Comparative Analysis of Memristor Models and Memories Design
Comparative Analysis of Memristor Models and Memories Design
Abstract: The advent of the memristor breaks the scaling limitations of MOS technology and prevails over emer-
ging semiconductor devices. In this paper, various memristor models including behaviour, spice, and experimental
are investigated and compared with the memristor’s characteristic equations and fingerprints. It has brought to light
that most memristor models need a window function to resolve boundary conditions. Various challenges of availed
window functions are discussed with matlab’s simulated results. Biolek’s window is a most acceptable window
function for the memristor, since it limits boundaries growth as well as sticking of states at boundaries. Simmons
tunnel model of a memristor is the most accepted model of a memristor till now. The memristor is exploited very
frequently in memory designing and became a prominent candidate for futuristic memories. Here, several memory
structures utilizing the memristor are discussed. It is seen that a memristor-transistor hybrid memory cell has fast
read/write and low power operations. Whereas, a 1T1R structure provides very simple, nanoscale, and non-volatile
memory that has capabilities to replace conventional Flash memories. Moreover, the memristor is frequently used
in SRAM cell structures to make them have non-volatile memory. This paper contributes various aspects and re-
cent developments in memristor based circuits, which can enhance the ongoing requirements of modern designing
criterion.
The principal objectives of the recent research are to re- 2. Memristor
duce required power, area, and cost of semiconductor devices
with high performance, efficiency, and reliability, so that the In 1971, Leon[14] noticed an absent link between charge
growth of semiconductor technology is maintained and (q) and magnetic flux (φ); in order to relate these parameters,
Moore’s law is kept alive. In order to fulfil these require- he propounded a new nonlinear passive device called the Mem-
ments, many emerging and novel devices have come into exist- ristor[14]. It was acknowledged as the fourth fundamental elec-
ence. Some of them were extensions of the complementary met- tric circuit element after the capacitor (C), resistor (R), and in-
al–oxide–semiconductor (CMOS) technology such as carbon ductor (L). In fact, the memristor completed the symmetry as
nanotube field effect transistors (CNT-FETs)[1], Nanowire shown in Fig. 1 and produced the sixth possible relation, be-
FETs[2], III–V channel replacement devices[3] and tunnel sides the five availed relations among four basic electrical para-
FETs[4], others were developed beyond CMOS technology meters viz. charge (q), voltage (v), current (i), and magnetic
such as negative gate-capacitance FET[5], spin FET[6], NEMS flux (φ), all of which are summarized in Table 1. Memristor
switch[7], and all-spin logic devices[8]. Along with this pro- means in short memory + resistor; it shows dissipative resist-
gress, numerous information storing techniques have evolved ance as well as non-volatile memory capability. The memris-
in the domain of volatile and non-volatile memory (NVM) tor memorizes the last resistance or amount of charge that
and it is found that NVM emerged with more pace than volat- passed through it, corresponding to the magnitude and direc-
ile memory. The emerging memory devices are ferroelectric tion of applied voltage across the device. So the memristor
random access memory (Fe-RAM)[9, 10], phase change can be described by the state dependent Ohm Law. A memris-
memory (PCM)[10, 11], spin-transfer torque random access tor is not merely a non-linear resistor, but is a resistor with
memory (STT RAM)[10, 12], and resistive switching based charge as a state variable[15].
memory[13] apart from static RAM (SRAM), dynamic RAM The fundamental equations, those that characterize the
(DRAM) and flash memory[8, 10]. The key challenges faced in memristor, are Eqs. (1) and (2) as follows
memory design are high endurance (i.e. ability to maintain
Roff, Ron ratio greater than unity), large retention (i.e. ability v = R (w) i, (1)
to retain information for a long time with large data storage), dw
low cost, fast read/write speed, and compatibility with the = i, (2)
dt
CMOS process. Resistive RAM (RRAM), in other words a
memristor, is able to accept these challenges and play a cru- where w is a state variable proportional to amounts of charge
† Corresponding author. Email: [email protected], [email protected]
Received 17 August 2017, revised manuscript received 23 October 2017 ©2018 Chinese Institute of Electronics
074006-1
J. Semicond. 2018, 39(7) Jeetendra Singh et al.
0.6
100 Hz
V (V) 500 Hz
Voltage 0.4
1000 Hz
0.2
Current (mA)
0
dv = Rdi dq = Cdv
−0.2
(Coulomb)
t
q (t) = ∫−∞ i(τ)dτ
Current
Charge
I (A)
q
−0.4
−0.6
∫ v(τ)dτ
dφ = Ldi dφ = Mdq
−1.0 −0.5 0 0.5 1.0
φ (t) = −∞
t
Voltage (V)
Table 1. Description of four fundamental circuit elements.
S.No Circuit element Inventor Invention year Relating parameter Relation Symbol
1 Capacitor Ewald George Von Kleist 1745 Charge (q) and voltage (v) dq = Cdv
2 Resistor George Simon Ohm 1827 Voltage (v) and current (i) dv = Rdi
3 Inductor Michael Faraday 1831 Magnetic flux (φ) and current dφ = Ldi
4 Memristor Leon Chua 1971 Magnetic flux (ϕ) and charge (q) dφ = Mdq
5 Basic law of electricity Charge and current t dq
q (t) = ∫ i (τ) dτ or i = dt
−∞
t dφ
6 Basic law of magnetism Flux and voltage φ (t) = ∫ v (τ) dτ or v = dt
−∞
074006-2
J. Semicond. 2018, 39(7) Jeetendra Singh et al.
is the incremental memristance in Ohm units.
Similarly, mathematically the memductance can be ob- Platinum
tained from a flux controlled memristor, q = gM (φ).
