Title: Design of PCB and MCM for high speed digital systems
1Design of PCB and MCM for high speed digital
systems
The art of compromise
- Torstein Gleditsch
- SINTEF Electronics and Cybernetics
2What this course is and what it is not
- This course is an introductory course to
designing Printed Circuit Boards and Multi Chip
Modules for high frequency digital applications. - It will introduce you to the basic concepts of
transmission lines and how utilize this theory
into practical designs. - It will introduce you to modeling of lines,
drivers and receivers to help you get a better
understanding of the effects of the different
measures. - It will not make you a proficient SPICE user.
- At last it will make you understand why high
speed digital design is difficult and how to
improve a design.
3What is so special with high-speed digital
- Broad band
- Must cover every frequency from DC to GHz
- No Dirty tricks possible
- Higher order harmonics is necessary for edge
integrity. - High number of critical signals
- Digital signals is more tolerant to distortion
- Switching create high currents in short periods
4A digital signal
- Frequency 50MHz
- Risetime 1ns ( 10 - 90 )
- Falltime 1ns ( 10 - 90 )
- Bandwidth 350MHz
5Spectrum of a single pulse
10 ns risetime
1 ns risetime
6Spectrum of a pulse train
7The effect of missing harmonics
1st harmonic only
1st to 5th harmonic
1st and 3rd harmonic only
1st to 21st harmonic
8Why transmission line calculations
- Transmission line theory describe the behavior of
the signals on a PCB or MCM - Simple models like RC-calculation do not apply at
higher frequencies - It is possible to predict the quality of the
signal. - It is possible to experiment with different types
of termination.
9When do I have to take into account transmission
lines effects
- Thumb rule When the rise time is less 2.5 times
the time delay of the signal on the trace.
(Transit time) - Another If the trace length is larger than 1/7
of the largest wavelength of the signal. (At
upper frequency edge) - Example with two different effective dielectric
constants
Trace length (mm)
Trace length (mm)
Risetime (ns)
Risetime (ns)
10Reflections from a 1ns risetime signal
40 mm non terminated line
300 mm non terminated line
11Crosstalk
300 mm non terminated line
40 mm non terminated line
12Transmission Lines
13Basic Transmission Line Types
Microstrip
W
t
H
Stripline
er tand
W
H
t
14Dielectric constant (Relative permittivity) (er )
- Describe a materials ability to hold charge
compared to vacuum when used in a capacitor. - The permittivity of a vacuum is e0 8.85419
10-10 - The permittivity of a material is e e0 er
- If we put a dielectric material between two
capacitor plates of area A and a distance D the
capacitance in Farad is
?
A
F
C
D
15Loss tangent ( tand )
- Describe the materials resistance to change of
polarization - s is the conductivity of the dielectric at
frequency ?. (S/m) - e is the permittivity of the material. (F/m)
?
tan
?
2
f
?
?
16Magnetic permeability (m)
- The magnetic permeability describe a magnetic
property of a material. It is the ratio between
the magnetic flux density (B) and the external
field strength (H).
B
H/m
?
H
The relative permeability of a material is the
permeability relative to vacuum.
µ
µ
µ
0
r
?
4
µ
0
7
10
17Properties of glass reinforced materials
18Basic transmission line diagram
Transmission line
Driver
Termination resistor
19Some Transmission Line Properties
- Signal velocity (m/s)
- Influenced by dielectric constant of materials
- Impedance
- Mostly influenced by dielectric constant of
material, width of line and distance to ground
plane(s) - Loss
- Influenced by conductivity of conducting
material, frequency of signal and loss tangent of
the material. - Dispersion
- Influenced by frequency dependant dielectric
constant
20The four describing parameters
- R Resistance per unit length ohm / m
- C Capacitance F / m
- L Inductance H / m
- G Conductance S / m
21Signal velocity ( ? )
- The signal velocity ? is measured in m/s
- For a practical calculations the signal velocity
is only dependant on the relative dielectric
constant er - With c0 the velocity of light in vacuum, we get
c
0
v
m/s
?
e
r
22Impedance ( Z0 )
- Impedance is measured in Ohms
- Impedance describes the AC resistance a driver
will see when driving a signal into a
indefinitely long transmission line.
?
R
L
j
?
ohm
Z
0
G
C
j
?
For a loss less line this simplifies to
?
