Thanh T. Tran (Auth.) - High-Speed DSP and Analog System Design-Springer US (2010) PDF
Thanh T. Tran (Auth.) - High-Speed DSP and Analog System Design-Springer US (2010) PDF
Thanh T. Tran
This book covers the high-speed DSP and analog system design techniques
and highlights common pitfalls causing noise and electromagnetic interfer-
ence problems engineers have been facing for many years. The material in
this book originated from my high-speed DSP system design guide (Texas
Instruments SPRU 889), my system design courses at Rice University and
my experience in designing computers and DSP systems for more than 25
years. The book provides hands-on, practical advice for working engi-
neers, including:
• Tips on cost-efficient design and system simulation that minimize
late-stage redesign costs and product shipment delays.
• 11 easily-accessible chapters in 210 pages.
• Emphasis on good high-speed and analog design practices that mi-
nimize both component and system noise and ensure system de-
sign success.
• Guidelines to be used throughout the design process to reduce
noise and radiation and to avoid common pitfalls while improve
quality and reliability.
• Hand-on design examples focusing on audio, video, analog filters,
DDR memory, and power supplies.
The inclusion of analog systems and related issues cannot be found in oth-
er high-speed design books.
vii
Preface viii
Thanh T. Tran
Houston, Texas, 2010
ix
Contents
xi
xii Contents
Glossary .........................................................................................211
Index ..............................................................................................213
About The Author
xv
Challenges of DSP
Systems Design 1
With these difficulties, it’s necessary to rethink the traditional high speed
DSP design process. In the traditional approach, engineers focus on the
functional and performance aspects of the design. Noise and radiation are
considered only towards the later stages of the design process, if and when
prototype testing reveals problems. But today, noise problems are becom-
ing increasingly common and more than 70% of new designs fail first-time
EMC testing. As a result, it is essential to begin addressing these issues
from the very beginning of the design process. By investing a small
amount of time in the use of low-noise and low-radiation design methods
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_1, 1
© Springer Science+Business Media, LLC 2010
2 Chapter 1
at the beginning of the development cycle, this will generate a much more
cost efficient design by minimizing late-stage redesign costs and delays in
the product ship date.
DDR
RGB888/
YUV422
HDMI
HD Camera VIDEO
Receive IN RGB888 HD
VIDEO
OUT
Panel
1000
GMII
BaseT SPI
Ethernet Touch-
PHY Screen
1000 HDMI
BaseT
GMII
DSP/SoC HD
Ethernet
PHY Display
EMIF
NOR or NAND USB Port1
FLASH USB Port2
Mic Array McASPs
AUDIO COMM SDIO
Audio
IN/OUT IN/OUT WLAN
CODECs
UART
GPS
PCI
WLAN/ PCI SPI
Serial
UWB
ROM
CORE
POWER
Power SUPPLY
PLL OSC
IO/DDR
The coupling between a noise source and noise victim causes electrical
noise. Figure 1.2 shows a typical noise path. The noise source is typically a
fast-switching signal and the noise victim is the component carrying the
signal. The noise victim’s performance will be impacted by the noise.
Coupling takes place through the parasitic capacitances and mutual induct-
ances of the adjacent signals and circuits. Electromagnetic coupling occurs
when the signal traces become effective antennas, which radiate and gen-
erate interferences to the adjacent circuitries.
• Flicker Noise: Also known as 1/f noise is present in all active de-
vices. It is caused by traps where charge barriers are captured and
released randomly, causing random current fluctuations. Flicker
noise is a factor of semiconductor process technologies, so DSP
system design cannot reduce it at the source but focus should be
on mitigating its effects.
All of the blocks shown in Figure 1.3 from the ADC to the Amp stage are
very sensitive to noise so any interference coupled to any of the blocks will
propagate and generate unwanted audible sounds. Common audio design
problems include:
• Audio playback may cause a flicker on the video screen. This can
be corrected by increasing the isolation of the video and audio cir-
cuits. The best method is by using high power supply rejection ra-
tio (PSRR) linear regulators to isolate the audio CODEC and the
video encoder/decoder supplies. Also, manually route the critical
traces away from any of the switching signals to reduce the
crosstalk and interferences.
In any HD video systems, there are many wide high speed busses switch-
ing at a rate of 66MHz or higher and these buses generating broadband
noise and harmonics that cause radiations in the Gigahertz range. This
type of interference is very difficult to control because there are so many
of these busses on the board and it is not practical to terminate every signal
trace being routed from one point to another. The good news is that there
are good design practices to follow in order to minimize the interference.
REFERENCES
[1] Federal Communication Commission (2005) Unintentional Radiators
Title 47 (47CFR), Part 15 B.
http://www.fcc.gov/oet/info/rules/part15/part15-91905.pdf
[2] Hiers T and Ma R (1999) TMS320C6000 McBSP: I2S Interface.