Now, differentiating Eq. (11) w.r.t time D
dq d
= g (φ) . (12) w)
RON (D
−
dt dt M W TiO2-x
It can also be written as Oxygen
Platinum vacancies
dq d dφ
= gM (φ) , (13)
dt dφ dt Fig. 3. (Color online) Linear model of a memristor. Conducting and
insulating layers are modeled as two series resistances.
or
dgM (φ)
i (t) = v (t) , (14) simulation of new memristor based circuits. Pershin et al.[41]
dφ
developed a spice model of a memristor by introducing a
i (t) = W (φ) v (t) , (15) threshold voltage, also they built models for a memcapacitor
and a memductor. In order to understand and reveal the com-
where
plex coupled electric-ionic conduction phenomena of the
dgM (φ)
W (φ) = , (16) memristor, several models of a memristor have evolved. In
dφ this section, the linear memristor model[20], non-linear mem-
is the incremental memductance of Siemens units. ristor model[32], the Simmons tunnel memristor model[34],
It can be seen from Eqs. (10) and (16) that if the mem- the Yakopcic neuromorphic memristor model[37], and the
ristance is independent of state variable charge and the mem- TEAM model[38] are emphasized, since these models are help-
ductance (inverse of memristance) is independent of state vari- ful to perceive excellently the conduction behaviour and cir-
able flux, then the memristor will become a resistor and the cuit designing aspects of the memristor.
memductance will become simple conductance. There is no dif-
ference in memristance and resistance if the memristor is lin- 3.1. Linear drift model
ear. After the memristor theory, a hypothetical periodic table
of 25 linear and non-linear circuit elements was presented by The linear model of a memristor was developed based on
Chua in 1980[18]. This table has a memcapacitor, meminduct- the fabricated structure of HP’s memristor in this model, a
or, and memristor along with three fundamental circuit ele- thin semiconductor metal oxide film (TiO2), consisting of
ments (resistor, capacitor and inductor). The memcapacitor two regions one deficient in oxygen (TiO2−x), say the doped
and meminductor were generalized but these are still to be real- region, and another with equally proportionate oxygen, say
ized[19]; however, the rest have only just been envisaged. the insulating region, was sandwiched between two elec-
trodes[20], as shown in Fig. 3. The oxygen deficient region
(conducting) named RON and the insulating region ROFF were
3. Mathematical modeling of memristor considered in a series.
The memristor was fabricated for the first time and then ( )
w (t) w (t)
modeled in HP’s lab in 2008 after about 38 years of its postula- RT = RON + ROFF 1 − , (17)
D D
tion[20]. After this successful endeavour, several memristor
structures were fabricated and modeled for its better realiza- where w(t) is the conductive width, and D is the total dimen-
tion[21]. In their nonlinear micro and nano scale TiO2 model, sion of the device. The boundaries of the doped and un-
Yang et al.[31] demonstrated that the switching polarity, doped regions, i.e. the values of RON and ROFF will vary ac-
rectification and conduction characteristics are determined by cording to the magnitude and polarity of the applied voltage
the distribution of oxygen vacancies at the platinum and TiO2 across the device and the time interval up to which the
interfaces of the fabricated memristor. Contemporarily, based voltage is applied. The applied voltage causes the oxygen va-
on the Simmons tunnel equations[32], Picket et al.[34] built a cancies to drift with dopants mobility μV. Thus, the voltage
mathematical model of a memristor, which fully correlates across the memristor is modeled as Eq. (18). The Blanc and
the electrical and dynamic behaviour of the memristor and Staebler relation (19)[22] of a dopant’s velocity and electric
also sets a framework to describe the physical process in- field is used to derive the state derivative relation (20). In or-
volved in its resistive switchings. Yokopcic et al.[37] presen- der to obtain the memristance final relation (22) for the con-
ted an improved memristor model, which is valuable in the dition ROFF ≫ RON, conductive thickness w(t) is inserted from
designing of neuromorphic systems. Kvantisky et al.[38] Eq. (20) into Eq. (18).
tried to simplify the complexity of the Simmons tunnel mod-
{ [ ]}
el and gave a generic model, which is viably simulates many w (t) w (t)
memristive devices. Moreover, Rak et al.[40] added another V (t) = RON + ROFF 1 − i (t) , (18)
D D
memristor micromodel, which is thoughtful in design and
074006-3
J. Semicond. 2018, 39(7) Jeetendra Singh et al.
1.2 8 0.5
(a) (b) 18
1.1 0.4
1.1 0.4
1.0 6 0.3 16
x = w/D
14
0.9 0.2
0.8 0.2
0.8 2 0.1 12
0.7 0.1
0.7 0 0
x = w/D x = w/D 10
0.6 Memristnce M(q) (kΩ) −0.1 Memristnce M(q) (kΩ)
0.6 −0.1
0.5 −2 −0.2 8
0 10 20 30 40 0 10 20 30 40
t (ms) t (ms)
Fig. 4. (Color online) Excess growth of conductive thickness and the corresponding memristance from their limiting value for RON = 100 Ω,
ROFF = 16 kΩ, winitial = 5 nm, µv = 10−14 m2/(V·s), at v(t) = 2 V sinusoidal and 100 Hz (a) in saturation {x→1, M(q)→100 Ω} and (b) in de-
pletion {x→0, M(q)→16 kΩ}.