L
ohm
Z
0
C
23Loss ( a )
- Loss a is measured in dB / m
- Loss is the reduction of signal voltage along a
line due to resistance and leakage - The resistive losses is due to resistance in
conductor and ground plane. - The dielectric losses is due to the energy needed
to change polarization of the dielectric
material. - The radiation losses is the energy sent from the
conductor acting as an antenna.
24Resistive losses ( aC )
- The resistance is based on the cross-section of
the conductor and the bulk resistivity of the
conductor material
s bulk conductivity S/m w line width m t
line thickness m
1
ohm
R
w
t
?
The loss factor is then calculated by
8,68589
R
dB/m
?
2
Z
0
25Skin depth ( ds )
At higher frequencies only a thin layer of the
conductor transport the current. The thickness
where the current density is reduced to 1/e is
called the skin depth ds.
?
1
m
d
s
f
?
?
µ
26Resistance at higher frequencies
When the conductor is much thicker than the skin
depth one have to substitute the skin depth for
the thickness in the resistance formula
1
R
ohm
?
d
w
s
1
R
?
1
?
w
?
?
µ
f
The resistance has now become frequency dependent
!
27Dielectric loss ( aD )
The dielectric loss aD is due to the energy
needed to change the polarization of the
dielectric material. The conductance is then
S/m
?
?
G
2
f
C
tan
The dielectric loss is then
1
?
8,68589
G
Z
dB/m
0
D
2
28Radiation loss
- All lines are radiating more or less
- Radiation loss is often negligible from a signal
integrity standpoint but important from a EMC
standpoint. - Radiation loss is difficult to calculate
29Dispersion
- Dispersion is an effect caused by different
velocities for the different frequency
components. The result is a different phase
change for each of the frequency components. - The reason for this is that the dielectric
constant of the material vary with frequency. - Dispersion is one of many sources of signal
distortion. - It is not easy to calculate dispersion because
the lack of frequency dependant material data. - Dispersion may cause trouble when exact edge
position is important.
30Crosstalk
Crosstalk is coupling of a signal form one line
to another There are two key parameters used to
describe crosstalk Kf the forward crosstalk
coefficient Kb the backward crosstalk
coefficient (normally the largest) Td is the
transit time delay
(
)
1
L
m
-
-
K
C
Z
m
0
f
2
Z
0
L
m
-
C
Z
0
m
Z
0
-
K
b
4
T
d
31Reflections
In this case Z1 R2 25 ohm, and Z0 50
ohm. Then r -0.333 gt -1.66 overshoot
32Via modeling
- Inductance and capacitance of vias is difficult
to calculate without 3D field analysis tools. - A 1nH inductance and a 1pF capacitance can be
used as a start - Careful use of coaxial line equations can also
give an indication of the values.
33Building a good board
- Some hints for the high performance designer
34What is important for a good board
- Symmetrical around center
- Tolerant for etch variations (/- 1 mil is not
unusual) - Lines spaced for acceptable crosstalk
- Standard dielectric thickness
- Tolerant for variation in dielectric thickness
- Acceptable high loss. (The loss may be your
friend) - Acceptable total thickness
- Power and ground layers close together (typically
100um)
35Laminate tolerances
Dictionary Toleranse Tolerance Tykkelse
Thickness Klasse Class Glassvevtype
Glass fabric type
36Prepreg Tolerances
37Impedance tolerance to line width
As line width increases, dependence on absolute
tolerance decreases
Border lines on -1mils absolute line width
tolerance
38Impedance tolerance to laminate tolerance
39Minimum line widths vs. er and dielectric
thickness
Curves on equivalent dielectric thickness
408 layer high-speed lay-up example 1
Dictionary Kobber Copper rent pure
418 layer high-speed lay-up example 2
42Pack32
- A desktop calculator for transmission lines
43Idea behind the program
- Easy to use
- PC based ( Windows 95 or NT )
- Good enough results (better than textbook
formulas) - Easy to organize material, component and lay-ups
- Easy for the developers to add new functions.
- Data is stored one place, no tedious typing
- Fast, no calculation take more than one second
44Workflow for transmission line analysis
- 1. Enter material data if not in database
- 2. Define a Lay-up
- 3. Calculate the line properties for a single
line - 4. Adjust Lay-up and re-calculate until satisfied
- 5. Calculate for double lines to check for
crosstalk and dual line impedance. - 6. Generate a SPICE model for a critical net in
the design, single or double lines as appropriate - 7. Simulate and adjust line parameters (width,
spacing, lay-up) - 8. Use the obtained parameters as design rules
for your critical nets.
