CMOS Scaling Trends and Beyond
CMOS Scaling Trends and Beyond
G
Mark T. Bohr, ordon Moore famously predicted in his 1965 paper that the number of com-
Ian A. Young ponents per chip would continue to increase by a factor of two every year.1
Intel The goals of following Moore’s law are to decrease the cost per component and
reduce the power consumed per component. In 1975, Moore updated his earlier
prediction by forecasting that components per chip would increase by a factor of two every
two years, and that this would come from the combination of scaling component size and
increasing chip area.2 Back in 1965, the industry was producing chips using a minimum fea-
ture size of approximately 50 mm totaling about 50 components. Today’s leading chips use
a minimum feature size of approximately 10 nm and incorporate several billion transistors.
Robert Dennard and colleagues described in 1974 a scaling methodology for
metal-oxide-semiconductor field-effect transistors (MOSFETs) that would deliver consis-
tent improvements in transistor area, performance, and power reduction.3 The methodology
called for the scaling of transistor gate length, gate width, gate oxide thickness, and supply
voltage all by the same scale factor, and increasing channel doping by the inverse of the same
scale factor (see Figure 1). The result would be transistors with smaller area, higher drive cur-
rent (higher performance), and lower parasitic capacitance (lower active power). This method
for scaling MOSFET transistors is generally referred to as “classic” or “traditional” scaling and
was very successfully used by the industry up until the 130-nm generation in the early 2000s.
For the past 20 years, we have been developing new generations of process technologies
on a two-year cadence, and each generation scaled the minimum feature size by approxi-
mately 0.7 times to deliver an area scaling improvement of about 0.5 times (see Figure 2).
Thus, we have been doubling transistor density every two years. But recent technology gen-
erations (such as 14 nm and 10 nm) have taken longer to develop than the normal two-year
cadence, owing to increased process complexity and an increased number of photomasking
steps. Nonetheless, Intel’s 14-nm and 10-nm technologies have provided better-than-normal
transistor density improvements that keep us on pace with increasing transistor density at a
rate of doubling about every two years.
introducing innovations in transistor materials Figure 1. Traditional MOSFET scaling as described by Robert Dennard.
and structure to continue scaling.
One of the first significant innovations
10 10,000
was the introduction of strained silicon tran-
sistors on Intel’s 90-nm technology in 2003.4
This innovation used tensile stain in n-channel
1 1,000
MOS (NMOS) transistor channels to increase
electron mobility and compressive strain in
p-channel MOS (PMOS) channels to increase
Micron
nm
0.1 100
hole mobility (see Figure 3). Tensile strain was ~0.7x every
induced by adding a high-stress film above 2 years
the NMOS transistor. Compressive strain was
0.01 10
induced by replacing the PMOS source-drain
regions with epitaxial SiGe depositions. The
resultant increases in electron and hole mobil-
0.001 1
ity provided increased transistor drive currents 1970 1980 1990 2000 2010 2020 2030
without having to further scale the SiO2 gate
Figure 2. Minimum feature size scaling trend for Intel logic technologies.
oxide thickness. This strained silicon technique
has been adopted by all major semiconductor
companies and continues to be used on the lat-
est 10-nm technologies. on Intel’s 22-nm technology in 2011.6 Tradi-
The need to improve the transistor gate tional planar MOSFETs had been able to scale
dielectric to continue scaling could not be transistor gate length down to about 32 nm to
avoided, and Intel’s 45-nm technology in deliver good performance and density while also
2007 first introduced high-k metal gate tran- maintaining low off-state leakage. But scaling
sistors.5 The traditional SiO2 gate oxide was the gate length below 32 nm was problematic
replaced by a hafnium-based high-k dielectric. without sacrificing either performance or leak-
The high-k dielectric both reduced gate oxide age. A solution was to convert from a planar
leakage current and improved transistor drive transistor structure to a 3D FinFET structure in
current. The traditional doped-polysilicon gate which the gate electrode had better electrostatic
electrode was replaced by metal electrodes with control of the transistor channel formed in a tall
separate materials for NMOS and PMOS to narrow silicon fin (see Figure 5). This improved
provide optimal transistor threshold voltages. electrostatic control provided scaled transistors
The combination of high-k dielectric and metal with steeper sub-threshold slope (see Figure
gate electrodes (see Figure 4) was a revolution- 6a). Steeper sub-threshold slope either provided
ary process change that provided significant transistors with lower off-state leakage or allowed
improvements in transistor performance while threshold voltage to be reduced, which enabled
also reducing transistor leakage current. High-k improved performance at low operating voltage
metal gate transistors are now universally used (see Figure 6b). Operating integrated circuits
on advanced logic technologies. at a lower voltage is highly desired in order to
The next major transistor innovation was reduce active power consumption. All advanced
the introduction of FinFET (tri-gate) transistors logic technologies now use FinFET transistors
SiGe SiGe
(a) (b)
Figure 3. Channel strain techniques used on 90-nm generation transistors. (a) NMOS transistor
using SiN cap layer; tensile channel strain. (b) PMOS transistor using SiGe source-drain; compressive
channel strain.