Texas Instruments Inc’s Application Report, SPRA595.
[3] Franco S (2002) Design with Operational Amplifiers and Analog Inte-
grated Circuits. McGraw-Hill, New York.
[4] Nassar M, Gulati K, DeYoung M, Evans B, and Tinsley K (2008)
Mitigating Near-field Interference in Laptop Embedded Wireless
Transceivers. IEEE Journal of Signal Processing Systems.
Transmission Line (TL) Effects 2
Transmission line (TL) effects are one of the most common causes of
noise problems in high-speed DSP systems. When do traces become TLs
and how do TLs affect the system performance? A rule-of-thumb is that
traces become TLs when the signals on those traces have a rise-time (Tr)
less than twice the propagation delay (Tp). For example, if a delay from
the source to the load is 2nS, then any of the signals with a rise-time less
than 4nS becomes a TL. In this case, termination is required to guarantee
minimum overshoots and undershoots caused by reflections. Excessive TL
reflections can cause electromagnetic interference and random logic or
DSP false-triggering. As a result of these effects, the design may fail to get
the FCC certification or to fully function under all operating conditions
such as at high temperatures or over-voltage conditions.
There are two types of transmission lines, lossless and lossy. The ideal
lossless transmission line has zero resistance while a lossy TL has some
small series resistance that distorts and attenuates the propagating signals.
In practice, all TLs are lossy. Modeling of lossy TLs is a difficult chal-
lenge that is beyond the scope of this book. Since the focus of this book is
only on practical problem-solving methods, it assumes a lossless TL to
keep things simple. This is a reasonable assumption because in DSP sys-
tems where the operating frequency is less than 1GHz the losses on printed
circuit board traces are negligible compared to losses in the entire signal
chain, from analog to digital and back to analog.
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_2, 13
© Springer Science+Business Media, LLC 2010
14 Chapter 2
C1 C2 C3 C4
1
Vp = , (2.1)
LC
L
Zo = , (2.2)
C
where L is inductance per unit length and C is capacitance.
Another important property of the TL is the propagation delay, Td. The Eq.
(2.3) for Td is
1
Td = = LC . (2.3)
Vp
The source and load TL reflections depend on how well the output imped-
ance and the load impedance, respectively, are matched with the character-
istic impedance. The load and source reflection coefficients, Eq. (2.4) and
Eq. (2.5), are
Transmission Line (TL) Effects 15
Z S − ZO
ΓS = source _ reflections = , (2.4)
Z S + ZO
Z L − ZO
ΓL = load _ reflections = , (2.5)
Z L + ZO
where ZS and ZL are the source impedance and load impedance respec-
tively.
The following example shows the characteristics of a TL with no load and
with a 3V signal source driving the line.
Example 2.1: Calculate the voltage at the open ended load of the
transmission line below.
Zo 50
V initial = V clk =3 = 2V
Zs + Zo 25 + 50
Zs − Zo 25 − 50
ρs = = = −0.333
Zs + Zo 25 + 50
Z L − Zo
ρL = =1
Z L + Zo
16 Chapter 2
In Figure 2.3, the overshoot voltage can be calculated using a lattice dia-
gram [1] as follows.
As shown in Example 2.1, the reflections with a 3V source caused the sig-
nal to overshoot as high as 4V at the load as explained below:
• The voltage level at time T2, when the reflected signal arrives
at the source, depends on the source impedance and the char-
acteristic impedance of the TL.
• The voltage level at time T3, when the reflected signal arrives
at the load again, depends on the reflected voltage at T2 plus
the reflected voltage at time T3.
Figure 2.5 shows the waveforms at the load for both terminated and
unterminated circuits. As shown in the previous example, the terminated
TL has a zero reflection coefficient and therefore no ringing occurs on the
waveform as seen on the top graph of Figure 2.5. The problem is that in
high-speed digital design, adding a 50-ohm resistor to ground at the load is
not practical because this requires the buffer to drive too much current per
line. In this case, the current would be 3.3V / 50 = 66mA. A technique
known as parallel termination can be used to overcome this problem. It
consists of adding a small capacitor in series with the resistor at the load to
block DC. The RC combination should be much less than the rise and fall
times of the signal propagating on the trace.
tr
L2 , max = L 1
10 (2.6)
OSCILLOSCOPE
Design file: 5912CLK.TLN Designer: TI
BoardSim/LineSim, HyperLynx
Comment: NOTE: The signals recorded at the terminating loads is identical therefore only the magenta signal is shown.