Although the mobility of dopants and device length f (x) = 1 − (2x − 1)2p , (24)
show significant effects on the linear and nonlinear kinetics in-
side a memristor[23], it has been seen in the simulation results where x = w/D and ‘p’ is a positive integer, which controls
(see Fig. 4) of the linear model that the conducting width and the linearity and non-linearity of the window function and is
hence the memristance grow beyond their physical limits. known as the control parameter. This window does not re-
The physical limits of normalized conducing width and mem- solve the state sticking problem. Locking of the state is seen
ristance are taken to be 1 and 100 Ω in the saturation region in Fig. 5(a) in the linear memristor model by implementing
(a), whereas they are 0 and 16 kΩ in the depletion region (b). Jogelkar’s window function. To verify the locking state, a
It can be seen in the saturation region of Fig. 4(a) that the nor- voltage (±1 V sin2 (2πft)) is applied. Then it can be observed
malized conducting width or normalized state variable (x = that though at t = 25 s, the voltage changes its polarity but the
w/D) grows up to 1.2 and the memristance decreases −2 kΩ. state variable does not return back and remains stuck at 1.
In the depletion region of Fig. 4(b), ‘x’ reaches below ‘0’ and Further, this window function is generalized to consider flex-
the corresponding memristance extends above 16 kΩ, which ibility and non-linearity[26].
is not feasible. So this model requires a window function, c) Biolek’s window
which is multiplied with a derivative of state variable f(w, I) Biolek proposed a new window function (25) that re-
to limit these at the boundaries. There are many window func- solved the problem of the sticking state by including the mem-
tions available in the literature to achieve the nonlinear drift ristor current ‘i’ together with state variable ‘x’ and ‘p’[27], as
of dopants and the physical limit. shown in Fig. 5(b). It can be seen that the state variable var-
a) Strukov’s window ies according to the polarity of the applied voltage, i.e., it
Strukov’s window function (23)[20] restrict w, M(q) at the does not get stuck at any particular state. Here a positive cur-
boundaries but another problem arose of sticking states. It rent is correlated with the increasing doped width and negat-
gives a zero value at w = 0 and w = D, so the state derivative ive current correlated with decreasing the doped width.
or velocity becomes 0. Thus, even though voltage is reversed, ( )
the state variable does not return back. Means states are f (x) = 1 − x − sgn (−i) 2p , (25)
locked and this problem is known as the sticking states prob- where sgn(i) = 1, when i ≥ 0, sgn(i) = 0, when i ≤ 0.
lem. d) Prodromakis’s window
w (1 − w) This window function (26)[28] is well approximated by
f (w) = . (23) the non-linearity issue at the boundaries by utilizing the quad-
D2
ratic equation in its expression. State variable and mem-
074006-4
J. Semicond. 2018, 39(7) Jeetendra Singh et al.
0.5 0.5
Voltage (V)
Voltage (V)
0 0
−0.5 −0.5
Voltage Voltage
x x
−1.0 −1.0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
t (s) t (s)
Fig. 5. (Color online) (a) Boundary locked: state variable does not change its state as the polarity of applied voltage changes using Jogelkar’s
window. (b) Boundary unlocked: state variable changes its states as the polarity of applied voltage changes for Biolek’s window function.
problem, but the discontinuity problem persists due to the use
6 0.9 ( { [ ] }p)
0.8 f (w) = j 1 − 0.25 x − sgn (−i) 2 + 0.75 , (27)
4
0.7
2 0.6 sgn(i) = 1, when i ≥ 0, sgn(i) = 0, when i ≤ 0.
0 0.5 The stochastic variations of dopants are also modeled by
0 10 20 30 40
t (ms) a current as well as state dependent window function[30]. It
poses an exponential function, which provides an ability to real-
Normalize state variable (x)
14 0.4
0.3
12 3.2. Nonlinear memristor model
0.2
10 0.1
A 50 nm metal oxide of TiO2 film was fabricated
8 0
0 10 20 30 40 between two Platinum electrodes and experimentally manifes-
t (ms) ted memristive switching in nanoscale devices by Yang and
Fig. 6. (Color online) Prodromakis’s window function limits the Picket[31]. The oxygen vacancies (TiO2−x) were created in the
boundaries growth within the device’s physical limit and includes right half of TiO2 by an annealing process. Two metal/oxide in-
non-linearity at the (a) saturation and (b) depletion boundaries in the terfaces of the device were modeled by an ohmic interface (in
linear memristor model for the same parameters used in Fig. 4. case of heavy doping or TiO2−x) Fig. 7(a). and rectifying inter-
face (in case of low doping or TiO2) Fig. 7(b). The model has
a highly nonlinear behaviour, which was presented by cur-
ristance variations are shown in Figs. 6(a) saturation and 6(b)
rent-voltage relation (28). The first term of the equation charac-
depletion regions. It can be observed that both state variable
terizes the ON state and the second term is the estimation of
and the corresponding memristance are restricted within the
rectifier I–V expression and characterizes the OFF state of the
device’s physical dimensions. The parameters are kept the
memristor.
same in Fig. 4 for the simulation of the linear model without
[ ]
a window function; the state variable does not grow beyond i (t) = w(t)n β sinh (αv (t)) + χ exp (γv (t)) − 1 , (28)
‘1’ in saturation and below ‘0’ in depletion. Also the mem-
ristance is limited in between 100 to 8000 Ω in saturation and where α, β, γ, and χ are fitting parameter, ‘n’ determines the
8 to 16 kΩ in depletion; however, it also fails to resolve the effect of state variable on current. State derivative equation
sticking of states at boundaries, since f(0, 1) = 0. Here p and j (29) carries a window function f(w) along with a linear
are control parameters. voltage function g(v(t)).
{ [( )2 ]p} dw
w = a f (w) g (v (t)) . (29)
f (w) = j 1 − − 0.5 + 0.75 . (26) dt
D
(a) (b)
Platinum
Platinum
Memristor
TiO2−x
Rectifier
TiO2 φ2
V (t) φ1
V (t)
Fermi level
Fig. 7. (Color online) (a) Memristor device structure and (b) its equi-
valent non-linear circuit[31].
