Professional Documents
Culture Documents
= +2
10
I. INTRODUCTION
1927
tion
, where
(2)
[31].) Since the carrier scattering is most important in a relaof the channel near the maximum
tively short section
of the potential profile, and our goal is to study ultimate performance of sub-10-nm MOSFETs, we believe we can neglect
scattering. This allows us to use Schrdinger equation for the
electron transport description.
Recently, a more general approach based on nonequilibrium
(Keldysh) Greens functions was used to calculate some
characteristics of nanoscale MOSFETs [24], [25], [27]. However, since the possible elastic scattering (due to impurities
and interface roughness) has been neglected in these works as
well, that approach, being much more complex, is equivalent to
the Schrdinger equation solution. (The mean free path due to
inelastic electron-phonon scattering in Si at room temperature
is above 50 nm [28], [33], [34], and is evidently negligible for
sub-10-nm devices.) Indeed, we have checked [30] that this
simple approach allows to reproduce virtually perfectly the
results shown in Fig. 2 of [25].
For transistors with ultrathin channels, we had to take into
account strong lateral confinement in the direction (in Fig. 1,
axis ) perpendicular to the channel plane. As a result of this
confinement, only the first subband is populated by electrons
from two of six conduction valleys of silicon (with
, and
). The wavefunction
of these electrons is taken in the form
(1)
that is an accurate approximation for our range of channel thicknm) [35]. Substitution of (1) into the Schrdinger
ness (
equation leads to an effective 1-D equation for the wave func-
1928
(a)
(a)
(b)
(b)
0 h i
barrier lowering (DIBL). Unfortunately, numerical modeling indicates that for our transistor model (Fig. 1) the 1-D approximation does not always give the correct result even in the limit
. The reason for this discrepancy is that the profile of the
typically features a hump near
effective 1-D potential
the source, at positive gate voltages (Fig. 3). It is created by uncompensated space charge of electrons injected from the electrodes into the undoped channel, in the region where the gate
field is effectively screened by bulk source and drain. (In this
sense our model is not a well-tempered MOSFET [36].) It is
rather this hump, than the following long potential plateau, that
provides the real electron transport bottleneck and therefore de.
termines the transport at
(c)
Fig. 4. Drain IV curves of double-gate, ballistic MOSFETs for three values
of gate length L. Besides L, all the parameters are the same as for Fig. 2. The
drain-to-source voltage is 0.5 V, the gate voltage V is changed with a step of
0.05 V.
1929
(a)
(b)
(c)
Fig. 5. Subthreshold curves of nanoscale MOSFETs for the same three values
of gate length L as in Fig. 4, each for ten values of drain-source voltage V (with
a step of 0.05 V). Transistor parameters are the same as for Fig. 2, besides a
thicker oxide (t
2:5 nm). Dotted lines show the ideal slope of the curves
(60 mV/decade). Dashed lines show the estimated gate oxide leakage current.
1930
(a)
!1
At this point we have to acknowledge that as becomes comparable with channel thickness (as it does for
nm), the
of electron quantum constep-like increase of the energy
finement in the direction perpendicular to the channel plane,
that follows from our 1-D version of the Schrdinger equation
and is visible in Fig. 3, does not exactly reflect the real situation.
, leading to
In reality, this step is smoothed by a distance
a slight reduction of the potential bump and hence reducing the
transport bottleneck. This is why our results for this value of
for
may be not completely accurate. We plan to explore
this effect, using the 2-D version of the Schrdinger equation,
in the near future.
IV. PARAMETER SENSITIVITY
From the point of view of the present-day CMOS technology,
the results presented above seem very optimistic: to summarize,
they indicate that physics allows the MOSFET gate length to be
nm, still retaining the high perforscaled down to at least
mance necessary for most logic and memory circuits. However,
these results also show two major challenges on the way to approaching these physical limits in practice.
The first problem is the rapidly increasing sensitivity of
transistor characteristics to unavoidable random spreads of their
geometrical parameters, due to fabrication uncertainties. For
example, Fig. 7 shows the magnitude of decrease (roll-off)
(defined as the value of
of the threshold gate voltage
providing a certain small drain current) with a reduction of the
gate length. The plot shows that a small uncertainty in gate
dimension may lead to a large variation of . In practically
due to random fluctuations
useful devices, the variations of
should be small in comparison with the power supply
of
necessary for driving the device. For example,
voltage
for
nm, in order to keep fluctuations of
below a
reasonable limit of 50 mV, the critical dimension should be
(b)
Fig. 8. Threshold voltage dependence on (a) the channel thickness t and (b) the
oxide thickness t .
