Chapter 2

Temperature Effects in Semiconductors

The changes in temperature described in the previous chapter affect the speed,
power, and reliability of our systems. Throughout this book, we will examine all
three of these metrics, though the majority of our discussion will be on how
temperature affects the speed performance. In this chapter, we discuss the problem
of temperature variation at the device and circuit level. In Sect. 2.1, we provide a
background on the material dependences on temperature. In Sect. 2.2, the normal
and reverse temperature dependence regimes are described. In Sect. 2.3, we explore
how these dependences change with technology scaling and the introduction of new
processing materials, such as high-k dielectrics and metal gates.

2.1

Material Temperature Dependences

In this section we provide details about the impact of temperature on the MOSFET
energy band gap, carrier density, mobility, carrier diffusion, velocity saturation,
current density, threshold voltage, leakage current, interconnect resistance, and
electromigration.

2.1.1

Energy Band Gap

Temperature affects the properties of electronic systems in a number of fundamental ways. The most fundamental of properties is the energy band gap, Eg, which is
affected by temperature according to the Varshni equation [1]
Eg T Eg 0 

aE T 2
T bE

(2.1)

where Eg(0) is the band gap energy at absolute zero on the Kelvin scale in the given
material, and aE and bE are material-specific constants. Table 2.1 [2] provides these
D. Wolpert and P. Ampadu, Managing Temperature Effects in Nanoscale
Adaptive Systems, DOI 10.1007/978-1-4614-0748-5_2,
# Springer Science+Business Media, LLC 2012

15

16

2 Temperature Effects in Semiconductors

Table 2.1 Varshni equation

constants for GaAs, Si,
and Ge [2]

Material
GaAs
Si
Ge

Eg(0) (eV)
1.519
1.170
0.7437

aE (eV/K)
5.41*104
4.73*104
4.77*104

bE (K)
204
636
235

Fig. 2.1 Energy band gap

temperature dependence of
GaAs, Si, and Ge

constants for three material structures. Table 2.1 and (2.1) are used to generate
Fig. 2.1, which shows how the band gaps of the three materials decrease as
temperature increases (the labeled points are the band gap of each material at
room temperature).

2.1.2

Carrier Density

Carrier densities affect electrical and thermal conductivity, and are a function of the
effective density of states in the appropriate band (conduction for n-type, valence
for p-type), the Fermi energy level in the material (which is a function of temperature and dopant concentrations), and the temperature as given by the following
equations:
n N C e

EC EF
kT

(2.2)

2.1 Material Temperature Dependences

17

Fig. 2.2 Temperature dependence of n in a doped semiconductor

p N V e

EF EV
kT

(2.3)

where n is the electron density, p is the hole density, NC is the density of states
in the conduction band, NV is the density of states in the valence band, EC is
the conduction band energy level, EV is the valence band energy level, EF
is the Fermi energy level, k 1.381023 J/K is the Boltzmann constant, and T is
temperature.
The temperature dependence of carrier density is shown in Fig. 2.2 for a doped
material. In the ionization region, there is only enough latent energy in the material
to push a few of the dopant carriers into the conduction band. In the extrinsic region,
which is the desired region of operation, the carrier concentration is flat over a wide
range of temperatures; in this region, all of the dopant carriers have been energized
into the conduction band (i.e. n  ND) and there is very little thermal generation of
additional carriers. As the temperature increases, the extrinsic region turns into the
intrinsic region, and the number of thermally generated carriers exceeds the number
of donor carriers. The intrinsic carrier concentration in a material ni is generally
much smaller than the dopant carrier concentration at room temperature, but
ni (np) has a very strong temperature dependence [2]
Eg0

ni / T 1:5 e 2kT

(2.4)

18

2 Temperature Effects in Semiconductors

where Eg0 is the energy band gap at T 0 K. Depending upon the dopant
concentration, the number of thermally generated carriers can exceed the number
of dopant-generated carriers, increasing the potential for thermal variation problems.

