Beruflich Dokumente
Kultur Dokumente
Massoud Pedram
Dept. of EE
University of Southern California
Outline
Background
Noise in Digital Systems
Capacitive Crosstalk
Inductance Effects
Ground Bounce
IR Drop
Skin Effect
Electromigration
EMI
M. Pedram
On and Off Chip Clock Frequencies
M. Pedram
Sources of Noise
M. Pedram
Noise in Digital Systems
Noise Crosstalk
Signals may be corrupted due One signal interfering with
to many factors: another signal
Power supply noise Capacitive cross talk
Crosstalk between RC lines on a chip
Inter-symbol interference Floating
Real noise (thermal and shot) Driven
Parameter variation
Coupling between LC
Independent random variables transmission lines
Power supply noise Near end
M. Pedram
Noise Sources
VN K NVs VNI
The first component represents those noise sources that are proportional to signal
amplitude e.g., crosstalk and signal-induced power supply noise
The second component represents noise sources that are independent of signal
amplitude e.g., transmitter and receiver offsets and unrelated power supply noise
M. Pedram
Proportional and Independent Noise
Sources
Some of the noise sources Some noise is independent
are proportional to signal of the signal swing
swing
If you increase the signal Receiver or transmitter
swing, you increase the noise sensitivity
Crosstalk Receiver or Transmitter
Inter-symbol interference offsets
Signal return ratio (Zr/Z0) Independent power supply
Signal-induced power supply noise
noise
Reference offsets
Need to cancel these sources
M. Pedram
Power Supply Noise
The power supply network
has parasitic elements
On-chip : resistive
Off-chip: inductive
Current draw across these
elements induces a noise
voltage
I
RI L
t
Instantaneous current is
what matters
Can be many times larger
than the DC current
10W chip draws 4A at 2.5V
Peak current maybe 10-20A
M. Pedram
High Speed Challenges
Signal quality of a single net: reflections and
distortions from impedance discontinuities in
the signal or return paths
Crosstalk between multiple nets: with
ideal return paths
M. Pedram
Don'ts and Do’s …
Don’ts:
Base a final design on a rule of thumb estimate or even an
analytical approximation
Use rules of thumb and analytical approximations that are
inconsistent or lack fidelity
Automatically and immediately reach for a field solver in
every case
Do’s:
Know your rules of thumb
Know where to locate a range of analytical approximations,
each representing a non-dominated tradeoff point in terms of
accuracy and efficiency
Check rule of thumb, analytical expression and field
simulation for acceptable consistency and high fidelity
Use a field solver where and when absolute accuracy is
important
M. Pedram
Maxwell’s Equations
M. Pedram
Capacitive Crosstalk
What Is Crosstalk
Crosstalk is a disturbance caused by the electric or
magnetic fields of some signal affecting another
signal in an adjacent circuit
In an telephone circuit, crosstalk can result in your
hearing part of a voice conversation from another
circuit.
The phenomenon that causes crosstalk is called
electromagnetic interference (EMI). It can occur in
microcircuits within computers and audio equipment
as well as within network circuits
M. Pedram
Capacitive Crosstalk to a Floating Line
Capacitive coupling between
on-chip lines
Coupling over shared signal
returns K2C
2Cc
2Cc Co
On adjacent layers
Cc
KC Va KC (Vb Vc)
Cc Co
M. Pedram
Capacitive Coupling to a Driven Line
M. Pedram
Miller Effect and Wire Delays
Capacitive crosstalk can affect the RC delay of signals
propagating down the line
If aggressor(s) switch in opposite direction of the victim, then
the effective capacitance is double the coupling capacitance
(due to the “Miller effect”) i.e., Ceff = 2Cc
If aggressor(s) switch in the same direction, the effective
capacitance becomes zero (capacitive crosstalk is eliminated)
i.e., Ceff=0
This can result in a large variation in the wire propagation
delay
It is also a major cause of timing noise (skew and jitter) in
VLSI circuits
M. Pedram
Parallel Striplines
M. Pedram
Devgan’s Metric
Aggressor
Ramp input
voltage source
Victim
Cij
Vn V par (n) Rn
tr , net ( j )
idesc(n)
jadj (i)
It determines an upper bound on the peak node voltage in the victim line in
terms of:
Coupling capacitance, Cij
Interconnect resistance (plus source resistance, if present) of the victim net, Rn
Rise time of at the output of the aggressor net driver, tr,net
M. Pedram
Example Calculation of Devgan’s Metric
Agb
N2
Aga
N1
N0
Agc
N3