3500.0
3000.0
V
o
l 2500.0
t
a
g 2000.0
e
-
m 1500.0
V
-
1000.0
500.0
0.000
-500.0
0.000 5.000 10.000 15.000 20.000 25.000 30.000 35.000 40.000 45.000 50.000
Time (ns)
Figure 2.9 shows an example of one clock output driving two loads con-
nected using a daisy-chain topology. The distance from the source to the
first load (1st SDRAM) is the same as the distance from the first load to the
second load (2nd SDRAM). In this case, the reflections coming from the
second load distort the clock signal at the first load. The best way to mini-
mize this distortion is by adding a parallel termination at the second load to
reduce the impedance mismatch and therefore reduce the reflections as
shown in Figures 2.10 and 2.11. This system still requires a series termi-
nation at the source to control the edge rate of the whole signal trace. This
resistor needs to be small so that the source and sink currents are large
enough to drive two loads. In this example, the series termination resistor
is 10 ohms.
A d d e d R C
2 nd S D R A M P a r a ll e l
1 s t S D R A M T e r m i n a ti o n
D M 6 4 2 E V M
10 O h m s
S e rie s
T erm .
C L O C K
O U T F R O M D S P
In Figure 2.12, the voltage at the load is slowly charged up to the maxi-
mum amplitude of the clock signal. Initially, the load looks like a short cir-
cuit. Once the capacitor is fully charged, the load becomes an open circuit.
The source resistor Zs controls the rise and fall-times. Higher source resis-
tance yields slower rise-time. The load voltage at any instant of time, t,
greater than the propagation delay time, can be calculated using the fol-
lowing equation:
Figure 2.15 shows the simulation result at U2 and Figure 2.16 shows the
actual scope measurement in the lab.
24 Chapter 2
OSCILLOSCOPE
Des ign file: 507201C.HY P Des igner: TI
B oardS im/LineS im , Hy perLy nx
5.000
4.000
V
o 3.000
l
t
a 2.000
g
e
- 1.000
V
-
0.000
-1.000
-2.000
-3.000
0.000 40.00 80.00 120.00 160.00 200.00
Tim e (ns)
Figure 2.18 shows an audio clock that transmits by U17 and receives by
U3. The design has a 20-ohm series termination resistor but no parallel
termination at the load. This demonstrates the concept discussed earlier.
OSCILLOSCOPE
Design file: 507201C.HYP Designer: TI
BoardSim/LineSim, HyperLynx
5.000
4.000
V
o 3.000
l
t
a 2.000
g
e
- 1.000
V
-
0.000
-1.000
-2.000
-3.000
0.000 40.00 80.00 120.00 160.00 200.00
Tim e (ns)
The lab measurement shown in Figure 2.20 correlates with the simulation
very well. The 22-ohm series resistor can be modified to lower the over-
shoots and undershoots. But since the overshoots are less than 0.5V, they
are acceptable in this case.
Ground
grid Current
Return
Input
Signal
Signal trace is routed between the two ground paths of the grid
Figure 2.22 also shows another example of using a ground grid where the
signal is being routed diagonally. As shown in this figure, the current re-
turn has to travel on a zig zag pattern back to the source and creates a large
current return loop that yields higher inductance and lower capacitance per
unit length. In this case, the characteristic impedance is higher than using a
continuous ground plane and higher than the case where the signal is
routed in parallel with the ground grid as shown in Figure 2.21.
Current
Ground
Return
gr id
Input
Signal
Signal trace is route d diagonally
So, if ground grid is a required in a design, the best approach is to route the
high speed signals right on top of the grids and parallel to the grid to en-
sure the smallest current return loops. This lowers the characteristic im-
pedance to the level equivalent to the impedance of the continuous ground
plane. This is very difficult to accomplish since complex board has many
high speed traces. Therefore, continuous ground plane is still the best
method to keep characteristic impedance and EMI low.
• Slow down the signal edge rate by lowering the buffer drive
strength if it is not affecting the timing margins. Remember a
Transmission Line (TL) Effects 29
REFERENCES
[1] Hall S, et al (2000) High Speed Digital System Design. John Wiley &
Sons, New York.
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_3, 31
© Springer Science+Business Media, LLC 2010
32 Chapter 3
The current return density and the amount of crosstalk can be estimated as
shown in Figures 3.3 and 3.4. Based on the equations shown in the figures,
the spacing between the traces and the distance that they run in parallel de-
termines the amount of crosstalk. Obviously, moving the traces further
from each other will reduce the crosstalk.
current
density curve
signal trace
H
Ground
I 1
ID = , where I is the signal current. (3.1)
πH ⎛D⎞
2
1+ ⎜ ⎟
⎝H⎠
H
Ground
K
Crosstalk = 2
, (3.2)
⎡D⎤
1+ ⎢ ⎥
⎣H ⎦
34 Chapter 3
where D is the distance between the traces, H is the height of the signal to
the reference plane and K is the coupling constant less than 1.