2.1.3

Mobility

We pay particular attention to the temperature and electric field dependence of

mobility, as mobility is one of the two main factors (the other is threshold voltage)
resulting in the MOSFET temperature behavior shown later in this chapter.
The carrier mobility, m (cm2/Vs), describes the drift velocity of a particle in an
applied electric field. Under small to moderate electric fields, m vd/x where vd is
the drift velocity, and x is the electric field. MOSFET mobility has very complex
temperature dependence, defined by the interplay of the following four scattering
parameters: phonon scattering mph, surface roughness scattering msr, bulk charge
Coulombic scattering mcb, and interface charge Coulombic scattering mint [3].
Each of these scattering parameters is related to the temperature of the material,
T, and the effective transverse electric field in the channel, xeff, which is
approximated as [4, 5]
xeff


Vgs VT
Qinv Qb


eSi
6Tox

(2.5)

where  is a constant (  0.4 in PMOS devices and   0.5 in NMOS devices),

Qinv is the inversion layer charge density, Qb is the substrate depletion charge
density, and eSi 11.7 is the relative permittivity of Silicon. This approximation
is not very convenient for circuit analysis, so xeff is also approximated in terms of
the gate-source voltage Vgs, the threshold voltage VT, and gate oxide thickness Tox.
The Berkeley Short-Channel IGFET Model (BSIM), one of the most widely
used simulation models, combines these four scattering parameters into an effective
mobility, meff [3] using Matthiessens rule
1
1
1
1
1


/

mph T; Eeff msr T; Eeff mcb T; Eeff mint T; Eeff

meff T; Eeff

(2.6)

Phonon scattering refers to the potential for an electron to be scattered by a

lattice vibration. As temperature increases, lattice vibrations increase and the
probability of an electron being scattered by the lattice increases; thus, high
temperature mobilities are limited by phonon scattering (mph / T3/2), causing
mobility to decrease as temperature increases as shown in Fig. 2.3a. Surface
roughness scattering becomes dominant when high electric fields pull electrons
closer to the Si/SiO2 surface (msr / xeff2.1).

2.1 Material Temperature Dependences

19

Fig. 2.3 (a) Temperature dependence of electron and hole mobilities in Si for different dopant
concentrations [2], (b) Field dependence of mobility [7]

At low temperatures, electrons move more slowly, and lattice vibrations are
small as well; thus, the ion impurity forces which have little impact on high-energy
particles become the dominant limit to mobility. In this regime, decreasing temperature extends the amount of time electrons spend passing an impurity ion, causing
mobility to decrease as temperature decreases (mcb / T). This effect is emphasized

20

2 Temperature Effects in Semiconductors

in the high dopant concentration curves shown in Fig. 2.3a, where mobility
decreases with decreasing temperature (e.g. the mn 1.31017 dopant concentration
line below ~30 K).
The electric field dependence of mobility is shown in Fig. 2.3b. In bulk Coulombic scattering, increasing xeff increases the charge density in the channel; the
associated charge screening reduces the impact of mcb (/ xeff2). At low
temperatures, the interface charges have two conflicting dependences. Reduced
temperature reduces the carriers thermal velocity, which increases the impact of
interface charges; however, the reduced thermal velocity also reduces the screening
effect [6], and this reduction in screening dominates the temperature dependence
(mint / T 1). The electric field screening effect is also weakened by the reduced
thermal velocity (mint / xeff, not xeff2 as in the mcb limit). In this book, we consider
devices operating in the phonon scattering limit, with temperatures >200 K; thus,
mobility will decrease as temperature increases.
The temperature dependence of mobility plays a major role in temperatureaware system design, and is discussed in more detail in the next subsection. In
room temperature Si, the electron mobility, mn, is nearly three times as large as the
hole mobility, mp, with mn 1,350 cm2/Vs and mp 480 cm2/Vs.

2.1.4

Carrier Diffusion

Diffusion is the movement of particles from a region of high concentration to a

region of low concentration. Carrier diffusion coefficients Dn and Dp (for electrons
and holes, respectively) are related to mobility by the Einstein relationship
D kT

m
q

(2.7)

Here, q is the charge on an electron (1.61019 C), and kT/q is an important value
known as the thermal voltage, fT. At room temperature (300 K), fT 0.0259 V.
Dn and Dp in room temperature silicon are 36 and 12 cm2/s, respectively.

