1)
Fig. 2. Two random patterns from the channel region of transistors of the same technology node (22 nm) showing the random number and position of dopants.
The corresponding 3-D potential and electric field at the surface are illustrated in sequence.
process/doping method and the doping profile used (i.e., uni- in the next two sections, which are necessary to build our
form, step, retrograde, and Gaussian) [34]. For example, if a model. These variables are categorized as quantitative variables
Gaussian doping profile is assumed, then the concentration and (number of dopants and effective doping concentration) and
the corresponding probabilities si are given by the equations electrical variables (surface potential, electric field at the sur-
face, and threshold voltage). All these variables are represented
N0 − (z−z2c )2 a3 N0 − (zi −z2c )2
NA = √ e 2σ si = √ e 2σ as “leaves” in the flowchart (Fig. 1) to designate that they are
2πσ 2πσ the outputs of the model.
where zc is the center of the Gaussian doping profile, σ is its
deviation from the center, and a is the lattice constant (the min- D. Modeling Quantitative Variables
imum possible distance between two dopants). In Section IV,
it becomes more obvious why we have to multiply NA (x, y, z) The quantitative variables give intuitive information about
by the term a3 to determine si . the number of dopants in the channel region, which is an
Note that the maximum depletion width Wdm given by important and critical quantity for nanoscaled transistors.
(1) is not valid for nonuniform doping profile, but generally, 1) Number of Dopants Dp : The first random variable Dp
there are analytical expressions for Wdm for some common denotes the “total number of ionized dopants in the channel
profiles. Furthermore, for doping profiles where it is difficult region” and is defined as the sum of Xi over all the N mesh
to analytically determine Wdm , a technology computer-aided vertices
design device simulator (Medici, Taurus, Sentaurus) can be
used to extract Wdm before the analysis starts.
N
Dp = Xi . (3)
In order to understand the concept of our framework, let
i=1
us consider an example [Fig. 2(a) and (d)] of two differ-
ent transistor instances at the same technology node (W = 2) Effective Doping Concentration Cp : The “effective dop-
L = 22 nm), and let us assume uniform doping concentra- ing concentration” is defined as the ratio of the total number
tion (NA = 1019 cm−3 ). These channel region instances are of ionized dopants in the channel region to the volume of this
generated by performing two iterations, as shown in Fig. 1. region. It is evident that the effective doping concentration of
In addition, we define a vector {X} of the aforementioned each transistor is also a random variable that is related to the
random variables Xi : {X} = { X1 X2 · · · XN }T . In the predefined random variable Dp by
subsequent sections, this vector will be very useful for the
formulation of the problem.
1
N
Dp
Having discussed the tools that are required to build a 3-D Cp = = Xi . (4)
analytical model for RDF, various random variables are defined W · L · Wdm N a3 i=1
PANAGOPOULOS AND ROY: ANALYTICAL MODEL FOR RDF-INDUCED THRESHOLD VOLTAGE VARIATIONS 395
4 2
c1 = (Vbi +Vds )− (1−4/π 2 )(1+6Tox /Wdm )Φ(x, y).
π π
UB is an order of magnitude smaller than US and UD and has
been ignored in this analysis. The complete derivation is given
Fig. 4. Two-dimensional discrete lattice at Si-SiO2 interface used for the
in [34, Appendix A]. conversion of the electrical variables to matrix form.