There are two types of crosstalk, forward and backward. Forward
crosstalk, also known as capacitive coupled crosstalk. This occurs when
the current flows in the same direction as the source. With backward
crosstalk, which is also called inductive coupled crosstalk, the coupling
current flows in the opposite direction of the source.
The following simulations [2] demonstrate the concept of reducing for-
ward and backward crosstalk by spacing the aggressor and victim traces.
The model simulates two parallel 5 mil wide, 12-inch long traces. The
source of the trace is connected to a DSP and the load to DDR memory. As
shown in Figure 3.5, D0 line is an aggressor and D1 line is a victim.
2000.0
1500.0
V
o 1000.0
l
t
a
g 500.0
e
-
m
V 0.000
-
-500.0
-1000.0
-1500.0
-2000.0
0.000 4.000 8.000 12.000 16.000 20.000
Tim e (ns)
Figure 3.6 shows simulation results. On the victim trace, the first negative-
going pulse, which has a –200mV peak, is the forward crosstalk. The posi-
tive-going pulse of 240mV is the backward crosstalk. The backward pulse
width is about 2 times the coupling region. In this case, the coupling region
is 3.54nS and the simulation also shows a 4nS backward crosstalk pulse.
The crosstalk Eq. (3.2) is
K
2
.
⎡D⎤
1+ ⎢ ⎥
⎣H ⎦
Let’s assume that K=1, D=5 mils and H=10 mils as in the simulation. The
maximum crosstalk is then calculated as follows:
K 1
Max. Crosstalk = 2
= 2
= 0.8V.
⎡D⎤ ⎡5⎤
1+ ⎢ ⎥ 1+ ⎢ ⎥
⎣H ⎦ ⎣10 ⎦
2500.0
2000.0
1500.0
1000.0
V
o
l
t
a
g 500.0
e
-
m
V
-
0.000
-500.0
-1000.0
-1500.0
-2000.0
0.000 2.000 4.000 6.000 8.000 10.000 12.000 14.000 16.000 18.000 20.000
Ti m e (n s)
• For board layout, analyze the critical signals and minimize the
coupling regions.
C
λ= , (3.3)
f
where C is the speed of light or 3x108 m/s and f is the frequency in Hertz.
The equation shows that a 100MHz clock signal has a wavelength of 3 me-
ters or 9.84 feet. A good rule for minimizing radiation is making sure that
the trace length is not longer than the wavelength divided by 20. So in the
case of the 100MHz clock signal, the signal length should be kept below
0.15m or 0.492 feet. Keeping the traces below 0.5 feet is easy, but the
squarewave clock signal consists of multiple harmonics and each of the
harmonics can radiate even when the traces are very short. Here is an ex-
ample.
Example 3.1:
Let f = 500MHz, the 5th harmonic of the 100MHz clock,
Effects of Crosstalk 37
C 3x108
λ= = = 0.6m .
f 500 x106
A rule-of-thumb for determining when the clock trace becomes an effec-
tive antenna is taking the wavelength divided by 20. The maximum length
the 500MHz clock is
C 3x108
λ= = = 0.03m or 3cm which is 1.18 inches.
20 f 20(500 x106 )
What this means is that, depending on the rise and fall-times of the
100MHz signal, the energy of the 5th harmonic can radiate and interfere
with the adjacent circuits when this signal trace is longer than 1.18 inches.
The energy of the harmonics depends on the rise and fall-times of the sig-
nal as shown in Figure 3.8. In this figure, it is assumed that the clock
waveform has a 50% duty cycle and rise and fall-times are equal. With
these assumptions, only odd harmonics of the clock are present. The am-
plitude of the harmonics starts decaying at the first pole frequency, f1, at a
rate of 20dB/decade and then increases to 40dB/decade at the second pole
frequency f2. The equations for f1 and f2 are
1
f1 = , where Pw is the high-time of the waveform, (3.4)
Pw
1
and f2 = , where Tr is the rise-time of the waveform. (3.5)
πTr
T r P w
T
1
2 0 d B /d e c a d e
π Pw 1
Amplitude
π T r
4 0 d B /d e c a d e
H ig h e r T r r e d u c e s
a m p li t u d e o f h a r m o n i c s
1 /T (3 ) (5 )
F r e q u e n c y ( h a r m o n ic )
To illustrate how a digital clock waveform can generate crosstalk that de-
grades the video quality of a time shifting system [3], let’s take a look at a
system diagram shown in Figure 3.9 where many critical components are
placed on the same printed circuit board. In this design, the clock signals
ranging from 18.4MHz to 100MHz are being routed to all the subsystems
(modem, audio CODEC, video encoder/decoder, CPU, MPEG encode and
MPEG-2 decode).