2.1.5

Velocity Saturation

Although saturation velocity has been recently found to be a dominant temperaturedependent parameter, notable work had been performed in this area as far back as
1970 [8] using device lengths of 10 mm. In the BSIM4 device model, the impact of
temperature on velocity saturation vsat is modeled by [9]
vsat vsat0  1  avsat T  T0 

(2.8)

2.1 Material Temperature Dependences

21

where vsat0 is the saturation velocity at nominal temperature (T0) and avsat is the
saturation velocity temperature coefficient. Qualitatively, velocity saturation is the
point at which increases in energy no longer cause carrier velocity to increase; instead,
the additional energy is lost to phonon generation through lattice interactions.
In the results presented in this book, devices operate in the velocity saturation
regime; thus, the impact of temperature on saturation velocity (increasing temperature decreases vsat) is one of the most important criteria affecting the overall impact
of temperature on device current, as will be shown later in this chapter.

2.1.6

Current Density

The temperature dependence of the carrier concentrations, mobilities and diffusion

coefficients affect the temperature behavior of the carrier current densities, with the
carrier densities defined by the following formulas [10]:
JN qmn nx qDn rn

(2.9)

JP qmp px  qDp rp

(2.10)

where JN and JP are the electron and hole current densities, respectively. The first
term in each equation is the drift component of the total current, with mn and
mp corresponding to the electron and hole mobilities, respectively; x is the electric
field. The second term in each equation is the diffusion component of the total
current, with n and p corresponding to the electron and hole concentration
gradients (if there is no concentration gradient, there is no diffusion). The temperature dependent parameter in the second term is the diffusion coefficient. Increased
temperature increases particle kinetic energy, increasing the diffusion component of
total current. The drift component of the total current has two temperature dependent
parameters, the mobility and the carrier density. The mobility term was shown in
Fig. 2.3 to decrease as temperature increases (in the lattice vibration-limited case)
while the carrier density remains nearly fixed with temperature over the extrinsic
range (our intended range of operation), as indicated by Fig. 2.2. Thus, we determine
that the drift component of the total current decreases as temperature increases.
The drift and diffusion currents have opposing temperature dependencies,
which causes the net current change to depend on the applied electric field. In the
high-field (drift-dominated) case, current decreases as temperature increases; in
the low-field (diffusion-dominated) case, current increases as temperature
increases. However, if the system in question is a multi-voltage system, and the
system has both drift- and diffusion-dominated components, the impact of temperature variation may become less well-defined. The difference between a driftdominated system and a diffusion-dominated system is defined by the threshold
voltage, VT. We will show that the temperature dependences of mobility and
threshold voltage result in some very interesting device behavior.

22

2.1.7

2 Temperature Effects in Semiconductors

Threshold Voltage

The MOSFET threshold voltage is given by [2]

VT VFB 2fF g

p
2fF

(2.11)

where VFB fgs(Qss/Cox) is the flat band voltage, with the gate-substrate contact
potential fgs fTln(NANG/ni2), NA and NG are the substrate and gate doping
concentrations, respectively, Qss the surface charge density, and Cox the oxide
capacitance; g Cox(2qeSiNA)0.5 is a body effect parameter, with eSi the relative
permittivity of Si; fF fTln(NA/ni) is the Fermi energy with the thermal voltage
fT kT/q, and ni the intrinsic carrier concentration of Si.
Of the parameters in (2.11), fgs and fF vary with temperature (each contains fT
and ni terms). The threshold voltage temperature dependence VT/T may thus be
written as [11]
@VT @fgs
@f
g @fF

2 F p
@T
@T
@T
2fF @T

(2.12)

where the temperature dependencies of fgs and fF are [11]



@fgs 1
EG0 3kT
fgs

T
q
@T
q

(2.13)




@fF 1
EG0 3kT
fF 

T
2q
@T
2q

(2.14)

Filanovsky [11] used empirical parameters from a 0.35 mm CMOS technology to

determine that the three terms in (2.12) are 3.1, 2.7, and 0.43 mV/K, resulting in
a net threshold temperature coefficient of 0.83 mV/K. The threshold voltage in a
MOSFET is commonly modeled to decrease linearly with increasing temperature;
the parameter is plotted in Fig. 2.4 over a range of oxide thicknesses d and dopant
concentrations NA.