45Material database
Sorry! You have not got the English version!
- For transmission line calculation,
- one need
- Electrical conductivity
- Dielectric constant
- Dielectric loss factor
- Magnetic permeability
46Lay-up definition
Sorry! Norwegian text!
47Single line analysis
Stripline
Microstrip
Buried microstrip
48Dual Line Analysis
49Exercise
- Design a lay-up with these properties
- 50 Ohms impedance
- FR-4 Dielectric
- No Solder resist
- 4 Signal layers
- Two Power layers
- Two Ground layers
- No more than 3dB loss for a 10cm line at 1GHz
- Tolerant for /- 1 mil etch error.
50Analog simulation of digital signals
51The two main simulation approaches
- Time domain simulators
- Easy to model nonlinear circuits such as digital
drivers - Has no possibility to handle frequency dependent
parameters - Convergence problems
- Relatively slow simulation
- Typical product SPICE
- Frequency domain simulators
- Easy to model circuits with frequency dependent
parameters - Only possible to model linear circuits
- Fast and easy convergence
- Typical product HP-EEsof Series IV Linear
simulator - What we really need does not exist
52Aim-SPICE
- Based on Berkeley SPICE version 3.E1
- Good user interface
- Good post processor
- Cheap
- Student version is limited but usable
- Designed for Windows95 / NT
- Text based, no schematic editor
53A SPICE example
- First line
- .MODEL LineModel LTRA L460N C94P R10 LEN .04
- V1 in 0 PULSE(0,10,10n,1n,1n,10n,20n)
- R1 in lin 50
- O1 lin 0 out 0 LineModel
- R2 out 0 50
54General
- All values are in basic SI units Ohm, Farad,
Henry,Volt, Ampere - Postfixes are allowed m 10-3 , u 10-6 , n
10-9 , p 10-12 k 103 , meg 106 , g 109 - Element names must be unique and is interpreted
by ASCII value - Node names can be numbers or names and is
interpreted by ASCII - Node 0 is the system ground and MUST be
connected. - The first line MUST be the model name
- Blank lines are allowed
- On the following pages some important elements is
shown, refer to the online SPICE manual for
details and more options.
55Passive elements
- Resistor
- Rxxx ltN1gt ltN2gt ltValuegt
- Capacitor
- Cxxx ltN1gt ltN2gt ltValuegt
- Inductor
- Lxxx ltN1gt ltN2gt ltValuegt
56Transmission lines
- Loss-less transmission line
- Txxx ltN1gt ltGndgt ltN2gt ltGndgt ltZ0Valuegt TdValue
- Lossy transmission line
- .MODEL LineModel LTRA L460N C94P R10 LEN .04
- Oxxx ltN1gt ltGndgt ltN2gt ltGndgt ltNamegt
57Voltage sources
- Voltage source
- Vxxx ltNgt ltN-gt ltTypegt
- Some useful Types for transient analysis
- For a pure DC source
- DC ltValuegt
- For a pulse train generator
- PULSE(ltHigh voltagegt,ltLow voltagegt,ltDelaygt,ltRise-t
imegt,ltFall-timegt,ltHigh timegt,ltCycle timegt)
58Exercise
- Simulate a 15cm long microstrip with these specs
- 125 um wide
- 17.5um thick
- Dielectric is polyimide
- Dielectric thickness is 100um
- Source impedance is 70 Ohm
- Termination resistance is 50 Ohms
- Rise and fall times is 0.8ns, frequency is 200MHz
- 1. What is the maximum over / undershoot
- 2. Insert a capacitor of 50pF in parallel with
the terminator, see what happens.
59Subcircuits
- SPICE have a nice feature called sub-circuits
which work like subroutines. - The syntax is
- .SUBCKT ltNamegt ltNode1gt ltNode2gt ....Node n
- ltSpice code with the node names as I/Ogt
- .ENDS
- You call this subcircuit with
- Xxxx ltN1gt ltN2gt ... Nn ltNamegt
60Distributed models
An other way of simulating transmission lines is
by using distributed models.