• The Video Decoder receives the baseband video signal from the
Tuner and digitizes the signal to prepare for digital signal com-
pression. This video data rate for an analog TV channel with
640x480 30 frames per second resolution is around 147Mbits/s
(data rate = 640 x 480 x 16 bits/pixel x 30 frames/second =
147Mbits/s).
Effects of Crosstalk 39
• The CPU then reads the compressed data from the hard drive and
sends the data to the MPEG-2 Decoder [4]. The Decoder decom-
presses the data from 2Mbits/s bit rate back to 147Mbits/s bit rate
and sends it to the Video Encoder. The Encoder converts the digi-
tal decoded data to analog signals and displays it on a TV screen.
In this time shifting system design, the crosstalk can occur anywhere
within a system even though the system was fully FCC certified, so it is
very difficult to find the root causes of the problem. For example, any of
clocks can generate harmonics that radiate to the antenna input and inter-
fere with the TV channel. To illustrate this effect, Figure 3.11 shows a
video screen with horizontal lines generated by the third harmonic
(55.2MHz) of the 18.4MHz clock radiating to the antenna input. 55.2MHz
harmonic happens to be within Channel 2 (54MHz to 60MHz) of the
NTSC spectrum [5] and causes interferences that can’t be rejected by the
Tuner because the Tuner can’t distinguish between the in-band noise and
the actual TV signal.
In this case, the best way to get rid of the interference is to reduce the en-
ergy radiated by the 3rd harmonic of the 18.4MHz clock. Figure 3.8 shows
that increasing the rise-time of the signal, assuming that this is not causing
any setup and hold time violations, attenuates the harmonic amplitude and
reduces the radiation. The two ways to reduce the rise-time are lowering
the clock buffer slew rate if possible or adding a series termination resistor
at the output of the clock buffer as shown in Figure 3.12.
3.3 SUMMARY
As highlighted in sections 3.1 and 3.2, the current return paths and the sig-
nal rise-time play the key role in generating crosstalk that interferes with
the adjacent circuits and causes random system failures and or system per-
formance degradation.
• Slow down the rise and fall-times if possible. Increasing the rise-
time reduces the power spectral density of the harmonics as shown
in Figure 3.8.
• Keep high speed signals as short as possible and make sure that
the wavelengths of the third and fifth harmonics of the signal are
much less than the wavelength divided by 20.
• Always route the high speed signals away from any critical high
impedance traces. High impedance traces are the input traces to
the video and audio amplifiers. Also, space the traces at least one
width apart from each other to reduce the coupling; for example a
5 mil trace should have a gap of at least 5 mils to another trace.
REFERENCES
[6] Texas Instruments Inc (2009) Spread Spectrum Clocking Using the
CDCS502/503 Application Report, SCAA103.
http://focus.ti.com/lit/an/scaa103/scaa103.pdf.
Power Supply Design
Considerations 4
Power supply design is perhaps the most challenging aspect of the entire
process of controlling noise and radiation in high-speed DSP design. This
is largely because of the complexity of the dynamic load switching condi-
tions. These include the DSP going into or out of low power modes, exces-
sive in-rush current due to bus contention and charging decoupling capaci-
tors, large voltage droop due to inadequate decoupling and layout,
oscillations that overload the linear regulator output, and high current
switching noise generated by switching voltage regulators. A clean and
stable power supply design is required for all DSP systems to guarantee
system stability. This chapter outlines the importance of proper power
supply design and the methods to minimize unwanted noise.
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_4, 45
© Springer Science+Business Media, LLC 2010
46 Chapter 4
• If the output voltage drops due to higher current load, the error
amplifier configured as a negative feedback circuit compares the
Regulated Output divided by the resistors R1 and R2 to the Refer-
ence Voltage and drives higher base current Ib to maintain regula-
tion. As shown in Eq. (4.3), the output current Ie is increased with
the increase in the base current Ib.
Ic Ie
DC Input T1 Regulated Output
R1
Ib Error Amp
C2
+
R2
Reference Voltage
R3
Like any other feedback systems, if there are changes in the input voltage
and the load current, the system takes some time to stabilize and this time
typically is specified in the component data sheet under the transient re-
sponse section. The major issue with linear regulator is the power dissipa-
tion across the transistor T1 for high output current applications. The
power dissipation is
PT1 = (Vin – Vout) x Ie (4.4)
For example, if the input voltage is 12V and the output is regulated at 5V
as shown in the Design Example 4.1. For the 1A output current, the power
dissipation across the transistor T1 is
PT1 = (12 – 5) x 1 = 7 Watts.