R2 N2
R1 C2b
N0 N1
tr , net (b)
R3 N3
C1a
R0=Rs
tr , net (a) C3c
tr , net (c)
M. Pedram
Example of Devgan’s metric (cont’d)
R2 N2
R1 C2b
N0 N1
tr , net (b)
R3 N3
C1a
R0=Rs
tr , net (a) C3c
tr , net (c)
C C2b C3c
V1 ( R1 R0 ) 1a
tr , net (a) tr , net (b) tr , net (c)
C2b C3c
V2 V1 R2 ; V3 V1 R3
tr , net (b) tr , net (c)
M. Pedram
Matrix Equations for Devgan’s Metric
where
M. Pedram
Matrix Equations (Cont’d)
M. Pedram
Pros/Cons of Devgan’s Metric
CT Lij
Vn V par (n) Rn
tr , net ( j ) Sij
idesc(n)
jadj (i)
Advantages
Convex and separable in Sij
The optimal spacing problem (OSP) becomes a convex and separable
mathematical program
Disadvantages
No self capacitance or aggressor net resistance
Unbounded behavior for decreasing rise time and/or increasing coupling
length
M. Pedram
Vittal’s Approach
Gives a generalization of the Devgan’s metric
Accounts for the self capacitance and
resistance of the aggressor net
No unbounded behavior
Replaces the “driver rise time of the aggressor net”
with an “effective rise time of the aggressor net as
seen by the victim net” as follows:
tr , net ( j )
tr , net ( j ), net (n) Dnet ( j ) Dnet (n)
2
M. Pedram
Vittal’s Metric
CT Lij
Vn V par (n) Rn
tr , net ( j ), net (n) Sij
idesc(n)
jadj (i )
M. Pedram
Kuhlmann’s approach
M. Pedram
Kuhlmann’s approach (cont’d)
Model the input voltage of the aggressor as an exponential function
Next we use V2(t) in the noise matrix equations which shows better results
than the Devgan’s metric
M. Pedram
Capacitive Crosstalk Countermeasures
Routing rules should be put in place to limit the
magnitude of the capacitive crosstalk between any
pair of signals
Floating signals should be avoided, and keeper
devices should be placed on dynamic signals to
reduce susceptibility to cross talk
Signal rise time should be made as long as possible,
subject to timing constraints, to minimize effect of
crosstalk on driven nodes
Sensitive signals should be separated from full swing
signals or even shielded by conductors on either side
that are tied to power or ground
M. Pedram
Coupled Transmission Lines
M. Pedram
Crosstalk between Transmission Lines
A signal transition on one
transmission line induces forward
and reverse traveling waves on
adjacent transmission lines
tx=dx/n tx
Near-end crosstalk:
VC ( x, t ) krx VA ( x, t ) VA ( x, t 2t x )
Far-end crosstalk:
VA ( x, t )
VD ( x, t ) k fxt x
t
M. Pedram
Crosstalk (Cont’d)
Inductive and capacitive
coupling add at the near end of
the line
Both waves are positive
Pulse begins at beginning of
coupled section
Inductive and capacitive
coupling subtract at the far end
of the line
In a homogeneous medium
cancellation is exact
Narrow pulse coincident with
wave on aggressor
Reverse and forward coupling coefficients:
k klx k klx
krx cx k fx cx
4 2
Cc M
kcx ; klx
Cc Co L
M. Pedram
TX Line Crosstalk Countermeasures
High swing signals should not be routed on
lines immediately adjacent to low swing or
other sensitive signals
The capacitive and inductive coupling
coefficients should be matched to eliminate
forward (far-end) crosstalk
Place all signal lines between a pair of return
planes
If the forward coupling coefficient is nonzero,
avoid long parallel runs of transmission lines
When possible, both ends of all transmission
lines should be terminated in the characteristic
impedance of the line
M. Pedram
Countermeasures (Cont’d)
Signal rise times should be made as long as possible
subject to timing constraints
This reduces far-end crosstalk, which is directly proportional
to the signal derivative and this is reduced directly as the
derivative is reduced
Near-end crosstalk is proportional to the maximum difference
in the signal level across the coupled length and thus is
reduced proportionally once the rise time is greater than tx,
the coupled time
Signals on adjacent layers should be routed in
perpendicular directions
This results in zero inductive coupling and negligible
capacitance coupling
The two lines of a differential pair should be spaced
close together and far from the lines of other pairs to
keep the fields local to the pair
M. Pedram
Inductance Effects
Introduction
On-chip inductance effects have become increasingly significant
because:
For performance considerations, some global signal and clock wires are
routed with large widths and thickness at the top levels of the metal to
minimize delays. This decreases the resistance of the wires, making their
inductive impedance comparable to the resistive part.