2.1.8

Leakage Current

Subthreshold leakage current Isub is exponentially dependent on temperature, as

shown in Fig. 2.5; a common rule of thumb is that leakage current doubles for every
10 C increase in temperature [12]. When VGS 0, Isub may be represented by the
Shockley diode model
 VDS

Isub I0 e fT  1

(2.15)

2.1 Material Temperature Dependences

23

Fig. 2.4 Change in threshold voltage temperature dependence at room temperature vs. dopant
concentration, with oxide thickness d [2]

Fig. 2.5 Temperature dependence of subthreshold leakage current (VGS 0 V) [14]

I0 ATe

1:12
 2f
T

(2.16)

where I0 is the reverse saturation current [12], A is a constant, and VDS is the drainsource voltage. Recalling that fT kT/q, we see that I0 is responsible for the
exponential temperature dependence shown in Fig. 2.5.

24

2 Temperature Effects in Semiconductors

The temperature dependence of gate leakage current has been shown to be very
minor compared to that of subthreshold leakage current [13].

2.1.9

Interconnect Resistance

The interconnect resistance R is related to temperature by

RT R0 1 aR T  T0 

(2.17)

where T is the temperature, R0 is the resistance at nominal temperature T0, and aR is

an empirical term named the temperature coefficient of resistance. Al and Cu
interconnects have similar values of aR0.004308 and 0.00401, respectively.
Over the military-specified temperature range, Al wire resistances can change by
up to 77.5% while Cu wire resistances can change by up to 72.2%. Interconnect
resistance increases with increasing temperature, complicating evaluation of the
impact of temperature on interconnect linksin these applications, the MOSFET
currents may either increase or decrease in temperature (as explored in the next
subsection), which means that the impact of temperature on interconnect resistance
can either add to the system temperature dependence or reduce the temperature
dependence, depending on the operating conditions.

2.1.10 Electromigration
Electromigration is a failure mechanism caused by high-energy electrons impacting
the atoms in a material and causing them to shift position. It is most problematic in
areas of high current density. This can form a positive feedback path can form
where electromigration will cause an atom to move down a wire, slightly narrowing
the wire width at that location and increasing the current density; this increased
current density then further increases electromigration, causing more atoms to be
displaced. This brings about two failure mechanisms: (1) the narrowing of the wire
will increase wire resistance, which may cause a timing failure if a signal can no
longer propagate within the clock period, or (2) electromigration will continue until
the wire completely breaks, allowing no further current flow and resulting in
functional failure.
Electromigrations impact on a systems reliability is measured in terms of a
mean time to failure (MTTF) using Blacks equation [15]
MTTF Aj  J nj  e kT

Ea

(2.18)

2.2 Normal and Reverse Temperature Dependence

25

where Aj is a constant related to the cross-sectional area of a wire, J is the current

density, n is a constant scaling factor, Ea is the activation energy, k is the Boltzmann
constant, and T is temperature. Thus, the MTTF is exponentially dependent on
temperature.

2.2

Normal and Reverse Temperature Dependence

Changes in temperature affect system speed, power, and reliability by altering the
threshold voltage [11], mobility [11], and saturation velocity [16] in each device.
The resulting changes in device current can lead to failures in timing, cause systems
to exceed power or energy budgets, and result in communication errors between IP
cores. The temperature relationships for MOSFET mobility, threshold voltage, and
velocity saturation are related to temperature using the following empirical
expressions [17]:
mT m0 T=T0 am

(2.19)

VT T VT0 aVT T  T0

(2.20)

vsat T vsat0 avsat T  T0

(2.21)

where T is the temperature; T0 is the nominal temperature; m0, VT0, and vsat0 are the
mobility, threshold voltage, and saturation velocity at T0, respectively; am, aVT, and
avsat are empirical parameters named the mobility temperature exponent, threshold
voltage temperature coefficient, and saturation velocity temperature coefficient,
respectively, where am  1.3, aVT  3 mV/ C, and avsat  97 m/(s C).
Two temperature dependencies exist: the normal dependence (ND) region, where
drain current (ID) decreases with increasing temperature, and the reverse dependence (RD) region, where ID increases with increasing temperature [18]. Between
the two regions, there is a supply voltage where the impact of temperature on delay
is minimized. This is referred to as the temperature-insensitive voltage VINS [19],
and as technology scales this voltage approaches nominal voltage.
In the temperature region of concern (between 55 C and 125 C, the range of
military operating temperatures [20]), m, VT, and vsat all decrease with increasing
temperature. Examining the velocity-saturated MOSFET drain current ID(T) [21]
we see that decreasing vsat decreases ID, while decreasing VT increases ID [22].
ID T vsat T  W  Ps  VGS  VT Ta