This power dissipation generates a lot of heat and increases the device op-
erating temperature to the point where heatsink is required to keep the de-
vice temperature under the maximum allowable limits. As the current re-
quirement increases with higher performance DSP and the system becomes
smaller and smaller, it is no longer practical to use the linear regulators to
generate all the supply voltages. In this case, it is best to use switching
power supply for the main power and linear regulators to provide clean
low noise supplies to the noise sensitive circuits such as analog and mixed
analog/digital data converter circuits as shown in Figure 4.2.
• As shown in the LM317 data sheet [1], this device has a Ripple
Rejection specification of 62dB minimum when using a 10uF ca-
pacitor to decouple the Adjust pin.
D1 1N4002
U1 LM317
+12V IN +5V ANALOG
IN OUT
ADJ
D2 1N4002
C1 100n R1 240 C3 1u
C2 10u R2
⎛ R ⎞
Vout = Vref ⎜⎜1 + 2 ⎟⎟ + (I adj R2 ) , (4.5)
⎝ R1 ⎠
⎛ R ⎞
(
5 = 1.25⎜1 + 2 ⎟ + 50 x10− 6 xR2 )
⎝ 240 ⎠
⎛ R ⎞
5 = 1.25⎜1 + 2 ⎟ ,
⎝ 240 ⎠
R2 = 720 ohms. Put the complete circuit shown in Figure 4.3 in the
circuit simulator [2] and the results are shown in Figure 4.4 where
the output is regulated at +5V when the input is +12V.
Error Amp
Pulse Width Modulator (PWM)
- 4 +
+
5 6
R5
R6
Figure 4.6 demonstrates how the system compensates for higher output
voltage. Higher output voltage leads to lower duty cycle signal and there-
fore lower the output voltage back to the regulated level.
52 Chapter 4
Figure 4.7 shows that the duty cycle of the signal increases when there is a
decrease in the output voltage. This again forces the system back into
regulation.
INPUT:
VIN (input voltage) 10.8 12.0 13.2 V
OUTPUT:
Refer to [4] to calculate all the component values shown in Figure 4.8.
The performance of this buck converter depends on the component values,
the component placements and the layout. Here are some important points
to remember;
VCC_12V
C1 C2a C2b
+ R3 R20
100K 1% 100K, 1% 22uF 25V 22uF 25V
9
5
6
7
8
25uF C7
U54A
S1 D 1
S2 D 2
S3 D 3
D4
PAD
100nF 1 16 4 Q1
EN HDRV G
2 15 Si7860DP
FB SW C6 VCC_3.3V
3 14
1
2
3
COMP BOOT L1
4 13 100nF
VDD LDRV 800nH
5 12
UVLO BP
6 11 C9a C9b C9c
RT SS_SEL
9
5
6
7
8
R9 + + +
100K 1% 7 10
ILIM PGOOD
S1 D 1
S2 D 2
S3 D 3
D4
PAD
4
P w rP d
1
2
3
C4 R2
R11
8.2nF 2.32K 1% 10K, 1%
R10 C14
C5
49.9 1% 1nF
220pF
R5
2.2K 1%
Power Supply
14
+ 1
3
2
-
Load C
7
Current return path
Power Supply
14
+ 1
3
2
-
Load Ca
7
Low noise with high power sup- Switching noise may cause EMI
ply rejection ratio problems or video and audio arti-
facts
Excellent choice for video, au- Excellent choice for the core CPU
dio, analog and PLL circuits. and the IO power.
Table 4.2 helps determine which power supply solution is a better fit for
the application. The next step is to determine the current consumptions and
whether or not power sequencing is required. In general, DSPs have a
minimum of two power supply rails, Core and IO. The sequence of ramp-
ing the Core and IO voltages can affect the startup current consumption so
refer to the device data manuals to help design a robust power supply for a
particular DSP. Here are recommended rules for selecting/designing a DSP
power supply.
CORE Voltage Regulator Design:
• The final design step for the Core voltage regulator is whether
or not a heatsink is required.
IO Voltage Regulator Design:
IO voltage regulator design depends on the external loads in the specific
application. For fast switching signals, the IO currents are supplied by the
decoupling capacitors, not by the regulator itself due the parasitic induc-
tance associated with the power supply trace or plane. The dynamic cur-
rent calculation will be shown in the decoupling section. The following
guidelines provide a conservative method to design an IO voltage regulator
for the DSP itself. It should be noted that this method applies to the DSP
power alone as opposed to the entire system.
• Count the number of outputs from the DSP. All GPIOs should
be considered as outputs.