Z R jL
As the clock frequency increases and the rise times decrease, electrical
signals comprise more and more high-frequency components, making the
inductance effects more significant.
With the increase of chip size, it is fairly typical hat many wires are long and
run in parallel, which increases the inductive cross talk and delay.
With the push of performance, some low-resistively metals,e.g. Cu wires,
have been explored to replace Al in order to minimize wire RC delays. This
could make the wire inductive reactance larger than the resistance.
Examples of Inductance effect
Overshoot/undershoot edges
Ldi/dt voltage drop
Long range cross talk
Frequency dependent resistance
M. Pedram
Adverse Effects of the Inductance
Signal quality
Inductive discontinuity and ringing
Ground bounce
Collapse of the power and ground rail voltage
Crosstalk
Distributed mutual inductance in uniform
transmission lines and lumped mutual inductance
(SSO)
Electromagnetic interference (EMI)
Generation of common mode noise currents
M. Pedram
Clock Frequency vs. Equivalent Inductance
At 1 watt
power
dissipation, to
achieve 100
MHz clock
requires an
equivalent lead
inductance on
the order of
half a nH
M. Pedram
What Affects the Magnetic Field Lines
Magnetic field lines exist around
all current carrying conductors
M. Pedram
Types of Inductance
Inductances may be classified as:
Self inductance
Mutual inductance
Or they may be classified as:
Partial Inductance
Loop inductance
Loop self inductance
Loop Mutual inductance
Effective (or total) inductance
M. Pedram
Self and Mutual Inductances
Inductance is related to the number of field lines around the
conductor per ampere of current
Different types of inductance:
Self or mutual
Loop or partial
Effective (total)
Self inductance
# of field lines/amp of its own current
Mutual inductance
# of field lines/amp of another line’s current
M. Pedram
First-order Equation for Self Inductance
Radius=r
m0 d 2 d 3
Lself ln
2 r 4
b
m d 2 d 1
Lself 0 ln ln( )
2 bc 2
Note that is obtained from lookup table. For more information visit:
http://home.san.rr.com/bushnell/self_inductance.htm
M. Pedram
First-order Equation for Mutual Inductance
Radius=r
m d 2d
s d Lmutual 0 ln
2 s
2
m d 2d s s
sd Lmutual 0 ln 1
2 s d 2d
M. Pedram
Voltage Across the Inductance
V
When the number of field lines around a conductor changes,
a voltage is induced:
N (LI ) I dI
V L L
t t t dt
M. Pedram
Partial Inductance
Inductance of
this part
Wire a La
LM
Wire b (return path)
Lb
M. Pedram
Loop Self Inductance
The loop inductance accounts for the presence of the return
path current
M. Pedram
First-order Equation for Loop Self
Inductance
signal Radius=r
d
m0 d s
Lloop, self ln( )
r
d
s
Lloop, self m0 d
M. Pedram
w
Loop Mutual Inductance
M. Pedram
First-order Equation for Loop Mutual
Inductance
a
b
h
m0 a 2b 2
Lloop, mutual
2 3/ 2
a 2 h2
For more information visit:
http://www.physics.uq.edu.au/people/ficek/ph348/sols/sol4/node4.html
M. Pedram
Induced Voltages due to a Fixed dI/dt
D
self dI a
R Va, noise Lself
dt
a
a
mutual dIb
Va, noise Lmutual
b
dt
M. Pedram
Effective Inductance
Effective (a.k.a. total or net) inductance is the total
number of field lines per ampere of current which are
around a section of a loop
Wire a La
LM
Wire b (return path)
Lb
LM
Wire b(return path)
Lb
La LM Lb LM La Lb 2 LM
M. Pedram
Ground Bounce
Ground Bounce
Digital logic requires a stable, quiet, DC supply voltage while drawing a
large AC current with very high-frequency components comparable to
signal rise times (about 200A from power supply and derivative at the
point of use over 200GA/s)
The supply and ground are distributed over a network with inductive
and resistive components
The current causes IR drop across the resistive component
The derivative of current causes Ldi/dt across the inductive components
M. Pedram
Example of Ground Bounce Effect
M. Pedram
Other Names for Ground Bounce
Other names are:
I noise
Switching noise
dI/dt noise
SSO (simultaneous switching output) noise
SSN (simultaneous switching noise)
M. Pedram
A Typical Power Supply Network
M. Pedram
Typical Load Current
For a given clock domain, load is usually periodic with the clock
May stop or start in a single cycle
With multiple clock domains, they may drift into phase
reinforcing one another
Load is often resistive, varying linearly with supply voltage