NiSi
Metal
PolySi
SiO2 High-κ
SiGe SiGe SiGe
SiGe
Silicon Silicon
(a) (b)
Figure 4. Comparison of transistor structures. (a) 65-nm generation transistor using SiO2 dielectric;
polysilicon gate electrode. (b) 45-nm generation transistor using hafnium-based dielectric; metal
gate electrode.
for their good density and superior low-voltage 70-nm transistor gate pitch, 42-nm fin pitch,
performance compared to planar transistors. As 52-nm interconnect pitch, double patterning
Figure 7 shows, when traditional MOSFET scal- techniques, and a 6-T SRAM bitcell area of
ing ran out of steam in the early 2000s, innova- 0.0588 mm2.7 This technology took longer to
tions such as strained silicon, high-k metal gate, develop and get ready for volume manufactur-
and FinFETs were needed, and we must now ing due to the increased process complexity
continually invent new transistor materials and and mask count: about 2.5 years instead of the
structures to continue scaling. normal 2-year cadence. But this technology
also provided better-than-normal area scal-
Recent Logic Technologies ing. Instead of the 0.5 times area scaling that
Intel’s 14-nm logic technology started vol- new technology generations normally provide,
ume production early in 2014. This was Intel’s 14-nm technology provided about 0.37
Intel’s second-generation FinFET technol- times logic area scaling compared to the previ-
ogy, and it used advanced features such as ous 22-nm technology (see Figure 8).
22 IEEE Micro
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on May 14,2025 at 05:19:43 UTC from IEEE Xplore. Restrictions apply.
Gate Gate
Drain
Source
Oxide Oxide
Silicon Silicon
substrate substrate
(a) (b)
Figure 5. Comparison of transistor structures. (a) Planar transistor. (b) FinFET transistor.
10 2.0
Transistor gate delay (normalized)
1 1.8
Channel current (normalized)
1.6
0.1 37%
faster 32-nm planar
Planar 1.4
0.01 FinFET
1.2
0.001 18%
1.0
faster
0.0001 22-nm
Reduced 0.8 FinFET
leakage current
1E-05 0.6
0.0 0.2 0.4 0.6 0.8 1.0 0.5 0.6 0.7 0.8 0.9 1.0 1.1
(a) Gate voltage (V) (b) Operating voltage (V)
Figure 6. Comparison of planar versus FinFET transistor electrical characteristics. (a) Channel
current versus gate voltage. (b) Transistor gate delay versus operating voltage.
90 nm 65 nm 45 nm 32 nm 22 nm 14 nm
2003 2005 2007 2009 2011 2014
Metal
SiGe SiGe
High-j SiGe
SiGe
Silicon
Strained silicon
FinFET
1
technologies have been scaling transistor area,
45 nm 0.49x
and Figure 9b shows the trend of increasing
wafer cost due to increased process complexity.
32 nm 0.45x Figure 9c shows how the cost per transistor con-
tinues to come down due to better-than-normal
Logic area (relative)
24 IEEE Micro
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on May 14,2025 at 05:19:43 UTC from IEEE Xplore. Restrictions apply.
mm2/transistor $/mm2 $/transistor
(normalized) (normalized) (normalized)
Log scale
Log scale
Log scale
45 nm
32 nm
22 nm
7 nm
10 nm
14 nm
45 nm
45 nm
32 nm
32 nm
22 nm
22 nm
7 nm
7 nm
10 nm
10 nm
14 nm
14 nm
(a) (b) (c)
Figure 9. Trends for improving logic transistor area and cost per transistor. (a) Area per transistor. (b) Wafer cost. (c) Cost per
transistor.