Platinum TiO2 width TiO2-x
Fig. 9. Different work function of two metal electrodes resulting in a
(a) (b)
+ Platinum trapezoidal barrier in thin insulating film[36].
Gon
( ) [ ( ) ]
TiO2-x
w − aoff |i|
Rs
dw i w
= foff sinh exp −exp − − , i > 0,
dt ioff wc b wc
(32)
( ) [ ( ) ]
dw i w − aon |i| w
= fon sinh exp −exp − − , i < 0.
dt ion wc b wc
− Platinum (33)
This model is an accurate but quite complex model of
Fig. 8. (a) Memristor device structure of the Simmons tunnel
the memristor since current voltage equations are not explicit
model[34]. (b) Spice modeling of Simmons tunnel’s tunneling width
to each other.
by capacitor and state variable derivatives for OFF and ON switch-
ing by current sources Goff and Gon[35].
3.3.1. Simmons spice model with rectangular barrier
two similar electrodes (30)[32], and also for dissimilar elec- A spice model on the basis of the Simmons tunnel memris-
trodes[33]. The derived formulae deployed a rectangular barri- tor model for similar electrodes was developed[35]. Here the bar-
er for similar electrodes and a trapezoidal barrier for dissimil- rier of the tunneling gap is rectangular since similar elec-
ar electrodes including image forces. Based on these expres- trodes have an equal work function that causes the same barri-
sions, the Simmons tunnel model for a memristor was de- er height. In this spice model, state derivative Eqs. (32) and
veloped[34]. A conductive channel of TiO2−x was created by (33) were modeled by two different current controlled
the electro-formation method and a thin electric tunneling sources Goff and Gon. Whereas increasing and decreasing the
gap was formed at one interface, see Fig. 8(a). The tunneling tunneling width was modeled by the voltage across capacitor
gap ‘w’ was modeled by a voltage source vg and the conduct- Fig. 8(b). In this spice model, it was seen that the ON switch-
ive channel was modeled by series resistance Rs, so the tunnel- ing current was 20% more than that of experimental results
ing current through the device is expressed as Eq. (36) with for a certain voltage.
gap voltage.
√ 3.3.2. Simmons spice model with trapezoidal barrier
j0 A
−B √φ (
) −B φI +evg
i=
φI e I − φI + e vg e
, (30)
∆w 2 The spice model of the Simmons tunnel for similar elec-
trodes[35] was further extended to dissimilar electrodes[36].
e , w = 1.2λw , ∆w = w − w , φ = φ – Two different electrodes with different work functions create
where j0 = 2πh 1 φ0 2 1 I 0
√ a trapezoidal barrier instead of a rectangular barrier for the
( w +w ) ( ) ( w (w−w ) )
e vg 1 w 2 − 1.15λw∆w ln w2 (w−w1 ) , B = 4π∆wh 2m , same electrodes, as shown in Fig. 9. The extended model was
( 1 2
) based on electric tunnel effects of a dissimilar electrode de-
2
e ln
λ = 8πkε w , 2 w2 = w1 + w 1 − 9.2λ
. rived by Simmons[33]. In this model, a parallel memcapacitor
0 3φ0 + 4λ − 2evg
with a memristor was accounted and showed that memcapacit-
The gap voltage is obtained by the following relation
ance is an inherent property of the memristor. The operating
v = vg + i (t) Rs , (31) frequencies determined the behaviour of the model as a memca-
pacitor or memristor.
074006-6
J. Semicond. 2018, 39(7) Jeetendra Singh et al.
( ) This model does not have a scaling factor in its window,
Ap eV(t) − eVp , V (t) > Vp ,
( ) but this model can be extended for an insulating layer separ-
g (V (t)) = −An e −V(t) −e ,
V V (t) < −Vn , (36)
n
ated by dissimilar electrodes. Its spice model is the same as
0, −Vn ⩽ V (t) ⩽ Vp . that of the Simmons tunnel model except it bore two addition-
al anti parallel diodes in Fig. 10. The TEAM model is further
In this model, spikes obtained by a linear DC sweep in-
extended as the voltage threshold adaptive memristor
put were more similar to neural spikes rather than a sinusoid-
(VTEAM) model[39] and makes it a voltage controlled device.
al input. Thus, it was more emphasized for the neuromorphic
application. The model’s validity was tested with TiO2, Si,
Ag, and the chalcogenide based memristor. Its validity was 3.6. Some more spice memristor models
also tested with spintronic devices, RRAM devices, and PCM
In order to simulate the behaviour of a memristor, spice
(phase change memory).
plays a key role and hence several spice models are available
in the literature. A SPICE macro model for the memristor can
3.5. Threshold adaptive memristor (TEAM) model be helpful in designing and testing of new circuits built with
a memristor[40]. Although this model was quite fast and stable
TEAM model tried to simplify the complexity of the Sim-
within the valid range of the parameters, it was unable to pick
mons tunnel model. It has two distinct current, voltage rela-
the initial state of the memristor. Soft and hard switching are
tions one for linear drift (37) and another for nonlinear drift
separately considered by another voltage controlled SPICE
(38)[38]. Here the state derivative Eq. (40) is modeled as the
model[41]. Instead of the window function, a ternary function
multiplication of two polynomials, one is the function of the
is exploited to limit the boundaries and thus the memristance
device current and the other is a function of state variable w.