2) Electric Field EZ : The electric field is a vector and
consists of three components (x, y, and z directions). Even maximum surface potential across the “lane.” Such formulation
though all these components can be calculated from the pro- is described mathematically as
posed model, the most important for the determination of the
threshold voltage is the component vertical to the Si-SiO2 Yi (y) = max (Vt (x, y)) (12)
x=(0...L)
interface (z-component). Hence, the electric field is given by
the following equation: for all y along the channel width. Finally, since we have
assumed that all the “lanes” have the same width dy, the
N
transistor’s threshold voltage is defined as the average of all the
EZ (x, y) = EZ,i (x, y) − EZ,i (x, y) (9) lane threshold voltages [23]
i=1
W W
where EZ,i (x, y) and EZ,i (x, y) are defined as 1 1 εsi EZ
Vth = Yi (y)dy = max Ψ+ dy. (13)
W W x=0...L Cox
qzi Xi qzi Xi 0 0
EZ,i (x, y) = EZ,i (x, y) = . (10)
4πεSi Ri3 4πεSi Ri3 Equation (12) is the analytical expression for the threshold
voltage. Alternatively, the threshold voltage can be defined by
3) Threshold Voltage Vth : In the final step of the proposed averaging all over the channel. In this case, the definition of Vth
framework, we define the threshold voltage Vth . The definition is given by
of the threshold voltage is not unique; it can be extracted from
current–voltage characteristics [8], [9], or it can be computed W L
1 εsi EZ (x, y)
using percolation paths, which connect the source and the drain Vth,lin = Ψ(x, y) + dy. (14)
terminals [5], [8]. WL Cox
0 0
Before discussing the modeling of Vth , we have to define
the random variable Vt (x, y), which represents the gate voltage This definition can be used when Vds is small (i.e., when
that should be applied in order to invert the channel locally. linear threshold voltage Vth,lin is of interest). On the other
Since the electrostatic potential Ψ and the electric field EZ at hand, the proposed framework does not limit us using other
the surface are known [(6) and (9)], the local Vt can be defined definitions that potentially improve the accuracy. Equation (12)
as the voltage drop across the bulk plus the voltage drop across is more general than (14) but more complex for statistical
the oxide analysis. Note that, for large transistors, (12) and (14) become
equivalent; thus, either of them can be used. Note that all the
εsi EZ (x, y) equations defined in this section are functions of the input
Vt (x, y) = Ψ + Vox = Φ(x, y) + U + (11)
Cox parameters W , L, NA , and Tox . Now, we can present the
random variables previously defined in discrete form.
where Cox = ox /Tox . In this paper, we simplify the definition
of threshold voltage by assuming that the electrons move from
drain to source contact laterally. This assumption reduces fur- III. S OLUTIONS IN D ISCRETE F ORM
ther the complexity and has minor impact on the accuracy of In this section, we discretize the electrical variables in order
the model, as can be observed from the obtained results shown to perform easier statistical analysis and obtain numerical solu-
later. Such simplification is allowed by dividing the channel in tions. In order to discretize Φ and EZ , we use a 2-D discrete
parallel “lanes” and assuming that electrons do not change their lattice similar to the one shown in Fig. 4. In this case, the
“lane” during their transport from source to drain contacts [23]. discrete form of Φ and EZ can be written in matrix form as
This is equivalent to a 1-D channel (Landauer model), in which
the threshold voltage is easily determined by just taking the {Φ} = [A] · {X} {EZ } = [B] · {X} (15)
PANAGOPOULOS AND ROY: ANALYTICAL MODEL FOR RDF-INDUCED THRESHOLD VOLTAGE VARIATIONS 397
where Φ(i,j) and EZ(i,j) are the potential and electric fields on depends on the input parameters W , L, and NA . For instance,
the Si-SiO2 interface around the mesh node (i, j), respectively. an increase in N translates to an increase in either the transistor
Note that K and L are the numbers of discrete lattice points size (W and/or L) or the depletion width (Wdm ), which, in
in the x and y directions, respectively (Fig. 4). In addition, turn, depends on input NA . Since the random variables Xi
this representation becomes exact (with the continuous case are independent, the distribution of Dp follows the Binomial
previously described) only in the limit, where the grid distances distribution, which approaches the Poisson distribution as N
a and b approach zero. However, as long as a and b are increases. Moreover, Dp approaches the normal distribution
smaller than the spatial scale on which Φ and EZ varies, this (Dp ∼ Normal(μDp , σDp )) as N becomes very large, accord-
representation is reasonably accurate. The size of the matrices ing to central limit theorem (CLT) [35]. Note that μDp and σDp
[A] and [B] is (KL)xN , and their corresponding elements are increase with the enlargement of the channel region (W , L,
of the form and Wdm ), whereas SDp decreases (approaches to zero) as the
q channel region increases. This verifies the general intuition that
α(κ,λ)i = (1/Ri (xκ , yλ ) − 1/Ri (xκ , yλ )) (16) large transistors are uniform.