32 nm
22 nm
14+ 14++
Transistor performance (log scale)
14 nm 10++
10++
10+ 10++ 10+
Higher 10 nm Better 10 nm
performance 14++ 10+ performance/W 14++
Lower
14+
14+ power
10 nm
14 nm
14 nm
22 nm
22 nm
32 nm 32 nm
2009 2011 2013 2015 2017 2019 2021 2009 2011 2013 2015 2017 2019 2021 2009 2011 2013 2015 2017 2019 2021
(a) (b) (c)
Figure 10. Trends for improving transistor performance and reducing active power. (a) Transistor performance. (b) Dynamic
capacitance. (c) Performance per watt.
is, steeper current turn-on) and can therefore research efforts are exploring logic technologies
operate at a lower power supply voltage than a going beyond CMOS,10 with an objective to
MOSFET. Figure 13 shows drain current ver- complement CMOS rather than to replace it.
sus gate voltage simulation results for nanowire The goal of Beyond-CMOS research is to iden-
TFETs implemented with different III-V semi- tify and enable an integrated circuit technology
conductor materials. that will be more energy efficient than CMOS.
While the success of information technol- If this happens, it will support the continuation
ogy progress in the past 50 years was based on of Moore’s law.
Moore’s law1,2 scaling and mostly one underlying Beyond-CMOS research efforts have been
technology—CMOS transistors—present-day underway for 10 years, being funded in the US
Drain Drain
Gate Gate
Source Source
Oxide Oxide
Silicon Silicon
substrate substrate
(a) (b)
Figure 11. Comparison of transistor structures. (a) FinFET transistor. (b) Nanowire transistor.
1.0E-05
VDS=0.5V, VG Swing = 0.5V
105
InSb In0.7Ga0.3AS,
12Å EOT, 1.6XREXT
InGaAs
Electron mobility (cm2/Vs)
>30X
improvement 1.0E-07 +80%
103 +55%
Simulated:
Si In0.7Ga0.3AS,
8Å EOT, REXT
102 1.0E-08
5x1011 1012 1013 2x1013 0.2 0.3 0.4 0.5 0.6 0.7
(a) Sheet carrier density (cm-2) (b) ION (mA/µm)
Figure 12. Comparison of III-V and silicon transistor electrical characteristics. (a) Electron mobility
versus carrier density. (b) Off-current versus on-current.
S D InAs NMOSFET
Over GaSb/InAs Si NMOS~PMOS
Ef the 1.E+02 TFET
G Ef
S D barrier N-
Vg P-
1.E+01 GeSn Het-j
MOSFET N++ N++
N and P-TFET
ID (uA/um)
Filled 1.E+00
states
dec
de
V/
S D
m
Filters 5nm-NW
<40
1.E-02
63
Fermi tail
G Vds =0.3V
S D Ef Through Ioff =0.1nA/um
Ef the 1.E-03
TFET P++ N++
barrier
Filled Vg 1.E-04
states 0.0 0.1 0.2 0.3 0.4 0.5
Vg (V)
(a) (b)
Figure 13. Comparison of Tunneling FET and MOSFET transistors. (a) Transistor structures and channel current modulation
techniques. (b) Drive current versus gate voltage electrical characteristics.
26 IEEE Micro
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on May 14,2025 at 05:19:43 UTC from IEEE Xplore. Restrictions apply.
104 Co HP CMOS High-performance CMOS
n ASL LV CMOS Low-voltage CMOS
sta CSL
ntE
ne VdWFET Van der Waals FET
rgy
*D STO HomJTFET Homojunction III-V TFET
103 ela
y HetJTFET Heterojunction III-V TFET
GnrTFET Graphene nanoribbon TFET
SpinFET FEFET ThinFET 2D hetero interlayer TFET
HP CMOS
Energy (fJ)
Figure 14. Simulated switching energy and delay for a 32-bit arithmetic logic unit circuit for CMOS
and for various beyond-CMOS device options.
