Transcript MOSFET SCALING

Potential Ways forward: Nanoelectronics
> 15 nm
Strained Si/SiGe
~ 10 nm, CNT
(ckts challenging)
~ 5 nm, SiNW
(Low mobility)
Multiple gates
Better architecture
(Moore’s Law, ~5-7 yrs)
Reversible Computing
Tunneling transistors
Neuromorphic Computing
Non-equilibrium Switching
Spintronics, Q. Computing
MQCA
Reconfigurationable MQCA
(RAMA)
Excitonic BEC (BisFET)
~2 nm, Org. Molecules
(Reproducibility, Gating)
Better Materials
(More Moore, ~7-15 yrs)
New Physics of Computation
(Beyond Moore >15 yrs)
Various FETs – real and proposed
Many of them are dominated by QM flow of electrons
(Beyond Drift-Diffusion)
NRI’s chart
Must understand Performance Metrics
How to compare technologies?
For logic: Switching delay vs ON/OFF ratio (error rate)
Various Candidate switches
Data plotted by Kerry Bernstein, IBM
What about CMOS with new materials?
The material ‘zoo’ !!
2 nm
Silicon Nanowires
(Low m < 100 cm2/Vs)
5 nm
S
O
U
R
C
E
Organic Molecules ?
(Reproducibility/
Gateability)
D
R
A
I
N
INSULATOR
Source
VG
VD
I
CNTs (m ~ 10,000cm2/Vs)
Hard to align into a circuit!
< 10 nm
Top Gate
Drain
Channel
Bottom Gate
Strained Si, SiGe
(m ~ 270cm2/Vs)
15 nm
Just different DOS, m’s and C’s?
Theory: Tseng
Expt: Moon
THINGS TO WORRY ABOUT
Electrostatics more Complicated
2D electrostatics of planar source/drain in ribbons
S
O
U
R
C
E
D
R
A
I
N
INSULATOR
VG
VD
3D coaxial electrostatics for wires/tubes
Energy Quantization ‘depletes’ charge
Electron is removed from surface
Decreases Gate Cap, surface scattering
Quantization of other parameters (‘speed’)
Transmission Line with a difference !
Quantum R = h/2e2
Influences electron speed/current
Quantum C = 4pme2/h ~ 1/a0
Influences electrostatics/charge
Quantum L
Influences RF properties/signal speed
Atomicity and Chemistry
Atomicity and Chemistry: Can help !!
Nanotube Data
Williams group
Characterizing single molecular
traps using noise patterns (UCLA)
Ballistic Quantum Flow of electrons
FET currents today at ~50% of ballistic limit
SiNW currents today at ~90% of ballistic limit
Beyond Newtonian Drift-Diffusion
Ballistic Flow (what is mobility?)
ECE 5502/6502
Datta, Purdue
QM: Tunneling – can hurt or help ..
Tunneling makes oxides
ineffective
But not restrained by ‘kT’ !!
So, Subthreshold swing can
be small  low power?
QM Wave flow (Interference)
Quantum Computation
schemes rely on similar
issues (“Entanglement”)
Scattering and Blockade
Pauli exclusion and Coulomb’s Law (Poisson)
An electron should not feel a potential due to itself, ignored in U ~ q2/C
A BIGGER CHANGE – of Emphasis !!
Is this difference
in perspective
just an intellectual
exercise?
DOES INDUSTRY
CARE?
TOP DOWN
ECE 6163
BOTTOM UP
ECE 6502
Newtonian Physics
Bulk Bands
Quantum Physics
Atomistics
Change in Tech Emphasis
(electrostatics Es to energetics EN )
Most developments in MOSFETs
were about controlling the fields
Emerging MOSFETs
beset by energetics
LDD, Halo, high-k, …
Pentium 2000, 50W/cm2
2025 > 40 MW/cm2 !!
Energy dissipated in bit switching
VDD
Gate
Source
Drain
Channel
Eb = MkTln(pon/poff), M = 1 + CD/Cox
Fundamental limit: pon/poff = 2, M = 1
Fundamental switching energy per bit
Eb = kTln2 (Landauer, Shannon)
To repeat operate and give directionality
Eb = 3kTln2 (Zhirnov et al)
Boltzmann
Fundamental vs Practical Limits
Boltzmann
Heisenberg
We dissipate ~ 10,000 kT today !!!
Why?
(1) Actual circuit has 300 million devices
 pon/poff ~ 105  Eb = 5hkTln10
Coulomb
Carnot
(2) h becoming larger (Cox )
(3) Not just one bit but many!!
22 nm node (2016)
n ~ 3 x 109/cm2
t = CV/I ~ 150 fs
P = 93 W/cm2
K. Roy group
Importance of Subthreshold swing
Impact Ionization
(~5mV/decade !)
Reliability issues
S = 60mV/decade
For ON-OFF ~ 105
need VG ~ 5 x 60 = 300 mV/dec
S determines switching voltage and energy
NEMFETs
(0.1mV/decade !)
Slow, hysteresis
Controlling S
• Increase ‘q’ by collective motion (e.g. relay)
S-1 = Sel-1 + Sconf-1
Ghosh, Rakshit, Datta, NL ’03
• Effectively reduce N through interactions
Salahuddin, Datta
More generally:
• Negative capacitance (M < 1)
Salahuddin, Datta
Vb = NMkBTln(pon/poff)
q
• Reduce “T” through Tunneling
Appenzeller et al, PRL
• Nonequilibrium switching
pon/poff ≠ e
qVb/kBT
• Other state variables (spin, strain etc): q  “Q” ?
Decreasing S is a ‘Holy Grail’ in switching
Search for mV switch
Reduces power dissipation
Also
Makes computation and communication compatible…
(Yablonovitch)
Pulse Shaping (Adiabaticity)
Ionic flow in
neuronal axons
Dissipation comes from
sources faster than RC
(Phase lagging charges dissipate)
So, chop off fast components
Signal propagates
by paying energy cost
Nonequilibrium Switching
Moving electrons without
a drain bias by using an
energy source away from
equilibrium (e.g. KINESIN)
This is energy efficient
Can do logic with this!
Emerging Frontiers:
Si Computers vs C Computers
C computers:
• Use heavier bits (no tunneling)
 Better packing
• Utilize Thermal energy
(Eliminating it is costly!)
• Communication distances small
Ralph Cavin, SRC
“In Carbo” vs “In Silico” – More energy efficient
Better Algorithms
Parallel Algorithms
Amdahl’s Law
Better Architecture (Fanout, 3D)
FO: 4
FO: several thousand !
ECE 663