(2.22)

Where W is the device width, Ps is a technology-specific constant, VGS is the MOSFET

gate-source voltage, and a is a technology-specific exponent. The temperature dependence of the MOSFET drain current, dID/dT, can be determined by the sum of
the impacts of vsat and VT on ID, composed of four values(1) the change in velocity

26

2 Temperature Effects in Semiconductors

saturation for a change in temperature, dvsat/dT, (2) the change in threshold

voltage for a change in temperature, dVT/dT, (3) the change in device current for
a change in velocity saturation, ID/vsat, and (4) the change in device current
for a change in threshold voltage, ID/VT:



dID
dID
dID
@ID dvsat @ID dvT



dT Tot
dT vsat dT vT @vsat dT
@vT dT

(2.23)

dID/dT|vsat is negative, and dID/dT|VT is positive. At nominal voltage in conventional CMOS technologies, the magnitude of dID/dT|vsat is greater than the
magnitude of dID/dT|VT; thus, circuits at nominal voltages become slower as
temperature increases. However, as VGS approaches VT, a change in VT has a larger
impact on ID; thus, at lower supply voltages, the magnitude of dID/dT|vsat is less
than the magnitude of dID/dT|VT, and device delay decreases as temperature
increases (the reverse temperature dependence). VINS occurs where dID/dT|Tot
approaches zero, with dID/dT|vsat  dID/dT|VT; however, because vsat and VT
differ between NMOS and PMOS devices, each type of device has a different
value of VINS. The dependence regions are shown in Fig. 2.6 for plots of the
current through diode-connected PMOS and NMOS devices in a 90 nm technology
model [23] over the range of military operating temperatures. In Fig. 2.6a,
VINS occurs in the shaded regions, with higher voltages exhibiting the normal
temperature dependence and lower voltages exhibiting the reverse temperature
dependence.
The reverse temperature dependence is occasionally referred to as temperature
inversion, while the normal and reverse temperature dependences are also referred
to as negative (for normal dependence) and positive (for reverse dependence)
current-temperature (I-T) slopes. In this document, we will use the  I-T slope
terminology as shorthand for the normal and reverse temperature dependences.
The difference between the 125 C and 55 C endpoints of Fig. 2.6a is presented
in Fig. 2.6b. In Fig. 2.6b, VINS is indicated in each device by the minimum points in
each curve; the absolute minimum for a 1:1 sizing ratio occurs at 345 mV,
corresponding to an 18% total change in current over the entire 180 C range of
ambient temperatures.

2.2.1

Discovery of the Normal and Reverse

Temperature Dependences

This book is by no means the first document to report on the reverse temperature
dependence. Indeed, what we name the reverse temperature dependence (i.e. the
increasing of electrical conduction with increasing temperature) was first discovered
by Faraday with his silver sulphide experiments mentioned in the previous chapter.
However, the mechanism detected by Faraday was quite different than the

2.2 Normal and Reverse Temperature Dependence

27

Fig. 2.6 (a) Device current across a range of temperatures and supply voltages in a 90 nm
technology, (b) temperature change from 125 C to 55 C

mechanisms causing the normal and reverse temperature dependences in our

modern Silicon electronics. The first recorded mention of the reversal of the temperature dependence describing the trade-off between mobility and threshold voltage is
attributed to C. Park, et al., from Motorola in 1995, in a conference paper exploring
the impact of temperature on integrated circuits at very low voltages [18]. In the time
since, the reversal of the temperature dependence has been explored in great detail
[11, 24, 25], including magazines, patents, and journal papers, and is now being
considered in industry-standard tools [26].