4πεSi
qzi
b(κ,λ)i = 1/Ri3 (xκ , yλ ) − 1/Ri3 (xκ , yλ ) . (17)
4πεSi B. Effective Doping Concentration Cp
Using the preceding formulation for Φ and EZ , we can define Following the same arguments used for the random variable
the discrete form of Vt that will allow us to calculate the Vth of Dp , the random variable Cp also approaches the normal distri-
a transistor numerically. Substituting (15) into (11), the local bution (Cp ∼ Normal(μCp , σCp )) as N increases (CLT). The
threshold voltage {Vt } can be written as a function of the mean value and the standard deviation are given by
random vector {X} as √
N
1 s st
εSi εSi μCp = E[Xi ] = 3 = NA σCp = √ (21)
{Vt } = {Ψ} + {EZ } = U + [A] + [B] {X}. N a3 a 3
a N.
Cox Cox i=1
(18) √
The spread of Cp is SCp = t/N s ∝ 1/ N . With the aid of
Finally, the {Vt } vector is utilized for the determination of (21), the probability s of a Bernoulli random variable Xi is
Vth for low [(14)] and high Vds [(13)], i.e., determined and is a function of the doping concentration NA ,
i.e., s = a3 NA . In contrast to the mean of Dp , the mean of Cp
1 1
K Λ Λ
Vth,lin = Vt Vth = max Vt(i,j) . (19) is constant and is equal to NA , which is the nominal doping
KΛ i=1 j=1 Λ j=1 i=1...K concentration. We also observe that both σCp and SCp approach
to zero as N becomes very large.
Note that all the variables defined in this section depends on the
parameter matrices [A] and [B] and the oxide thickness Tox .
For a specific technology node, [A] and [B] are calculated only C. Surface Potential Φ
once, and their size depends on the size of the device under Having discussed the statistics of the quantitative variables
analysis. A more accurate representation can be easily achieved Dp and Cp , we will next present the statistical analysis of the
by making the grid finer. In the next section, we elaborate the electrical variable Φ. Since Φ is a vector of random variables,
statistical analysis of the random variables. its mean and variance are also expected to be vectors with the
elements, mean, and variance of Φ(i,j) , respectively. Using the
IV. S TATISTICAL A NALYSIS definition of Φ [(15)], we derive the mean of Φ as
T
In this section, we calculate the mean, the variance, and
N
N
the distributions of the random variables previously defined. In E[Φ] = s α(1,1)r · · · α(K,L)r ≡ [MΦ ]. (22)
addition, a discussion on the theoretical results is presented. r=1 r=1
The joint pdf for the potential vector using the known constant
matrices ME and ΣE is given approximately by the multi-
variablenormal pdf [35]: f = exp(−(1/2)( − ME )T Σ−1 E ( −
ME ))/ (2π)K·L |ΣE |. Fig. 5. Statistics of the quantitative random variables. (a) Number of dopants
Note that the electric fields at neighboring grid points at the in the channel Dp . (b) Effective doping concentration Cp obtained for different
Si-SiO2 interface are correlated. This correlation is one of the values of NA and transistor size.
most important advantages of our model and is necessary for
accurate determination of the electric field and the threshold V. R ESULTS AND D ISCUSSIONS
voltage. In order to show the efficacy of the proposed model, we
performed 100 000 Monte Carlo simulations on various-sized
E. Threshold Voltage Vth transistors with different doping profiles and obtained distribu-
tions and statistical data for a number of parameters.
Finally, we derive expressions for the statistics of the thresh- In Fig. 5, the mean and the standard deviation of Dp and Cp
old voltage Vth . Using (19), the mean of the threshold voltage with different values of W and L are demonstrated. Fig. 5(a)
is derived as follows: verifies the linear relationship of μDp with W and the square
K Λ root of σDp with W , as captured by (20). We can also observe
1 εSi
E[Vth ] = U + MΦ + ME ≡ μVth . from Fig. 5(b) that μCp is constant and is equal to NA while
√
KΛ i=1 j=1 Cox
σCp is proportional to 1/ W L, as predicted by (21). Since,
(28) for very large transistors, σCp decreases and tends to zero, we
may conclude that the effective doping concentration is given
μVth is the average value of threshold voltage and depends on
by a single value NA (as in the 1-D analysis). Furthermore, we
the matrices MΦ and ME . Similarly, applying the covariance
observe that transistors with higher doping concentration suffer
operator on (19), we obtain the standard deviation of Vth
from larger variations in the number of dopants.