of the device. A methodology to develop various spice mod-
[ ] els for mem circuit elements, e.g. memristor, memcapacitor,
R − RON
v (t) = RON + OFF (w − won ) i (t) , (37) and meminductor, is well described by Biolek et al.[42]. Vour-
woff − won
( ) kas et al. gave a voltage controlled SPICE model[43] with a pro-
λ grammable threshold. This model provides the best fitting
woff −won
(w−won )
v (t) = RON e i (t) , (38) with the characterization data. Here, two SPICE sub-circuits
where w∈[won, woff], and RON and ROFF are bounded resist- are used to solve boundary conditions.
ances and satisfy the following relation
4. Memristor in memory design
ROFF
= eλ , (39)
RON The memristor, either as a crossbar or discrete has vital
applications in analog and digital electronics. It is widely
( i(t) )αoff used in the designing of neuromorphic circuits[44], chaotic
k − 1 foff (w) , 0 < ioff < i,
dw (t)
off ioff circuits[45], logic gates[46], variable gain amplifiers[47], oscill-
=
0, ( ) ioff < i < ion , (40)
dt
kon i(t) − 1 αon fon (w) , i < ion < 0,
ators[48, 49], polarity-dependent memory switch[50], multil-
i on evel memory[51], logic circuits[52], and redox resistive memor-
ies[53]. Non-volatile and nanoscale properties are a memris-
where ioff & ion are threshold currents and koff (+) ve, kon (−)
tor’s merits in comparison to the other memory cell structure
ve, αoff, αon are constants. foff (w) and fon (w) are proposed
and made the memristor more promising in memory design-
as the window function, which are not necessarily the same
[ ( )] ing. Here some memory cells are summarized that utilize a
a n d g i v e n a s foff (w) = exp −exp w−a off
, fon(w) =
[ ( w−a )] w c memristor in their design. The significant outcomes and vari-
exp −exp − wc
on
. ous parameters such as write/read time, energy consumed, stat-
074006-7
J. Semicond. 2018, 39(7) Jeetendra Singh et al.
0
0 5 10 15 20 25
BL BL' Memristance range (kΩ)
(b) 20
L1
NT1 OFF ON NT2
15
Memristance (kΩ)
AMB1 Memristor AMB2
A B
10
L2 Write '1' Write '0'
Fig. 11. Hybrid memory cell exploiting a memristor and two ambi- 5
polar transistors[54].
Tw
ic noise margin, and write/read delay of these memory cells 0
0 100 200 300 400 500 600
are briefed and compared to traditional memories. t (ns)
Fig. 12. (a) Variation of the write time of hybrid memory cell for a
4.1. Hybrid memory cell range of memristance. (b) Consecutive Write ‘1’ and Write ‘0’ opera-
tion with changing memristance values.
A hybrid memory cell consists of two ambipolar transist-
ors and a memristor[54], see Fig. 11. This cell has a fast
WL
read/write operation and operates at the low temperature as BL
compared to NAND/NOR-based Flash memory[55] of CMOS;
Table 2 compares a hybrid memory cell with NAND/NOR
Memristor
Flash. An ambipolar transistor will behave as a NMOS transist-
or if its polarity gate (shown by an arrow) voltage is ‘0’, and
behaves as a PMOS transistor if polarity gate voltage is ‘1’.
The write time of the hybrid memory cell is measured by ob- LL
serving the memristance of the memristor, so the write time
is the time, required to set the memristor to its desired mem- Fig. 13. Unit 1T1R cell of array structure[56].
ristance state. Fig. 12(a) shows the memristance range versus
write time variations, it can be seen that for the memristance ive read operations. So, the calculated read time for a hybrid
range 100–19 kΩ the required write time is 219 ns. To write memory cell is 0.095 ns for N = 100 and TW = 219 ns. The
‘1’ initially the memristor should be in OFF state for node drawback of this memory cell was that it required a refresh
voltage VA > VB, and this condition came when BL = VDD operation after successive read operations because of a low
and BL’ = GND. The conductive width moves from left to threshold voltage and thus the low voltage across the memristor.
right and memristance goes from ROFF to RON. Similarly, a
write ‘0’ operation is performed when BL = GND and BL’ = 4.2. 1T1R memory cell and array
VDD. This makes node voltage VB > VA and memristance go
from RON to ROFF. The memristance variation corresponding The One Transistor One Memristor (1T1R) shown in
to write ‘1’ and write ‘0’ operation is shown in Fig. 12(b). Fig. 13 is the basic structure of the large memory array[56]
Read operation or data of the memory cell can be obtained by that provides high density, non-volatility, low power consump-
sensing the voltage difference between bit lines in high mem- tion, and has great potential to replace traditional Flash. The
ristance state ROFF or low memristance state RON of the memris- 1T1R has the same structure as the DRAM, the only differ-
tor. Therefore, consecutive read operations occur, when mem- ence is that it has a memristor in place of the capacitor. To per-
ristance changes its state. Thus, the read time is given as form the read operation the load line (LL) charges the bit line
T R = T W /2N, (41) (BL) via a memristor and access transistor. The write opera-
tion is performed by changing the memristance state of the
where TW is the write time and N is the number of consecut-
memristor. In the array, the Word Line (WL) is used to se-
074006-8
J. Semicond. 2018, 39(7) Jeetendra Singh et al.
Energy/op (pJ/bit)
Energy/op (pJ/bit)
DAC energy Sense amplifier load
TiO2 cell energy 1.0 Decoder overload
TiO2 cell energy + refresh energy
8
0.5
0 0
1 2 3 4 5 0 1 2 3 4 5
n n
Fig. 14. (Color online) Energy consumption in several components of TiO2 based RRAM array for multiple (n) bit storage during (a) write
operation, and (b) read operation.