28

2.3

2 Temperature Effects in Semiconductors

Temperature and Technology Scaling

VINS occurs at increasingly larger supply voltages as technology is scaled, and is

fast approaching the nominal supply voltage VNOM (which is steadily decreasing
as technology scales) as shown in Table 2.2, particularly with the introduction
of high-k dielectrics to replace SiO2 [27] (high-k dielectrics reduce m and change
m/T [28], altering the balance of the m and VT impacts). The change in VINS as
technology scales is caused by changes in the threshold voltage and saturation
velocity [29]. The data in Table 2.2 were generated using predictive technology
models [23], with the 45, 32, and 22 nm data points using high-k dielectric/metal
gate models.
As VINS approaches VNOM, adaptive systems which vary supply voltage to reduce
energy consumption or improve reliability will have operating voltages in both the
normal and reverse temperature dependence regions, making it unclear if circuits
will increase in speed or decrease in speed as chip temperatures increase, and
exacerbating problems associated with inter-die temperature variation; solutions
to this issue are discussed in Chap. 3.
The current dependencies for a 22 nm technology with high-k dielectrics and
metal gates are shown in Fig. 2.7. As shown, VINS in the PMOS device increases
from ~375 mV at 90 nm to ~575 mV at 22 nm, with the nominal supply voltage
decreasing from 1.2 V at 90 nm to 0.8 V at 22 nm. The 22 nm NMOS device is in
the reverse temperature dependence region even at the nominal supply voltage.

2.3.1

The Reverse Temperature Effect and High-k/Metal

Gate Technology

VT, m, vsat and nominal supply voltage are all technology dependent parameters,
with predicted values available down to the 22 nm node [23, 30]. Use of high-k
dielectrics and metal gates to alleviate nanoscale gate leakage problems also alters
VT, m, and vsat [31, 32]. The combination of these changes makes it difficult to
determine the effect of temperature on device performance. Two dependences exist,
as mentioned in the prior subsection: a normal temperature dependence, where
current decreases as temperature increases, and a reverse temperature dependence
[18, 19], where current increases as temperature increases.
Table 2.2 VINS approaches
VNOM as technology scales

Technology (nm)
90
65
45
32
22

VNOM (V)
1.2
1.1
1.0
0.9
0.8

VINS
(V)
0.39
0.40
0.61
0.69
0.73

VINS/
VNOM
0.33
0.36
0.61
0.77
0.91

2.3 Temperature and Technology Scaling

29

Fig. 2.7 Temperature dependence of device current across a range of supply voltages in a 22 nm
high-k/metal gate technology

These parameters are further complicated by environmental requirements

(military specifications call for a range of 55125 C [20]) and intra-die temperature variation (shown to exceed 50 C [33]). To account for the wide range of
conditions, as well as process and voltage variations, variation-tolerant adaptive
systems have been used to guarantee functionality by adjusting operating voltages
and frequencies [3436]; however, these systems with multiple voltage modes
make the above-mentioned temperature effects even more difficult to determine.
For large gate overdrives (VGSVT > VINS), the temperature dependence of a
device is dominated by the dependence of vsat, while for small gate overdrives
(VGSVT < VINS), small changes in VT can cause large changes in current, resulting
in a temperature dependence dominated by VT. Further examination of these effects
in SiO2 dielectric, polysilicon gate devices is available in [18, 19].
In nanoscale devices, high-k dielectrics and metal gates have been introduced to
reduce gate leakage due to thinning gate oxides and reduce the depletion effects of
polysilicon gates [31, 32]; unfortunately, these techniques have the effect of
dramatically altering the temperature dependence of ID. The extent of the change
is shown in Fig. 2.8, which compares 45 nm predictive technology models [23]
of both SiO2/poly gate (dashed line) and high-k/metal gate (solid line) devices.
Each line in Fig. 2.8 shows the change in delay of an inverter (sizing ratio b 2)
from 55 C to 125 C. For example, at 0.62 V, the high-k/metal gate inverter delay
is unchanged from 55 C to 125 C, resulting in the 0.62 V point occurring on the
0% line. This 0% intersect point on each curve represents VINS. As shown, VINS
in the high-k/metal gate is 40% higher than in the SiO2/poly gate devices.