Λ The distributions of the threshold voltage for different tran-
s · t K
ε2 2εSi
σVth =
sistor sizes are shown in Fig. 6. We observe that the variations
AAT + Si 2
BB T
+ AB T
.
KΛ i=1 j=1 Cox Cox of the Vth due to RDF increase as the transistor size decreases.
(29) In addition, our model predicts the skew distribution of Vth
that has been observed by the authors in [22] and [29] and is
Note that σVth depends on the matrices A and B. attributed to the small number of dopants in the channel region.
PANAGOPOULOS AND ROY: ANALYTICAL MODEL FOR RDF-INDUCED THRESHOLD VOLTAGE VARIATIONS 399
Fig. 7. Logarithmic plot of the threshold distribution for short- and long-
channel transistors obtained from the proposed model for a sample set of
100 000 MOSFETs with Tox = 3 nm and NA = 5 × 1018 cm−3 .
Fig. 12. Three-dimensional band diagram for low and high Vth of two 20-nm transistors is illustrated in sequence, demonstrating their correlation.
Fig. 13. Comparison of the Vth for low and high Vds obtained from the proposed model for a sample set of 100 000 MOSFETs. (a) Standard deviation. (b)
Skewness. (c) Kurtosis.
the general trends of threshold fluctuations due to RDF, it fails case (1-D analysis). This convergence is achieved by increasing
to predict quantitatively the magnitude of these fluctuations and the transistor size to infinity. For mathematical simplicity and
higher order statistical moments. On the other hand, the 3-D without loss of generality, we suppose that the shape of the
“atomistic simulator” predicts quantitatively the magnitude of transistor channel is cylindrical (Fig. 15).
the fluctuations at the cost of extended simulation time. Starting from (7) and applying the average operator on it, we
Finally, the proposed model succeeds in capturing the 3-D obtain
effects, and it can be used for predicting electrical properties
NA
N
of devices in the nanometer and micrometer regime. It is a
E[Φi ] = air
complete and unified model that is able to capture the effect of a3 r=1
RDF for all the technology nodes. This is achieved mainly by ⎡W
effectively calculating the electrostatic potential, electric field, ρ 2π dm
∼ qNA ⎣ rdrdϕdz
and threshold voltage. =
4πεSi r2 + (z + b)2
0 0 0
⎤
A PPENDIX A Tox
rdrdϕdz ⎦.
In this section, we show that the variables Φ and EZ converge −
r2 + (z + b)2
to a value that is the same as that we know for the uniform −Wdm −Tox
402 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 58, NO. 2, FEBRUARY 2011
[28] Y. Li, C.-H. Hwang, T.-C. Yeh, and T.-Y. Li, “Large-scale atomistic ap- Kaushik Roy (S’83–M’83–SM’95–F’02) received
proach to random-dopant-induced characteristic variability in nanoscale the B.Tech. degree in electronics and electrical com-
CMOS digital and high-frequency integrated circuits,” in Proc. ICCAD, munications engineering from the Indian Institute of
Nov. 2008, pp. 278–285. Technology, Kharagpur, India, and the Ph.D. degree
[29] C. Millar, D. Reid, G. Roy, S. Roy, and A. Asenov, “Accurate statisti- from the electrical and computer engineering from
cal description of random dopant-induced threshold voltage variability,” the University of Illinois, Urbana, in 1990.