(a) WL
Vdd
Q2 Q4
D DN
Q5 Q6
BL BLN
Q1 Q3
Q7
CTRL1
Vss
RRAM1
CTRL2
800
600
600
DN (mV)
D (mV)
400
400
200 200
0 0
0 200 400 600 800 0 100 200 300 400 500 600 700 800
DN (mV) D (mV)
Fig. 15. (Color online) (a) Basic structure of 7T1R SRAM[57]. (b) WSNM of 7T1R SRAM for storing ‘0’. (c) WSNM of 7T1R SRAM for
storing ‘1’.
lect a particular cell and the Bit Line (BL) is shared in a store n = 3 bit/cell. The optimized HfO2 based array utilized
column to read and write. An energy and performance model 365 and 173 fJ/b energy for 1 and 200 ns write and read ac-
was developed and compared to the proposed architecture cess time respectively, when storing the 3 bit/cell. This
utilizing the TiO2 and HfO2 base memristor. The amount of memory cell requires a refresh operation after successive
energy consumed during a write and read operation in differ- read operations because of the low threshold voltage. So this
ent components of the TiO2 based RRAM array is shown in challenge can be accepted in the design of memristor based
Fig. 14. In order to store a 3 bit/cell in a RRAM array, it is de- memory cells.
termined that the write energy, 4.06 pJ/b is consumed for
100 ns write access time as shown in Fig. 14(a). Whereas the 4.3. Memristor in SRAM
read energy consumed by the RRAM array is shown in
Fig. 14(b), which needed 188 fJ/b for 1ns read access time to
SRAM is frequently used in today’s memory designing,
074006-9
J. Semicond. 2018, 39(7) Jeetendra Singh et al.
Table 3. Comparison of various parameters of existing memory structures extracted from[12, 56] .
Traditional memories Other emerging technologies
Parameter
DRAM SRAM NOR FLASH NAND FLASH FeRAM MRAM PCRAM STTRAM MEMRISTOR
Cell element 1T1C 6T 1T 1T 1T1C 1T1R 1T1R 1T1R 1T1R
Feature size (nm) 36–65 45 90 22 180 65 45 40 9
Density (Gbit/cm2) 0.8–13 0.4 1.2 52 0.14 1.2 12 15 154–309
Read time (ns) 2–10 0.2 15 100 45 35 12 35 8.5
Write time (ns) 2–10 0.2 107 106 65 35 100 35 10
Retention time 4–64 ms N/A 10 years 10 years 10 years > 10 years > 10 years > 10 years > 10 years
WL BL
(a) (b)
Vdd Vdd RRAM2
RRAM1
Restore Restore
Q2 Q4 Q2 Q4
D DN
Q5 Q6 Q5 Q6
BL BLN
Q1 Q3 Q1 Q3
Q7 Q7 Q8
Q9
CTRL1 CTRL1 WLL WLR
Vss
RRAM1
RRAM2
Vss SL
CTRL2 CTRL2
Fig. 16. More configurations of SRAM utilizing memristor. (a) 8T2R SRAM[58]. (b) 9T2R SRAM[59].
101 101
100 100
45 32 16 45 32 16
Feature size (nm) Feature size (nm)
Fig. 17. (Color online) (a) Average read delay, (b) average write delay of various considered SRAM cells configuration with different fea-
ture size.
but it lags because of its volatile nature and dissipates more en- ness of a memory cell, so in order to measure the Write SNM
ergy in standby mode due to leakage. The memristor geared (WSNM), the 7T1R butterfly curve is plotted in Figs. 15(b)
up its uses in memory designing by providing it non-volatil- and 15(c) to store ‘0’ and ‘1’ respectively at 32 nm techno-
ity, low leakage, and Instant ON features. There are various logy. The measured WSNM of Figs. 15 (a) and 15(b) are
ways by which a memristor is used in SRAM design 7T1R[57] 0.297 and 0.313 V, which are less than the WSNM of 6T
as shown in Fig. 1(a), 8T2R[58] Fig. 16(a), 9T2R[59] Fig. 1(b). SRAM which had a value of 0.390 V[60]. Whereas 8T2R and
7T1R which have one 1T1R connected to data the nodes of 9T2R have WSNMs of 0.322 and 0.332 V respectively, these
the conventional 6T SRAM cell, so when the power goes, the values are more than that of 7T1R. The WSNM of 7T1R can
state of the node is stored in RRAM1 or the memristor and re- be increased by increasing the memristance of the RRAM1.
stored back to the node through transistor Q7, when power 8T2R consists of two 1T1R connected to the data nodes
comes back. Since the Static Noise Margin (SNM) is found D and DN and each is programmed to a high resistance state