30

2 Temperature Effects in Semiconductors

Fig. 2.8 Effect of high-k dielectric and metal gate on temperature dependence

The normal dependence region is below the 0% line, and the reverse dependence
region is above the 0% line.
Fig. 2.9a shows the change in PMOS device current from 55 C to 125 C at the
45, 32, and 22 nm technology nodes (with nominal voltages of 1, 0.9, and 0.8 V,
respectively). As shown, VINS increases by ~40 mV per technology node, with VINS
at 22 nm equal to 0.56 V. The NMOS device response, shown in Fig. 2.9b, is in the
reverse temperature dependence region over the entire range of operating voltages
at the 32 and 22 nm nodes.
The PMOS and NMOS devices are combined into an inverter with b 2 in
Fig. 2.9c. As shown, VINS approaches 90% of nominal voltage in the 22 nm node. As b
increases, the stronger PMOS effect decreases VINS. Thus, adaptive voltage systems
may easily wind up straddling both temperature domains in nanoscale systems,
making temperature-aware design increasingly critical as technology scales.
Reverse temperature dependence at near nominal voltages complicates variationtolerant system design, which uses multiple supply voltages to adjust for changes
in process, voltage, and temperature. The additional complexity needed to account
for both normal and reverse temperature dependence depends on the available
design time information. If the system can be fully characterized at design time,
then the multiple dependences can be programmed into the voltage and frequency
look-up table entries [34] to ensure that the system adapts in the correct direction
given a change in temperature. For example, whereas a low-voltage system would
generally reduce the frequency as temperature increases, in the reverse dependence
region the system would have to reduce the frequency when temperature decreases.

2.3 Temperature and Technology Scaling

31

Fig. 2.9 Changes in (a) PMOS current, (b) NMOS current, and (c) inverter delay over the 55 C
to 125 C temperature range

The reverse temperature effect is particularly important to consider in adaptive

systems because of thermal runaway. In the normal dependence region, temperatures
are prevented from increasing to dangerous levels because the delay becomes so
large that the adaptive system is forced to reduce the clock frequency, reducing the
energy and therefore the temperature. In the reverse temperature dependence region,
circuits continue to speed up as temperature increases; there is no delay limit on high
temperature operation. The higher temperatures could result in thermal runaway
resulting from the exponential temperature dependence of leakage current [37], which
may already be dominating the total power consumption in the nanoscale regime [38].
If the temperature dependences are not known at design time (from tool
limitations, process variations, unknown IR drops, etc.), there are two options to
ensure variation-tolerance. The system may be designed with large enough
guardbands that it can operate correctly over the entire temperature range regardless
of the temperature dependence, though this will result in a large reduction in delay
performance. Another option is to use a temperature dependence sensor to determine the temperature dependence at each operating voltage; we propose the first
temperature dependence sensor in Chap. 3.

32

2 Temperature Effects in Semiconductors

References
1. Varshni YP (1967) Temperature dependence of the energy gap in semiconductors. Physica
34:149154
2. Sze SM (1981) Physics of semiconductor devices, 2nd ed. John Wiley and Sons, NY
3. Chain K, Huang JH, Duster J, Ko PK, Hu C (1997) A MOSFET electron mobility model of
wide temperature range (77400K) for IC simulation. Semicond Sci Technol 12:355358
4. Sabnis AG, Clemens JT (1979) Characterization of the electron mobility in the inverter Si
surface. Int Electron Devices Mtg 1821
5. Chen K, Wann HC, Dunster J, Ko PK, Hu C (1996) MOSFET carrier mobility model based on
gate oxide thickness, threshold and gate voltages. Solid-State Electronics 39:15151518
6. Jeon DS, Burk DE (1989) MOSFET electron inversion layer mobilitiesa physically based
semi-empirical model for a wide temperature range. IEEE Trans Electron Devices
36:14561463
7. Grabinski W, Bucher M, Sallese JM, Krummenacher F (2000) Compact modeling of ultra deep
submicron CMOS devices. Int Conf on Signals and Electronic Systems 1327
8. Fang FF, Fowler AB (1970) Hot electron effects and saturation velocities in Silicon inversion
layers. J Appl Phys 41:18251831
9. Cheng Y et al (1997) Modelling temperature effects of quarter micrometer MOSFETs in
BSIM3v3 for circuit simulation. Semicond Sci Technol 12:13491354
10. Pierret RF (1988) Semiconductor fundamentals, 2nd ed. Addison-Wesley, MA
11. Filanovsky IM, Allam A (2001) Mutual compensation of mobility and threshold voltage
temperature effects with applications in CMOS circuits. IEEE Trans Circuits and Syst I:
Fundamental Theory and Applications 48:876884
12. Oxner ES (1988) FET technology and application. CRC Press, NY
13. Agarwal A, Mukhopadhyay S, Raychowdhury A, Roy K, Kim CH (2006) Leakage power
analysis and reduction for nanoscale circuits. IEEE Micro 26:6880
14. Fallah F, Pedram M (2005) Standby and active leakage current control and minimization in
CMOS VLSI systems. IEICE Trans Electronics E88-C:509519
15. Black JR (1969) Electromigrationa brief survey and some recent results. IEEE Trans Electron
Devices 16:338347
16. Ku JC, Ismail Y (2007) On the scaling of temperature-dependent effects. IEEE Trans Computer-Aided Design of Integrated Circuits and Syst 26:18821888
17. Morshed TH et al (2009) BSIM4.6.4 MOSFET model users manual. [Online] http://wwwdevice.eecs.berkeley.edu/~bsim3/bsim4_arch_ftp.html
18. Park C et al (1995) Reversal of temperature dependence of integrated circuits operating at very
low voltages. Int Electron Devices Mtg 7174
19. Bellaouar A, Fridi A, Elmasry MI, Itoh K (1998) Supply voltage scaling for temperatureinsensitive CMOS circuit operation. IEEE Trans Circuits and Syst II: Analog and Digital
Signal Processing 45:415417
20. US Dept of Defense (2007) Integrated circuits (microcircuits) manufacturing, general specification, Std MIL-PRF-38535H. Washington DC
21. Sakurai T, Newton AR (1990) Alpha-power law MOSFET model and its applications to
CMOS inverter delay and other formulas. IEEE J Solid-State Circuits 25:584594
22. Shichman H, Hodges DA (1968) Modeling and simulation of insulated-gate field-effect
transistor switching circuits. IEEE J Solid-State Circuits SC-3:285289
23. Zhao W, Cao Y (2006) New generation of predictive technology model for sub-45nm early
design exploration. IEEE Trans Electron Devices 53:28162823
24. Lasbouygues B, Wilson R, Azemard N, Maurine P (2006) Timing analysis in presence of
supply voltage and temperature variations. ACM Int Symp on Physical Design 1016
25. Kumar R, Kursun V (2006) Reversed temperature-dependent propagation delay characteristics
in nanometer CMOS circuits. IEEE Trans Circuits and Syst II: Express Briefs 53:10781082