Selecting a high number of segments will give
excellent results but may cause problems in
initialization. The Student version can only
handle 10 of these segments due to the maximum of
50 elements.
61PACK model generation
- PACK can generate the models for you
62Pack Generated Distributed Model
63Exercise
- Make a distributed model of the last exercise and
compare results
64Exercise
- A board needs 60 Ohms lines in the inner layers.
- Design a board for 60 Ohm lines and generate a
single LTRA model of 5 cm. - Make a three segment line with a 60 Ohm
driver.After the first segment the line splits
into the two other segments. Each of the splits
is terminated in 60 ohms.
65Modeling of drivers and receivers
66Not all models have long hair and blue eyes
- Driver models are difficult to obtain, usually
you have three options - 1. Use vendor supplied SPICE models of the
output stage - 2. Use vendor supplied IBIS models
- 3. Roll your own simple models based on rise
time and output impedance.
67Vendor supplied SPICE models
- Difficult to obtain, vendors generally do not
want to disclose their design. - Most vendors require non disclosure agreement
- Difficult to run. Vendors usually use H-SPICE,
board designers do not. - Models are to detailed, unnecessary accuracy give
slow simulation. - Difficult to set the right environment for
drivers etc.
So, vendor supplied SPICE models is generally not
a good idea.
68Vendor supplied IBIS models
- IBIS solves most of the problems related to
vendor models except - Not all SPICE simulators run IBIS models
- Not all vendors supply IBIS models, although this
is rapidly increasing - IBIS models is sufficiently accurate for package
and module simulation. - So, use IBIS models (If you can)
69IBIS model
70IBIS I/V curves (ABT 244)
Pull-up curve ( Referenced to Vcc ) Y-axis
Current A X-axis Voltage V
Pull-down curve
71IBIS viewing tool
72Make your own models
- Simple models are sufficient to study
transmission line effects and to develop design
rules. - They are not sufficient for optimizing a specific
net. - A very simple model
73Exercise
Study the supplier SPICE model of an IBIS
driver. Adjust the data to fit an
SN74ALVCH16244 Compare with data sheets and the
timing article
74Routing of high speed signals
- There is no easy way
- ( but some auto-routers are really good )
75Main Issues for Routing
- Routing topologies and loading
- Impedance
- Rise and fall time degradation
- Signal Skew (Signal delay difference)
- First incident clocking
- Crosstalk
- Signal return path
- Termination
76Point to Point Topology
- This is the ideal topology but not very efficient
in that it requires a lot of extra buffers. - Always use this topology for critical clock
trees. - Use low skew clock buffers
- Skew can be compensated with delay lines
77Fan / Star topology
- This topology is routing efficient but put heavy
load on the driver, especially where several
lines fan out. - Use drivers with low output impedance.
- Terminate properly at each receiver or
reflections will propagate back and forth in the
net. - Be careful with the high power consumption of
many terminated lines
78T - Topology
- This split will cause a 33 negative reflection
if all lines are of same impedance. - All ends (also driver) should be terminated
properly for larger nets - The driver will have to drive twice the amount of
DC into the termination resistors.
79Daisy Chain
- The preferred topology for high speed digital.
- Terminate in both ends.
- Allow no stubs
- Keep tap load low.
- Keep distance between loads so high that the
signal can recover.
80Rise and fall time degradation
- Any lossy transmission line acts as a filter
which filters the high frequency components
first. Since the high frequency components is
damped more the signal will show slower rise
times at the end of the line.
81Signal skew
- Signal skew is the delay difference at inputs.
- Keeping low signal skew in a clock system is a
challenge and there are several effects to
consider. - Keep every signal trace equally long
- Use low skew drivers
- Rise time degradation may pose a problem in
calculated delay daisy chains. - Signals travel faster at outer layers, so trace
lengths must be compensated. Or - use only inner
layers. - Poor decoupling influence rise time.
- Use only first incident clocking or better,
reduce reflections to zero.
82Signal return path
- The high frequency signals follow a mirror image
of the trace on the ground plane. Do not degrade
this return by - Changing to layers which have another
ground-plane without placing a ground-ground via
close to the signal via. If the new reference
plane is a power plane a low inductance
decoupling is placed close to the signal via - Using split reference planes.
- Using one small via for several returns, this may
cause common mode problems.