to be an important metric in defining the stability and robust-
and a low resistance state. 9T2R have one more transistor in
074006-10
J. Semicond. 2018, 39(7) Jeetendra Singh et al.
addition with two 1T1R, which is used to restore the data ecutive report, 2015
from an access node. One end of both is connected to a single [9] Qazi M, Clinton M, Bartling S. A low-voltage 1 Mb FRAM in
bit line and the other end with a common source line, i.e. 0.13 μm CMOS featuring time-to-digital sensing for expanded
sources of Q7 and Q8 connected to the source line for program- operating margin. IEEE J Solid-State Circuits, 2012, 47(1): 141
[10] Waser R. Nanotechnology: volume 4: information technology
ming of RRAM. A comparison of an average read delay and
II. Weinheim: John Wiley & Sons, 2008
write delay arose in 6T SRAM, 7T1R SRAM, 8T2R SRAM, [11] Wong H S, Raoux S, Kim S, et al. Phase change memory. Proc
and 9T2R SRAM, which are plotted in Fig. 17 with various fea- IEEE, 2010, 98(12): 2201
tured sizes. It is found that the average read and write delays [12] Wang K L, Alzate J G, Amiri P K. Low-power non-volatile
for 6T SRAM[60] and 7T1R SRAM are almost comparable to spintronic memory: STT-RAM and beyond. J Phys D, 2013,
each other, whereas 9T2R shows maximum read and write 46(7): 074003
delay. [13] Lee M J, Lee C B, Lee D, et al. A fast, high-endurance and
scalable non-volatile memory device made from asymmetric
Ta2O5−x/TaO2−x bilayer structures. Nat Mater, 2011, 10(8): 625
5. Conclusion [14] Chua L. Memristor—the missing circuit element. IEEE Trans
Circuit Theory, 1971, 18(5): 507
Here, a comprehensive study of the memristor is pre- [15] Oster G. A note on memristors. IEEE Trans Circuits Syst,
sented, starting from its existence as a fourth fundamental 1974, 21(1): 152
electrical component to the memory applications stage. Vari- [16] Chua L O, Kang SM. Memristive devices and systems. Proc
ous window functions are implemented and discussed in or- IEEE, 1976, 64(2): 209
der to better understand their behaviour. It is found that Bio- [17] Adhikari S P, Sah M P, Kim H, et al. Three fingerprints of
lek’s window function removes the sticking of states and Pro- memristor. IEEE Trans Circuits Syst I, 2013, 60(11): 3008
dromakis’s window function provides better non-linearity at [18] Chua L. Device modeling via nonlinear circuit elements. IEEE
the boundaries. Its various models are compared in detail Trans Circuits Syst, 1980, 27(11): 1014
with their merits and demerits. The Simmons tunnel model is [19] Di Ventra M, Pershin Y V, Chua L O. Circuit elements with
memory: memristors, memcapacitors, and meminductors. Proc
one of the most accurate models among all the available mod-
IEEE, 2009, 97(10): 1717
els, but it is quite complex. A memristor’s nanoscale size
[20] Strukov D B, Snider G S, Stewart D R, et al. The missing
makes it more difficult to understand its basic tunneling mech- memristor found. Nature, 2008, 453(7191): 80
anism and coupled electric-ionic motion. So there is a need [21] Raj B, Vaidyanathan S. Advances in memristors, memristive
for an accurate mathematical model which has characterist- devices and systems. Springer International Publishing, 2017,
ics that best match with the experimental data and its beha- 701: 449
viour. Although there are several advantages of memristor- [22] Blanc J, Staebler D L. Electrocoloration in SrTiO3: vacancy
based memory cells, there are certain challenges remaining. drift and oxidation-reduction of transition metals. Phys Rev B,
Such as, it is seen that a hybrid memory cell requires an addi- 1971, 4(10): 3548
tional refresh operation. The 1T1R needed a read operation [23] Bansal P, Raj B. Memristor modeling and analysis for linear
before every write operation and SRAM cells with the memris- dopant drift kinetics. J Nanoeng Nanomanuf, 2016, 6(4): 306
[24] Strukov D B, Williams R S. Exponential ionic drift: fast
tor need to be explored. This paper is very helpful to research-
switching and low volatility of thin-film memristors. Appl
ers in building the basic concept of modeling and memory ap- Phys A, 2009, 94(3): 515
plications of the memristor. [25] Joglekar Y N, Wolf S J. The elusive memristor: properties of
basic electrical circuits. Eur J Phys, 2009, 30(4): 661
References [26] Bansal P, Raj B. Memristor: a versatile nonlinear model for
dopant drift and boundary issues. J Comput Theor Nanosci,
[1] Baughman R H, Zakhidov A A, De Heer W A. Carbon nan- 2017, 14(5): 2319
otubes–the route toward applications. Science, 2002, [27] Biolek Z, Biolek D, Biolkova V. SPICE model of memristor
297(5582): 787 with nonlinear dopant drift. Radioengineering, 2009, 18(2):
[2] Greytak A B, Lauhon L J, Gudiksen M S, et al. Growth and 210
transport properties of complementary germanium nanowire [28] Prodromakis T, Peh B P, Papavassiliou C, et al. A versatile
field-effect transistors. Appl Phys Lett, 2004, 84(21): 4176 memristor model with nonlinear dopant kinetics. IEEE Trans
[3] Heyns M, Tsai W. Ultimate scaling of CMOS logic devices Electron Devices, 2011, 58(9): 3099
with Ge and III–V materials. Mrs Bulletin, 2009, 34(7): 485 [29] Zha J, Huang H, Liu Y. A novel window function for memris-
[4] Knoch J, Appenzeller J. Modeling of high-performance p-type tor model with application in programming analog circuits.
III–V heterojunction tunnel FETs. IEEE Electron Device Lett, IEEE Trans Circuits Syst II, 2016, 63(5): 423
2010, 31(4): 305 [30] Singh J, Raj B. Frequency dependent window function for
[5] Salahuddin S, Datta S. Use of negative capacitance to provide stochastic nature of memristor model. 4th International Confer-
voltage amplification for low power nanoscale devices. Nano ence on Production and Industrial Engineering, CPIE, 2016: 19
Lett, 2008, 8(2): 405 [31] Yang J J, Pickett M D, Li X, et al. Memristive switching mech-
[6] Schliemann J, Egues J C, Loss D. Nonballistic spin-field-ef- anism for metal/oxide/metal nanodevices. Nat Nanotechnol,
fect transistor. Phys Rev Lett, 2003, 90(14): 146801 2008, 3(7): 429
[7] Jang W W, Yoon J B, Kim M S, et al. NEMS switch with [32] Simmons J G. Generalized formula for the electric tunnel ef-
30 nm-thick beam and 20 nm-thick air-gap for high density fect between similar electrodes separated by a thin insulating
non-volatile memory applications. Solid-State Electron, 2008, film. J Appl Phys, 1963, 34(6): 1793