IEEE Trans. Electron Device Lett., vol. 29, no. 8, pp. 946–948, Aug. 2008. He was with the Semiconductor Process and De-
[30] X.-W. Jiang, H.-X. Deng, J.-W. Luo, S.-S. Li, and L.-W. Wang, “A sign Center, Texas Instruments Incorporated, Dallas,
fully three-dimensional atomistic quantum mechanical study on random where he worked on field-programmable gate array
dopant-induced effects in 25-nm MOSFETs,” IEEE Trans. Electron De- architecture development and low-power circuit de-
vices, vol. 55, no. 7, pp. 1720–1726, Jul. 2008. sign. He joined the electrical and computer engineer-
[31] A. Asenov, G. Slavcheva, A. R. Brown, J. H. Davies, and S. Saini, ing faculty at Purdue University, West Lafayette, IN, in 1993, where he is
“Increase in the random dopant induced threshold fluctuations and low- currently a Professor and holds the Roscoe H. George Chair of Electrical &
ering in sub-100 nm MOSFETs due to quantum effects: A 3-D density- Computer Engineering. He supervised 50 Ph.D. students. He has published
gradient simulation study,” IEEE Trans. Electron Devices, vol. 48, no. 4, more than 500 papers in refereed journals and conference proceedings. He is
pp. 722–729, Apr. 2001. a coauthor of two books on low power CMOS VLSI design (John Wiley &
[32] S. Sze and K. Ng, Physics of Semiconductor Devices, 3rd ed. Hoboken, McGraw Hill). He is the holder of 15 patents. His research interests include
NJ: Wiley, 2008. spintronics, VLSI design/computer-aided design for nanoscale silicon and
[33] R. F. Pierret, Fundamental of Solid State Devices, 2nd ed: Addition- nonsilicon technologies, low-power electronics for portable computing and
Wesley, 1996. wireless communications, VLSI testing and verification, and reconfigurable
[34] Y. Taur and T. Ning, Fundamentals of Modern VLSI Devices. computing.
Cambridge, U.K.: Cambridge Univ. Press, 1998. Dr. Roy was a Research Visionary Board Member of Motorola Labo-
[35] A. Papoulis and S. U. Pillai, Probability, Random Variables and Stochastic ratories in 2002 and held the M.K. Gandhi Distinguished Visiting faculty
Processes, 4th ed. New York: McGraw-Hill, 2002. at Indian Institute of Technology (Bombay). He has been in the Editorial
[36] C. A. Balanis, Advanced Engineering Electromagnetics. Hoboken, NJ: Board of IEEE Design and Test, IEEE TRANSACTIONS ON CIRCUITS and
Wiley, 1989. SYSTEMS, and IEEE TRANSACTIONS ON VLSI SYSTEMS. He was a Guest
[37] S. Ramo, J. R. Whinnery, and T. van Duzer, Fields and Waves in Commu- Editor for Special Issue on Low-Power VLSI in the IEEE Design and Test
nication Electronics, 2nd ed. Hoboken, NJ: Wiley, 1984. (1994) and IEEE TRANSACTIONS ON VLSI SYSTEMS (June 2000), IEEE
Proceedings–Computers and Digital Techniques (July 2002). He was the
recipient of the National Science Foundation Career Development Award in
Georgios Panagopoulos (S’05) received the 1995; IBM Faculty Partnership Award; ATT/Lucent Foundation Award; 2005
Diploma (with honors) degree in computer and SRC Technical Excellence Award; SRC Inventors Award; Purdue College of
communication engineering from the University of Engineering Research Excellence Award; Humboldt Research Award in 2010;
Thessaly, Volos, Greece, in 2006. He is working best paper awards at the 1997 International Test Conference, IEEE 2000
toward the Ph.D. degree in electrical and computer International Symposium on Quality of IC Design, 2003 IEEE Latin American
engineering with the School of Electrical and Test Workshop, 2003 IEEE Nano, 2004 IEEE International Conference on
Computer Engineering, Purdue University, West Computer Design, 2006 IEEE/ACM International Symposium on Low Power
Lafayette, IN. Electronics & Design, and 2005 IEEE Circuits; System Society Outstanding
From 2006 to 2007, he was a Research Assis- Young Author Award (Chris Kim); and the 2006 IEEE TRANSACTIONS ON
tant designing analog circuits with the University VLSI SYSTEMS Best Paper Award. He is a Purdue University Faculty Scholar.
of Thessaly, Greece. His current research interests
include spintronics, device/circuit co-design for nanoscaled low-power and
reliable circuits, and design of process-variation-tolerant VLSI circuits and
systems.
Mr. Panagopoulos is the recipient of Academic Excellence Awards and
several scholarships during his studies.