83How To Design Delay Lines
- Remember loss
- Use meander structure, be careful to avoid
inductor effects. - Keep meander lines at least 4 times the distance
to ground plane apart to avoid distortion. - Preferably use inner layers, this reduce
distortion and keep inductance low.
84Line Termination
- Terminating lines driven with drivers that do not
allow termination
85Why terminate the signals
- Termination is a method to match the driver, line
and receiver in such way that no reflections are
generated. - Generally the ideal would be to do a parallel
termination in the line impedance - Terminate critical lines that are longer than 1/3
of the rise time. On FR-4 this mean that a 1ns
rise signal, the longest unterminated line is
5cm. - Reflections can cause double trigging, if using
non- terminated nets do not use data before the
reflections have settled.
86Series Termination
- Used to match driver impedance
- Slow rise and fall times (Some times good, for
instance to achieve low crosstalk) - Absorb reflections if matched to line
87Parallel Termination
- The termination of choice for high speed digital,
especially for ECL, PECL and other technologies
intended for termination. - High power consumption.
- 100 clean signals possible
88RC Termination
- AC termination, better for technologies not
intended for termination. - Low power consumption (No DC consumption)
- Be careful with inductive capacitors, select
capacitors with care.
89Thevenin Termination
- Ideal for bus termination or lines with 3-State
drivers. - 330 Ohm 220 Ohm is often used.
- Faster switching from 3-state
- Does NOT correctly terminate the line,
reflections will occur
90Diode Termination
- Kills large over and undershoot
- Not generally useful in high-speed systems. (A
design needing this type of termination have
probably greater problems elsewhere)
91Power distribution and de-coupling
- How to supply the right voltage at the right
place at the right time
92Power distribution
- The power and ground system serves two purposes.
- First to serve as a return path for the signal.
- Second to supply the devices with power.
- The high frequency components of today is using a
lot of power at a low voltage. The result is very
high currents. - A modern FPGA can draw several amperes the first
nanosecond after switching. - To meet these challenges one need a low
inductance, low resistance and high capacitance
power system.
93Why Decoupling
- The high switching currents in today's components
create a need for a local current supply with
sufficient charge to avoid a severe voltage drop
at the power pins. - When these resources are exhausted on need larger
supplies that are closer than the power supply. - Closely spaced power planes give a good high
frequency, low inductance decoupling but it can
not hold much charge.
94A Capacitor is Not a Capacitor
1
f
res
?
?
2
L
C
- A capacitor is a passive device dominated by
capacitance. - A 100nF capacitor can have C84nF, L1.3nH,
R0.13ohm - A capacitor is often described with its resonance
frequency - Different dielectrics NPO, X7R or Z5U have
different properties. - Usually thin and wide X7R chips is the best
choice for decoupling. - To be sure on have to measure the inductance
which is difficult - To compare two capacitors of same nominal
capacitance chose the one with the highest
resonance frequency. - So called high frequency capacitors is not
necessarily less inductive
95Selecting Decoupling Capacitors
- Selecting a decoupling capacitor is not easy but
some general - rules apply
- Select the smallest value that is sufficient to
decouple the device. If you do not know what is
sufficient choose one 100nF capacitor for each
power pin - Use only chip capacitors, leaded capacitors is to
inductive to be of any help. - Decouple in levels with the smallest value (i.e.
100nF) close to the power pin, the medium
values(i.e. 10mF) evenly distributed, the higher
values (i.e. 47mF tantalum) close to the
connector power pins.
96Where to place decoupling capacitors
- Get as close as possible to power and ground
pins. - Keep distance to pin very SHORT and the trace
WIDE. - Set the via as close as possible to the
capacitor, preferably in the pad. - One capacitor for every power (and ground) pin is
a good rule of thumb, you do not have to mount
all of them in production if the system works
with less. - Calculate inductance. If too high - forget the
capacitor. - Do not forget to place decoupling capacitors
close to termination resistors, it is important
to keep the reference voltages steady.
97Splitting power planes
- Be very careful with splitting ground planes.
- If you have to split, keep the split narrow and
secure an AC return path. - Splits act as slot antennas and will radiate
heavily. - Remember that digital signals are broad band and
that filters are narrow band. Filters over the
gap may cause problems.
98Microstrip Over a Slotted Ground Plane
1
2
3
4
IEEE Circuits Devices Nov. 1997