60 Chapter 4
The number of power supply rails required for DSP or SoC is increasing
constantly as more and more peripherals are being integrated. And manag-
ing these rails during powering-up or down of the DSP is a very difficult
task. Typically, a high performance DSP or SoC consists of at least 3
power supply rails, +1.8V for DDR2, +3.3V for data converters, +1.2V for
core. The internal logic has many voltage translations to enable all the
blocks communicating to each other. During power-up, if one power sup-
ply rail goes up before another for some period of time, the internal logic
can get to an unknown state which can cause internal bus contention and
the system to go unstable. Designers must refer to the DSP data sheet and
design in the power sequence if it is required. The problem is that the
62 Chapter 4
power supply sequence circuits shown in Figures 4.11 and 4.12 can only
guarantee the power up sequence of the power supply itself, not the whole
system. This is because the decoupling capacitors used around the DSP as
shown in Figure 4.10 affecting the time it takes for the power supply to
ramp up to the final operating voltage. This time can be calculated as fol-
lows:
⎛ dV ⎞
I power = Cdecoupling ⎜ ⎟,
⎝ dt ⎠
⎛ dV ⎞
dt = Cdecoupling ⎜ ⎟, (4.6)
⎜I ⎟
⎝ power ⎠
where Cdecoupling is the total decoupling capacitance, Ipower is the current
sourced from the power supply, dV is the change in voltage and dt is the
time it takes to get to the dV value.
As shown in the Eq. (4.6), for a given power supply, the time it takes to
reach the final voltage level depends on the total decoupling capacitance.
So, to guarantee a particular sequence, it is very important for designers to
do the following:
• Refer to the DSP data sheet and determine whether or not power
sequencing is necessary.
• Use Eq. (4.6) and calculate the ramp time for each power supply
rail and verify that the power up sequence was achieved at a system
level with all the decoupling capacitors installed on the board.
• Use a current probe and measure the power up currents (Core and
IO) to make sure that there are no bus contentions. Keep in mind
that all capacitors appear like a short circuit to ground when they
start from a zero volt state, so the surge current may be higher than
expected during start-up. Be sure to provide adequate margin in
the power supply design to avoid false-triggering the over-current
protection circuits. A good rule-of-thumb is adding a 50% margin
to the maximum current consumed in the design.
Power Supply Design Considerations 63
4.3 SUMMARY
As demonstrated in this chapter, selecting the right power supply architec-
tures for the DSP system including surrounding analog/digital circuits and
doing the system floorplan design are the two most important critical tasks
for designers to get done first before getting started on the actual imple-
mentation. Good power integrity is key for achieving a low noise and low
EMI system design and here is a list of recommendations to improve the
probability of success:
REFERENCES
The number one root cause of system related problems is due to inade-
quate power supply decoupling around the DSP and or other surrounding
circuits such as DDR, clocks, analog-to-digital and digital-to-analog con-
verters etc. The most challenging task for designers is to determine the
best decoupling techniques to achieve low noise and high performance. In
general, component manufactures provide a conservative recommendation
for power supply decoupling, but in many cases, it is not practical to fol-
low this recommendation because of PCB space availability, power con-
sumption, EMI or safety requirements. Also, component manufactures
always provide development platforms for designers to evaluate and these
platforms typically are a lot larger than the actual design and are not re-
quired to be FCC certified, so copying what was done on the development
platform is not a guaranteed that the design will be successful. This chap-
ter will discuss three important topics for designers: 1. a general rule-of-
thumb decoupling method, 2. an analytic decoupling method and 3. how
to make design tradeoffs to achieve the best noise performance possible.
Once designers select and design a power supply for the DSP, the next step
is to determine the decoupling capacitors needed to ensure that the power
supply droop under all dynamic operating conditions is lower than the
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_5, 67
© Springer Science+Business Media, LLC 2010
68 Chapter 5
U19A
14
+ 1
3
2
-
Load Cap
7
14081/SO
In Figure 5.1, the DSP labeled U19A is driving a capacitive load and is
switching at a fast rate. Now, let’s assume that the regulator is placed 5
inches away from the DSP and is routed with a 5 mil trace to the DSP.
During fast switching, the power supply trace becomes an open circuit be-
cause of the parasitic inductance associated with the trace. This generates a
large voltage droop at the pin of the DSP which can be estimated as fol-
lows [1].
dI
Droop = L( Max ) (5.1)
dt
where L is the parasitic inductance.
dI 1.52ΔV
Max = C (5.2)
dt (Tr ) 2
dI 1.52ΔV
Droop= L(Max ) = L[ C]
dt (Tr )2
⎡1.52(2.64) ⎤
= 75x10−9 ⎢ −9 2
x50x10−12 ⎥ = 3.76V .