1931
(a)
length
and the corresponding change of the
source-to-drain tunneling. This crossover leads also to the
for various tox (Fig. 7).
crossing of the plots
V. POWER SCALING
Another problem arising at
nm is the growing power
consumption. In order to analyze this effect we have used a
simple model [37] for the total power in a CMOS circuits, where
the total power is a sum
(3)
of dynamic (first term) and static (second term) power dissipation. Here is the switching activity factor of the -th circuit
is its total effective capacitance including that of all
block,
the interconnects and input capacitance of transistors, is the
is the total OFF current of all transisclock frequency, and
. In contrast with
tors biased by the power supply voltage
, the on-current
participates in (3) indirectly, via the speed requirement
(4)
is the fraction of the clock period
taken
where
by the capacitance recharging constant .
The model expressed by (3) and (4) is of course rather approximate; however, it captures the basic balance between static
and dynamic power. A major advantage of this model is that the
minimum value of the specific power per unit transistor width
, on just two parameters
depends, besides
(5)
(b)
Fig. 10. Minimum total power (solid lines) and optimum V
(dashed lines)
as functions of gate length L, for several values of the ON current and two values
of the parameter defined by (5).
where
is the total width of logic transistors in the -th ciris the total width of all transistors providing
cuit block, and
. In fact, combining (3)(5) we get
leakage current
(6)
, this ratio (i.e., the specific power per unit
For a fixed
channel width) may be readily found from the transistor charfrom fixed
we deteracteristics
necessary to sustain it.
mine the gate voltage value
is then determined as
, i.e., a result of gate
.
voltage reduction by the voltage swing equal to
Fig. 9 shows a typical dependence of the resulting power concalculated for several values of . At small
sumption on
voltages it is dominated by static power, because transistors
dynamic power domcannot be shut well, while at large
inates and leads to the growth of total power. As a result, as a
1932
function of
displays a clear minimum. Fig. 10 shows this
minimum value of power, as well as the corresponding optimal
, as a function of gate length , for two relatively
value of
large values of . (For smaller , the plots are virtually similar
, with
and
.)
to those for
The most important features of these results are that both the
and the optimum
increase
minimum specific power
nm, due to transistor performance degradation: as
below
is decreased, the MOSFET requires larger and larger voltage
to avoid the fast growth of static power if the device
swing
leads, of course,
has not been shut well. This increase of
to the growth of dynamic power and hence of the total .
The growth of specific power in sub-10-nm transistors may
considerably exacerbate the problem of total power consumption in silicon integrated circuits that is very serious even
without account of this effectsee, e.g., [38], [39].
VI. CONCLUSION
Calculations using a simple model of quasiultimate silicon
MOSFETs (Fig. 1) show that these devices with small but plaunm) may
sible values of gate oxide and channel thickness (
retain performance acceptable for both logic and memory applinm. However,
cations even if the gate length is reduced to
nm, power per unit channel
as soon as is decreased below
width starts to grow rapidly (Fig. 10). Even more importantly,
threshold voltage becomes an extremely sensitive function of
, and
(Figs. 7 and 8), creating very serious problems for
reproducible device fabrication.
Our recent work [30] shows that both these effects are more
general than this specific model transistors with thin source
and drain extensions show virtually the same behavior if their
.
channel length is in-between values and
ACKNOWLEDGMENT
The authors wish to thank D. Antoniadis, S. Cristoloveanu, S.
Datta, B. Doyle, D. Frank, C. Hu, Z. Krivokapic, M. Lundstrom,
P. Solomon, S. Tiwari, and Y. Taur, for numerous discussions
and valuable suggestions.
REFERENCES
[1] R. Sasajima, K. Fujimaru, and H. Matsumura, A metal-insulator tunnel
transistor with 16 nm channel length, Appl. Phys. Lett., vol. 74, pp.
32153217, May 1999.
[2] H. Kawaura, T. Sakamoto, and T. Baba, Observation of source-to-drain
direct tunneling in 8 nm gate electrically variable shallow junction MOSFETs, Appl. Phys. Lett., vol. 76, pp. 38103812, May 2000.
[3] J. Appenzeller et al., Scheme for the fabrication of ultrashort channel
MOSFETs, Appl. Phys. Lett., vol. 77, pp. 298300, July 2000.
[4] F. Boeff et al., 16 nm planar NMOSFGET manufacturable within
state-of-the-art CMOS process thanks to specific design optimization,
in IEDM Tech. Dig., 2001, pp. 637640.
[5] B. Yu et al., 15 nm gate length planar CMOS transistor, in IEDM Tech.
Dig., 2001, pp. 937939.
[6]
, FinFET scaling to 10 nm gate length, in IEDM Tech. Dig., 2002,
pp. 251254.
[7] B. Doris et al., Extreme scaling with ultrathin Si channel MOSFETs,
in IEDM Tech. Dig., 2002, pp. 267270.
[8] A. Hokazono et al., 14 nm gate length CMOSFETs utilizing low
thermal budget process with poly-SiGe and Ni salicide, in IEDM Tech.
Dig., 2002, pp. 639642.
1933