in large part via the Semiconductor Research implement computing technologies. It enabled
Corporation (SRC).11 The expectation of this the setting of expectations for power and perfor-
industry–university research consortium 10 years mance and revealed some pathways for improve-
ago was that this field would produce a comput- ment. Experimental demonstrations have not
ing technology that is better than CMOS for the yet achieved the theoretical modeling projections
majority of its applications. Reality showed that, put forward in the benchmarking. One reason is
among many impressive proposals and demon- that each computing technology requires solving
strations, none of them beat CMOS. However, numerous fabrication challenges.14
they do possess many valuable features, such as The various materials implementations of the
low-power operation and non-volatility. Thus, TFET (see Figure 13) have shown that they have
the current vision is that beyond-CMOS circuits improved energy-delay product (and therefore
will replace CMOS in some critically important power and performance) over the future CMOS
computation or information processing applica- technology node that they are benchmarked
tions. They would be monolithically integrated against (see Figure 14): the International Tech-
with CMOS on the same chip or packaged nology Roadmap for Semiconductors prediction in
together in a multichip module. 2011 of the 2018 CMOS node. With a poten-
Another expectation was that beyond- tial three-times improvement in energy-delay
CMOS circuits would not require any product over CMOS, this is starting to be an
MOSFETs as part of their operation and could interesting device option, and it does not require
maybe even eliminate any charge currents in the a drastic change in circuit design for logic while
quest for energy efficiency. This did not come it offers some additional circuit functionality.
true: a thorough circuit analysis reveals that a The spintronic devices in Figure 14 oper-
MOSFET transistor is needed to supply power ate with a wide range of switching energy and
and for clocking and control of the logic cir- at slower switching speed compared to CMOS.
cuit operation. However, this does not preclude The spintronic devices that match the best
pursuing the key direction of beyond-CMOS CMOS switching energy use magnetoelectric
research, which is to discover and invent com- materials to do the switching of nanomagnets.
putation that can operate at significantly lower Although they are slower than CMOS, they
supply voltages than CMOS to enable dramatic have the added benefit of being non-volatile.
improvements in energy efficiency. Non-volatility in the logic device has the
To this end, beyond-CMOS benchmark- potential to provide energy efficiency benefits
ing12,13 (see Figure 14) was helpful in evaluating by taking advantage of it in the computing
the potential of various materials and devices to microarchitecture.
28 IEEE Micro
Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on May 14,2025 at 05:19:43 UTC from IEEE Xplore. Restrictions apply.
J. Nanoparticle Research, vol. 10, 2008, CMOS Computing,” to be published in
pp. 1–10. Proc. European Solid-State Device Research
12. D.E. Nikonov and I.A. Young, “Overview Conf., 2017.
of Beyond-CMOS Devices and a Uniform
Methodology for their Benchmarking,” Mark T. Bohr is an Intel Senior Fellow in the
Proc. IEEE, vol. 101, no. 12, 2013, pp. Logic Technology Development group at Intel.
2498–2533. His research interests include scaling logic tran-
13. D.E. Nikonov and I.A. Young, “Bench- sistors, interconnects, and memory cells. Bohr
marking of Beyond-CMOS Exploratory received a master’s degree in electrical engineer-
Devices for Logic Integrated Circuits,” IEEE ing from the University of Illinois. Contact him
J. Exploratory Solid-State Computational at [email protected].
Devices and Circuits, vol. 1, 2015, pp. 3–11.
14. I.P. Radu et al., “Spintronic Majority Ian A. Young is an Intel Senior Fellow in the
Gates,” IEEE Int’l Electron Devices Meeting Components Research group at Intel. His
Technical Digest, 2015, p. 32.5.1–32.5.4. research interests include novel embedded
15. S. Borkar, “Electronics Beyond Nano-Scale memory and quantum nanoelectronic and
CMOS,” Proc. 43rd ACM/IEEE Design nanomagnetic devices for energy-efficient inte-
Automation Conf., 2006, pp. 807–808. grated circuits beyond-CMOS. Young received
16. I.A. Young and D.E. Nikonov, “Prin- a PhD in electrical engineering from the Uni-
cipals and Trends in Quantum Nano- versity of California at Berkeley. Contact him at
Electronics and Nano-Magnetics for Beyond [email protected].
www.computer.org/annals