Upload
others
View
22
Download
0
Embed Size (px)
Citation preview
1. CMOS Scaling
Dmitri Nikonov
Thanks to Kelin Kuhn
Nikonov 1. CMOS
Beyond CMOS computing
1
Outline
Moore’s law = scaling
Performance improvement with scaling
Latest: tri-gate transistors
Fundamental limits to scaling
2 Nikonov 1. CMOS
Moore’s Law
Nikonov 1. CMOS 3
Transistor size becoming smaller
Moore’s Law
Nikonov 1. CMOS 4
Gordon Moore becoming wiser
Scaling Falsifies Predictions
Nikonov 1. CMOS 5
IEDM Plenary Session 1980 (Broers)
How Far Scaling Went
Nikonov 1. CMOS 6
1980 SRAM Cell: 1700 um2 22nm SRAM Cell: 0.092 um2
Small enough that a 2011 22nm SRAM cell is dwarfed by a 1980 SRAM cell CONTACT
10000X
K. Kuhn, MIT invited seminar (MTL), 45nm High-k + Metal Gate Logic Technology, 5-19-08 (images from archives Mark Bohr, 2007)
Metal Interconnects
Nikonov 1. CMOS 7
9 levels of metal
Contacted Gate Pitch
Nikonov 1. CMOS 8
Transistor gate pitch continues to scale 0.7x every 2 years.
Proves to be 4*F. F is the label for process generations.
M. Bohr, ISCC, 2009.
Economics of Moore’s Law
Nikonov 1. CMOS 9
“Doubling of number of transistors per chip every 2 years”.
Lowers cost per transistor.
Self-fulfilling prophesy.
Original paper: G.E. Moore, Electronics 19, 114 (1965)
103
104
105
106
107
108
109
1010
’70 ’80 ’90 ’00
10
10-6
10-5
10-4
10-3
10-2
10-1
10-7
$ per
Transistor
As the
number of
transistors
goes UP
Transistors
per Chip
’10
Price per
transistor
goes DOWN
100
Year
Classic Scaling
Nikonov 1. CMOS 10
R. Dennard, IEEE JSSC, 1974
Device or Circuit Parameter Scaling Factor
Device dimension tox, L, W 1/κ
Doping concentration Na κ
Voltage V 1/κ
Current I 1/κ
Capacitance εA/t 1/κ
Delay time/circuit VC/I 1/κ
Power dissipation/circuit VI 1/κ2
Power density VI/A 1
Classical MOSFET scaling was first described by Dennard in 1974
Dennard, IEEE JSSC, 1974
Process Evolution
Nikonov 1. CMOS 11
Inflection in Scaling
Nikonov 1. CMOS 12
THEN
Scaling drove down cost
Scaling drove performance
Performance constrained
Active power dominates
Independent design-process
NOW
Scaling drives down cost
Materials drive performance
Power constrained
Standby power dominates
Collaborative design-process
65nm 45nm 32nm
Images from Bai, Mistry, Natarajan, Auth IEDM/VLSI, 2004/7/8/12 (see course required reading)
22nm
Short Channel Effects (SCE)
Degradation of short channel effects
V2 > V1
Decreasing L
ID
VG
~DIBL (mV/V)
~ 1 / Subthreshold
Slope (mV/dec)
13
Gate Voltage (V)
“On”
Current
“Off”
Current
Channel
Current (normalized)
Basic ID-VG
Poor
SCE
Reduced Threshold
Voltage
14
Good
SCE
Electrostatics Benefits
Threshold Voltage
9
~DIBL (mV/V)
~Subthreshold
Slope (mV/dec)
~VTlin
http://en.wikipedia.org/wiki/Threshold_voltage
On Current, Off Current
9
IDsat
Ioff
~DIBL (mV/V)
~Subthreshold
Slope (mV/dec)
~VTlin
IDsat IDsat
Idsat (=Ion)
Io
ff
Performance from Scaling
Nikonov 1. CMOS 17
Natarajan, Intel, IEDM, 2008
Where Performance Comes From
Nikonov 1. CMOS 18
K.J. Kuhn, Moore's Law past 32nm: Future Challenges in Device Scaling, Solid State Devices and materials conference, plenary, October 2009
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
90 65 45 32 22
PM
OS
Ieff
@ 0
.7V
(No
rmal
ize
d)
Generation (nm)
Tri-gate
HK-MG
Strain
Classic
Strain and High-k + metal gate are key enablers past the 90nm node
Intel Announced First Tri-Gate (2011)
Nikonov 1. CMOS 19
Close-Up on Tri-gate Transistors
Nikonov 1. CMOS 20
Nomenclature of Non-Planar Devices
Nikonov 1. CMOS 21
Where Non-Planar Can Go?
Nikonov 1. CMOS 22
Working With Atomic Dimensions
Nikonov 1. CMOS 23
International Technology Roadmap for Semiconductors
Nikonov 1. CMOS 24
Limits of Computing
Moving from fundamental limits given by the laws of
physics to practical limitations. These limits tend to
be broken. Fundamental limits
Material limits
Device limits
14nm inverter
neuron
Interconnect Limits Transistor Limits
Meindl, Proc. IEEE 83, 619 (1995).
Meindl, Chen, Davis, Science 293, 2044 (2001).
Meindl, J. Vac. Sci. Technol. B 14(1), 192 (1996).
Nikonov 1. CMOS 25
Equations for Fundamental Limits
Meindl, Proc. IEEE 83, 619 (1995).
Thermodynamics: if energy is less
than thermal – bit errors
4sw
E kT
Quantum Mechanics: energy time uncertainty
sw swE t h
Relativity: Signal no faster than the
speed of light
0/L c
Nikonov 1. CMOS 26
Better Fundamental Limits
* Zhirnov et al., Proc. IEEE 91, 1934 (2003) and Nikonov and Bourianoff, JSNM 21, 497 (2008).
2 2
2
3
2sw
Ema
3sw
E kT
sw swE t
3.8nma 27fssw
t 20
1.2 10 J 78meVsw
E
a a a
a
a
source gate drain
electron
OFF
ON
electron
Thermodynamic limit on bit error ratio
Higher energy = faster switching
Energy of electron in a transistor
is limited by quantum confinement
Gate raised = confined in the source.
Gate lowered = can travel to drain.
Solving equations together gives
limits
Nikonov 1. CMOS 27
MOSFET Scales Towards the Limit
* Data courtesy of Robert Chau (Intel)
Current CMOS device scaling close to the ideal limits
0.01
0.1
1
10
100
0.001 0.01 0.1 1LGATE (mm)
Gate
Delay
(ps)
0.000001
0.00001
0.0001
0.001
0.01
0.1
1
10
100
0.001 0.01 0.1 1
LGATE (mm)
Switching
Energy
(fJ)
Nikonov 1. CMOS 28
How long is left for Moore’s law
Intel’s generation to HVM
2013 14nm
2017 7nm
2021 3.5nm
ITRS start of production
2012 32nm
2018 15nm
2024 7.5nm
2030 3.8nm
Scaling might end between 2021 and 2030
But it is NOT the end of Moore’s law:
better architectures, 3D circuits.
Nikonov 1. CMOS 29
Summary
Moore’s law = 0.7 size every 2 years
Despite trends, Intel developers manage to improve performance
Tri-gate transistors = major advance
Fundamental laws limit size scaling to ~4nm
30 Nikonov 1. CMOS