Some loads are high impedance, constant independent of supply
voltage
M. Pedram
Local Loads and Signal Loads
Local loads pass current from a
point in the local power network
to one in the local ground network
Current can be supplied from a
nearby bypass capacitor
Signal loads connect a point in the
power network via a signal lead to
a distant point in the ground
network
Usually due to unbalanced signaling
Current must return over a long path
Bypass capacitors are not effective
M. Pedram
Inductive (Off-chip) Power Supply
Noise
Each section of the supply network is an LC circuit
1/ 2
Has a resonant (natural) frequency, LC
LC
Inductor carries DC current ( LC)
Capacitor supplies AC current ( LC)
At LC, the LC impedance is infinite, thus, a small current
will cause large voltage oscillations
Size the bypass capacitor to
Supply cycle to cycle AC current with acceptable ripple
Handle inductor start/stop transient
M. Pedram
Response of an LC Section to Typical
Supply Current
Over a clock
cycle,
inductor
current is
essentially
constant, Iavg ;
Load current
varies
considerably;
Capacitor
current is the
difference
between the
two;
Capacitor
voltage
ripples due to
this AC
current
M. Pedram
Bypass Capacitor to Handle Normal Load
in a Cycle
max
M. Pedram
Putting It Together: Sizing the Bypass
Capacitor
Approximate capacitance
requirement by summing the
independent requirements
M. Pedram
The Truth about Bypass Capacitors
Most capacitors are only capacitors
at low frequencies
Capacitors have parasitic series
resistance and inductance
Every pico-farad has its very own
nano-Henry
Two key breakpoints
LC frequency
RC frequency
Capacitors are ineffective at
bypassing currents above either of
these frequencies
M. Pedram
Impedance vs. Frequency for Typical
Capacitors
M. Pedram
Controlling the Ground Bounce
Wire a La
LM
Wire b(return path)
Lb
Ground bounce
dIb dI
Vb Leff ,b ( Lb LM ) b
dt dt
M. Pedram
IR Drop
Power Distribution Trends
The on-chip power distribution problem is
getting much harder as technology evolves
Combination of
lower voltages
higher current density
thinner metal layers
larger chips
We are quickly approaching the point where
peripheral bonding will not be adequate for
high-performance chips
M. Pedram
Logic Current Profile
Why does on-chip logic produce a ‘ spikey’ current profile?
Consider the logic that generates the current
Current is drawn to charge gate and wire capacitance Q=CV,
E=0.5 CV2
Typical behavior includes
circuit idle before clock edge
very little current
exponential clock amplification just before clock edge
exponential ramp up in current
flip-flops are clocked
current depends on activity factor
Fanout in a logic circuit
exponential ramp up in current
Fan-in or selection in a logic circuit
drop in current
M. Pedram
Logic Current Profile Examples (Con’t)
A memory or register array
fanout in decoder
selection of word-line
amplification in S/A
selection in multiplexer
A carry-lookahead adder
modest fan-out in PGK blocks
fan-out in carry tree
fan-in in carry tree, but with amplification
amplification to drive output
Control logic
much like the decoder
fan-out in the first few stages
then tree fanning in to flip-flops
often a long, serial tail
Overall current profile is the superposition of all of these profiles
M. Pedram
Worst-Case vs. Average Logic Current
Profile
Current drawn is often very data dependent
e.g., a data path may switch 64-bits from all 0s to
all 1s
on average only 1/4 of the bits will have this
transition
when there is a transition at all
For noise analysis we must consider worst-
case power
cannot allow possible, but unlikely events to cause
system failure
For battery life we may consider average
power
M. Pedram
IR Drops
Power distribution network is designed to keep IR drop on VDD
and GND networks within limits
e.g., for 10% supply variation, can drop at most 5% on each supply
Networks are usually designed specifically for the loads of a
given chip. However, we can gain insight into the process by
considering a uniform load
For example, suppose current densities are
– J peak = 0.3A/mm 2
– J avg = 0.05A/mm 2
For a peripheral bonded chip, VDD and GND are usually
distributed by combs with interdigitated fingers
– a hierarchy of such combs is often used
How much of a metal layer (or how many layers) do we need to
distribute this power?