References

33

26. Dasdan A, Hom I (2006) Handling inverted temperature dependence in static timing analysis.
ACM Trans Design Automation of Electronic Syst 11:306324
27. Wolpert D, Ampadu P (2008) Normal and reverse temperature dependence in variationtolerant nanoscale systems with high-k dielectrics and metal gates. 3rd ACM Int Conf on
Nano-networks 15
28. Wong H, Iwai H (2006) On the scaling issues and high-k replacement of ultrathin gate
dielectrics for nanoscale MOS transistors. Microelectronic Engineering 83:18671904
29. Langen D, Ruckert U (2002) Extending scaling theory by adequately considering velocity
saturation. 15th Ann IEEE Int ASIC/SoC Conf 145149
30. Zhao W (2008) Personal communication
31. Guillaumot B et al (2002) 75nm damascene metal gate and high-k integration for advanced
CMOS devices. Int Electron Devices Mtg 355358
32. Cheng B et al (1999) The impact of high-k gate dielectrics and metal gate electrodes on sub100 nm MOSFETs. IEEE Trans Electron Devices 46:15371544
33. Sato T, Ichimiya J, Ono N, Hachiya K, Hashimoto M (2005) On-chip thermal gradient analysis
and temperature flattening for SoC design. IEICE Trans Fundamentals E88-A:33823389
34. Tschanz J et al (2007) Adaptive frequency and biasing techniques for tolerance to dynamic
temperature-voltage variations and aging. IEEE Int Solid-State Circuits Conf 292604
35. Elgebaly M, Sachdev M (2007) Variation-aware adaptive voltage scaling system. IEEE Trans
Very Large Scale Integr Syst 15:560571
36. Martin S, Flautner K, Mudge T, Blaauw D (2002) Combined dynamic voltage scaling and
adaptive body biasing for lower power microprocessors under dynamic workloads. IEEE/
ACM Int Conf on Computer-Aided Design 721725
37. Lee CC, de Groot J (2006) On the thermal stability margins of high-leakage current packaged
devices. 8th Electronics Packaging Technology Conf 487491
38. Kim NS et al (2003) Leakage current: Moores law meets static power. Computer 36:6875

http://www.springer.com/978-1-4614-0747-8