52(10): 1578 [33] Simmons J G. Electric tunnel effect between dissimilar elec-
[8] International Technology Roadmap for Semiconductor 2.0, ex-
trodes separated by a thin insulating film. J Appl Phys, 1963,
074006-11
J. Semicond. 2018, 39(7) Jeetendra Singh et al.
34(9): 2581 5(1): 59
[34] Pickett M D, Strukov D B, Borghetti J L, et al. Switching dy- [48] Bao B C, Liu Z, Xu J P. Steady periodic memristor oscillator
namics in titanium dioxide memristive devices. J Appl Phys, with transient chaotic behaviours. Electron Lett, 2010, 46(3):
2009, 106(7): 074508 237
[35] Abdalla H, Pickett M D. SPICE modeling of memristors. 2011 [49] Talukdar A, Radwan A G, Salama K N. Generalized model for
IEEE International Symposium on Circuits and Systems memristor-based Wien family oscillators. Microelectron J,
(ISCAS), 2011: 1832 2011, 42(9): 1032
[36] Hassan A M, Fahmy H A, Rafat N H. Enhanced model of con- [50] Chandra S. On the discovery of a polarity-dependent memory
ductive filament-based memristor via including trapezoidal switch and/or memristor (memory resistor). IETE Techl Rev,
electron tunneling barrier effect. IEEE Trans Nanotechnol, 2010, 27(2): 179
2016, 15(3): 484 [51] Kim H, Sah M P, Yang C, et al. Memristor-based multilevel
[37] Yakopcic C, Taha T M, Subramanyam G, et al. A memristor memory. 2010 12th international workshop on Cellular Nano-
device model. IEEE Electron Device Lett, 2011, 32(10): 1436 scale Networks and Their Applications (CNNA), 2010: 1
[38] Kvatinsky S, Friedman E G, Kolodny A, et al. TEAM: [52] Robinett W, Pickett M, Borghetti J, et al. A memristor-based
threshold adaptive memristor model. IEEE Trans Circuits Syst nonvolatile latch circuit. Nanotechnology, 2010, 21(23):
I, 2013, 60(1): 211 235203
[39] Kvatinsky S, Ramadan M, Friedman E G, et al. VTEAM: a [53] Waser R, Dittmann R, Staikov G, et al. Redox-based resistive
general model for voltage-controlled memristors. IEEE Trans switching memories–nanoionic mechanisms, prospects, and
Circuits Syst II, 2015, 62(8): 786 challenges. Adv Mater, 2009, 21(25/26): 2632
[40] Rák Á, Cserey G. Macromodeling of the memristor in SPICE. [54] Junsangsri P, Lombardi F. Design of a hybrid memory cell us-
IEEE Trans Compur-Aid Des Integr Circuits Syst, 2010, 29(4): ing memristance and ambipolarity. IEEE Trans Nanotechnol,
632 2013, 12(1): 71
[41] Pershin Y V, Di Ventra M. SPICE model of memristive [55] International Technology Roadmap for Semiconductors
devices with threshold. arXiv preprint arXiv: 1204.2600, 2012 (ITRS). [Online]. Available: www.itrs.net/Links/2011ITRS/
[42] Biolek D, Di Ventra M, Pershin Y V. Reliable SPICE simula- 2011Chapters/2011ERD, 2011: 6.
tions of memristors, memcapacitors and meminductors. arXiv [56] Zangeneh M, Joshi A. Design and optimization of nonvolatile
preprint arXiv: 1307.2717, 2013 multibit 1T1R resistive RAM. IEEE Trans Very Large Scale
[43] Vourkas I, Batsos A, Sirakoulis G C. SPICE modeling of non- Integr (VLSI) Syst, 2014, 22(8): 1815
linear memristive behavior. Int J Circuit Theory Appl, 2015, [57] Wei W, Namba K, Han J, et al. Design of a nonvolatile 7T1R
43(5): 553 SRAM cell for instant-on operation. IEEE Trans Nanotechnol,
[44] Liu C, Yang Q, Zhang C, et al. A memristor-based neur- 2014, 13(5): 905
omorphic engine with a current sensing scheme for artificial [58] Turkyilmaz O, Onkaraiah S, Reyboz M, et al. RRAM-based
neural network applications. 22nd Asia and South Pacific on FPGA for “Normally Off, Instantly On” applications. Journal
Design Automation Conference (ASP-DAC), 2017: 647 of Parallel and Distributed Computing, 2014, 74(6): 2441
[45] Muthuswamy B, Kokate P P. Memristor-based chaotic circuits. [59] Xue X, Jian W, Xie Y, et al. Novel RRAM programming tech-
IETE Tech Rev, 2009, 26(6): 417 nology for instant-on and high-security FPGAs. 2011 IEEE 9th
[46] Kvatinsky S, Kolodny A, Weiser U C, et al. Memristor-based International Conference on ASIC (ASICON), 2011: 291
IMPLY logic design procedure. IEEE 29th International Con- [60] Wang J, Nalam S, Calhoun B H. Analyzing static and dynamic
ference on Computer Design (ICCD), 2011: 142 write margin for nanometer SRAMs. Proceedings of the 2008
[47] Wey T A, Jemison W D. Variable gain amplifier circuit using International Symposium on Low Power Electronics & Design,
titanium dioxide memristors. IET Circuits, Devices Syst, 2011, 2008: 129
074006-12