M. Pedram
Power Distribution Network: The Top-level
View
GND
VDD
M. Pedram
Calculation of the IR Drop
2 i J pk L p rw
VIR i J pk Ap R p
4N 2k p
i 1 i 1
Lp / 2 J J pk rw L p 2
pk rw x
VIR dx
0 kp 8k p
M. Pedram
IR Drop Countermeasures
Use an area-bonded chip so that power
need not be distributed from the chip edge
Use more or thicker metal layers to
distribute the power to reduce the
effective resistivity, rw
Use on-chip bypass capacitors to flatten
the current profile so the distribution need
only be sized for average rather than the
peak current
M. Pedram
Skin Effect
Transmission Line Loss Models
Losses that remain constant with
frequency can be modeled as:
Dielectric constant
RDC
M. Pedram
Signal Propagation in the Frequency
Domain
M. Pedram
Low Loss Approximation
At low frequencies (towards the DC end), current is uniformly
distributed over the cross section of each conductor. We will refer to
these frequencies as the low-frequency region
As the frequency increases, due to the induced electric field, the
current distribution starts changing. There are three reasons :
Edge effect: the current tends to concentrate at the sharp edges of a
conductor which affects both the signal conductor and the ground
conductor
Proximity effect: particularly pronounced on the ground conductor, the
current of which tends to concentrate below the signal conductor as the
frequency increases
Skin effect: pronounced in the high-frequency region on both conductors,
where the current becomes concentrated in the thin layer at the surface
Microstrip
Stripline
M. Pedram
Skin Effect
The tendency of alternating current to flow near the surface of a
conductor, thereby restricting the current to a small part of the
total cross-sectional area and increasing the resistance to the flow
of current
High frequency current flows primarily on the surface of a
conductor with current density falling off exponentially with
depth into the conductor:
d
J exp
d
The skin depth, d, the depth where the current has fallen off to
exp(-1) of its normal value is given by:
1/ 2
d f m
M. Pedram
The Origin of Skin Depth
At DC, if each ring has the same current, which one has more
inductance?
At high frequency, which ring has more impedance?
If you were an electron, where would you want to be?
The center path has the highest inductance since it has all the field lines
of the outer ring plus the field lines that are between it and the outer
ring. In general, the self inductance of the current path will decrease as
you approach the outer edge
M. Pedram
Skin Depth Limited Current Distributions
M. Pedram
A Simple Approximation
Assume the signal path width is w and its thickness is t
M. Pedram
Resistance Modeling
M. Pedram
Inductance is Frequency Dependent
M. Pedram
Electromigration and
Electromagnetic Interface
Electromigration
Over long distances, current density must be kept low to avoid large IR
drops; Over short distances, current density must still be kept low to
avoid metal migration
Over time, wires that carry high current densities will fail as the metal is
eroded away
think of your wire as a fuse
Migration threshold varies depending on process, temperature, and
lifetime
If a metal layer is subjected to an average current density of greater
than about 1mA/mm2 (109 A/m2), the layer will slowly fail
This is often a factor on short power buses that connect from the main
bus to a point of high current use
To avoid metal migration the fraction kp of metal devoted to each
power bus must satisfy: 3
10 J aw Lp
kp
2t
where t is the thickness of the metal layer, Jaw is in A/mm2
It can also be a factor on the output of high-current drivers
Migration applies to vias as well as wires
M. Pedram
Electromagnetic Interference (EMI)
Large external electric and magnetic fields can couple into circuit nodes
and cause noise
For both types of fields, susceptibility is reduced by keeping a minimum
distance between a signal and its return
For E fields, the induced voltage is proportional to the distance between
the signal and return path
For B fields , the amount of flux coupled is proportional to the area of
the loop between the signal and the return path
Electromagnetic Interference is rarely a problem in digital circuits unless
they are operated near large sources of fields
Usually, the larger concern is the interference created by the digital
system
The same techniques that reduce susceptibility to external fields also
reduce emissions by the digital system
Most digital systems can be made “quiet” naturally by keeping signals
and return paths close together, balancing currents in external cables,
controlling rise time on signals, and properly terminating lines
M. Pedram
Appendix: How to Minimize Inductance
Effects
Dedicated ground wires
M. Pedram
How to Minimize Inductance Effects
(Cont’d)
Buffer Insertion
Rl Ll Cl
, , ; k 1
Splitting wires n kn / 2 n
G S S G G S G
G S S S S G G S S S S S S S S G
M. Pedram
How to Minimize Inductance Effects
(Cont’d)
Termination
Shunt-RC termination
Series_R termination
Diode termination
M. Pedram
Appendix II:
Values for Single Supply Noise vs. Distance
M. Pedram
On-chip Parasitic Capacitance
M. Pedram
Typical Transmission Line Parameters
and Coupling Coefficients
M. Pedram