Title: Lecture 3 Power Consumption and Limiting Factors
1Lecture 3Power Consumption and Limiting Factors
- CMOS Deep-Submicron Power Dissipation
- Limits to Low-Power Design near Si material
limit - Plots of limiting factors
- Summary
- Michael L. Bushnell
- CAIP Center and WINLAB, ECE Dept., Rutgers U.,
Piscataway, NJ
2CMOS Power Dissipation Short-Circuit Current
3Short Circuit Current (contd)
4Short Circuit Current (Concluded)
- When input and output have equal rise and fall
times, PSC is small. - If inverter is lightly loaded, so output tr, tf
are shorter than input tr, tf then PSC becomes
comparable to dynamic dissipation - Must make input tr, tf equal to output tr, tf
5Dynamic Power Dissipation
CL VDD2 T
- PD
- Energy transferred/transition
- Ave. Transitions/sec.
CL VDD2 2
2 T
6Alternate Energy/Transition Method
7Alternate Energy/Transition Method
8Capacitance Calculation
- Almost identical to methods in Digital VLSI
Course - Interconnect C combined parallel plate and
fringing C - W, H, L are width, height, length of metal wire
9Limits to Low-Power Design
- Moores Law Transistors/chip grows 1.5 X every
year - Power-delay product (P td) declined by 1/105
since late 1940s - Limits to low-power design
- Fundamental
- Material
- Device
- Circuit
- System
- Practical considerations
10Five Key Principles
- Using the lowest possible supply voltage
- Using the smallest geometry, highest frequency
devices but operating them at lowest possible f - Using parallelism and pipelining to lower
required f - Power management by disconnecting power when
system is idle - Designing systems with lowest requirements on
subsystem performance for given user functionality
11Fundamental Limits
- Due to basic principles of thermodynamics,
quantum mechanics, and electromagnetics,
independent of devices, materials, circuits - Thermodynamics (P1) At any node with an R to
ground, signal power Ps must exceed available
noise power Pavail - g 1 is a constant, en2 is open-circuit
mean-square V across R, k Boltzmanns constant,
T absolute temperature, B node
bandwidth - g 4 recommended, so Ps must be gt 0.104 eV (now
larger by 107 factor)
12Fundamental Limits (contd)
- Quantum Mechanics (P2) Heisenberg Uncertainty
Principle - In order to measure effect of switching
transition of Dt interval, it must involve energy
greater than h/Dt - P h / (Dt) 2
- h Plancks constant
- Electromagnetic Theory (iL2a(t)) Velocity of
high-speed pulse on an interconnect of length L
must always be less than c0 (speed of light in
free space) - t is the interconnect transit time
13Material Limits
- Independent of devices and circuits
- Consider semiconductor cube of undoped material
of dimension Dx, embedded in 3D matrix of such
cubes - Limit on switching energy and time (P3)
calculated in terms of electrostatic energy
stored in cube and transit time of a carrier
through the cube - P ½ em Ec2 ss3 td2
- ss carrier saturation velocity
- Ec self-ionizing electric field strength
- td cube transit time
14Material Limits (contd)
- Heat removal consideration (P4) Fouriers Law of
heat conduction - P p K ss DT td
- K thermal conductivity of semiconductor, A
surface area of heat flow, DT temperature
gradient - For Si, P/td 0.21 W/ns
- For GaAs, P/td 0.69 W/ns (unsuitable)
- Silicon-on-Insulator (SOI) Keq 0.029 KSi, 0.02
KSi, 0.013 KSi (most suitable)
15Material Limits (concluded)
- Interconnect material limit (speed-of-light)
(iL2b(t)) - Propagation time through interconnect of length L
of a material with relative dielectric constant
er must be
16Device Limits
- Independent of circuits.
- Minimum effective channel length Lmin gives limit
on switching energy - E ½ C0 Lmin2 V02
- Consider minimum transition times for Lmin 100
nm (P4) and tox 3 nm (P7) - MOSFET already pushing against Si material limits
- Limit on interconnect of length L RC response
time - Drive by a unit step, so t 0 to 90 response
time - t RC L2
- r / Hr sheet resistance, e / He sheet
capacitance - Limit on minimum interconnect response time
r e Hr He
17Circuit Limits
- Independent of system architecture.
- Must distinguish between logic 0 and 1
- VDD VDD,min
- b between 2 and 4, T 300 K, VDD,min 0.1 V
- This cannot be used because the threshold VT
would be so small that drain leakage in off
state would be unacceptable. Limit of VDD 1 V
appears likely
b k T q
18Circuit Limits (contd)
- Switching energy per transition (P9)
- E P td ½ Cro V02
- Cro total CL of ring oscillator stage
- Intrinsic gate delay time to charge/discharge
load capacitance Cro - td
- Leads to power
constraint - Z channel width
19Circuit Limits (concluded)
- Global interconnect limit (corner to corner)
(iL2c (t)) - Response time
- t (2.3 Rtr Rint) Cint
- Rtr Output resistance of driver, Rint lt 2.3 Rtr
20System Limits
- Chip architecture
- Power-delay product of CMOS technology for the
chip - Heat removal capacity of chip package
- Clock frequency
- Chip physical size
- Selected architecture
- Systolic array of 1024 identical square
macrocells, each one of dimension L - 5-Level clock distribution H-tree
- Maximum Manhattan distance for clock within
macrocell is L, for logic signal it is 2L
21System Limits (contd)
- Limit on logic gate dimension
- Arl1/2 Rrl M
- Gate logic area is logic limited for nw 4 (
wiring levels), pw 0.2 mm (wiring
pitch), ew 0.75 (wiring efficiency factor)
- Rrl 6, M 3, and for other values
- td tdrl
- Heat removal limit average power dissipation of
composite gate P must be less than cooling
capacity of packaging - P QA
- Q package cooling coefficient
- A substrate area occupied by critical path
composite gate
pw ew nw
Tcc ncp
22System Limits (concluded)
scp ncp a
Ccc ncp Crl
- P (1/2 Crl V02)-2 (1 ) td2 (
Q Arl)3 - Locus P11 (t)
- Crl MOSFET diffusion C wiring C gate C
- Ccc corner-to-corner interconnect C
- ncp random logic gates on critical path
- scp gt 1 (clock skew factor)
- Arl random logic area
- Q package cooling coefficient
- a probability that gate switches during a clock
interval
23Practical Limits
- Beyond a certain limit of scaling, cost of
designing, manufacturing, testing, and packaging
will cause the cost/function to level off and
start increasing - transistors/chip N F-2 D2 PE
- F feature size (0.0625 mm in 2020), D chip
area in 2020 (gt 50 mm2), PE packaging
efficiency in 2020
(1 transistor/minimum feature area) - Leads to possibility of 100 billion transistor
chips
24P Limits Plotted
P7 Device Transition Time P8 Missing P9
Transistor Switching Energy P10 Missing P11 Heat
Removal P12 Missing
P1 Thermodynamics P2 Heisenberg Uncertainty
Principle P3 Electrostatics P4 Heat Conduction P5
Missing P6 Missing
25Interconnect Limits Plotted
Speed of light Speed of light Wire length Missing
26Summary
- CMOS Deep-Submicron Power Dissipation all 3
factors (dynamic, short circuit, and leakage) are
now important - There are limits to Low-Power Design we are
near Si CMOS material limit