⎣ (2x10 ) ⎦
This example demonstrates that for a 5 inch trace, 2ns signal, 50pF load
and 3.3V IO, the maximum power supply droop is 3.76V. This level of
droop is certain to cause random system failures. To compensate, decoup-
ling capacitors are placed close to the DSP to provide the required charge
during switching. What is the best method to filter the noise from the DSP
system? Noise characteristics differ so much from system to system that no
one method guarantees low noise and low radiation for all cases. However,
designers can apply best practices outlined here to minimize the noise and
to improve the probability for success. Before going into the decoupling
techniques, it is important to understand the characteristics of the common
components (capacitors, inductors and ferrite beads) being used to filter
out the power supply noise.
CAP+
ESR
ESL
CAP-
The series equivalent circuit for a capacitor has three different compo-
nents: equivalent series resistance (ESR), equivalent series inductance
(ESL) and the capacitance itself. The self-resonant frequency happens at
the point where the impedance of the capacitor, C, is equal to the imped-
ance of the inductor, L.
1
Z C
, capacitor =
ωC
,
Z L
, inductor = ωL, where L is inductance.
At resonance, ZL is equal to Zc or
1
= ωL ,
ωC
1
ω
2
= ,
LC
1
ω= , where ω = 2πf.
LC
Therefore, the self-resonant frequency is
1
fR = . (5.3)
2π LC
As shown in the self-resonance equation, lower capacitance and lower in-
ductance yield a higher resonant frequency. For a given capacitance value,
choosing a smaller surface mount component achieves a higher self-
resonant frequency. Because a smaller component package typically has
lower parasitic and lead inductance. The whole decoupling concept is to
provide a low impedance path from the power supply to ground and to
shunt the unwanted RF energy. This means that choosing a capacitor with
high capacitance but with low inductance is very important. The problem
with this concept is that higher capacitance comes in a larger package
which yields higher parasitic inductance. In many cases, it is better to use
many capacitors with different values to decouple the DSP.
Power Supply Decoupling 71
Figure 5.3 shows the capacitor frequency response. For a particular ca-
pacitor, the impedance decreases with frequency and reaches the lowest
impedance point at resonant frequency, fR. For frequencies above the
resonant frequency, the impedance of the capacitor is dominated by the
parasitic inductor, ESL. This causes the impedance to increase with fre-
quency. It is recommended to operate in the capacitive region of the curve
as this region guarantees a close to ideal impedance response of the capaci-
tor.
There are many different types of capacitors and which type to use de-
pends on the voltage, temperature and frequency of the design. For exam-
ple, low frequency filtering requires a large electrolytic aluminum or tanta-
lum capacitor with a value of 10uF or higher, while high frequency
filtering needs a small film or ceramic capacitor with a value of less than
10uF. Selecting the wrong capacitor type can negatively affect the per-
formance of the system, so designers must carefully review the component
specifications and the applications before making the selection. Table 5.1
shows the electrical characteristics of different types of capacitors com-
monly being used.
1MHz for
¾ X7R/Y5R 1pF – 3.3uF Non-linear varia- 0.1uF sur-
tion with tempera- face mount
ture capacitor
¾ Z5U/Y5U 0.001pF – 10uF Poor temperature
and voltage
stability
IND+
DCR
IND-
1
fR = .
2π LC
Figure 5.5 shows the inductor frequency response. For a particular induc-
tor, the impedance increases with frequency and reaches the highest im-
pedance point at resonant frequency, fR. For frequencies above the reso-
nant frequency, the impedance of the inductor is dominated by the
parasitic capacitor, C, and this causes the impedance to decrease with fre-
quency. It is recommended to operate in the inductive region of the curve
as this region guarantees a close to ideal impedance response of the
74 Chapter 5
inductor. Like capacitors, there are different types of inductors and the
two main ones are air core and magnetic core. Air core is the coil with air
or insulating core and magnetic core is the coil wrapped around magnetic
materials such as iron and ferrite. Inductors are commonly being used in
RF and high power circuits but are rarely being designed in high speed
DSP systems. Because it is better and lower cost to use ferrite beads to
isolate and filter the noise in DSP systems.
Here are general rules for using inductors to filter noise in a DSP system:
Ferrite+
DCR
Ferrite-
10uF 10uF
0.01 0.01 0.01 0.01
0.01 uBGA
0.01
2 xI CoreMax
I C Re gion = xM (5.4)
N
If the maximum current specification is not available in the
data sheet, then estimate the maximum current by multiplying
the typical current by 2 as in Eq. (5.5).
4 xI CoreTyp
I C Re gion = xM (5.5)
N
• Calculate the total decoupling capacitance for the region by
applying Eq. (5.7) below.
dVCore
I C Re gion = CCore , (5.6)
dt
dt
CCore = I C Re gion , (5.7)
dVCore
where dt is the fastest rise-time in the region and dV is the
maximum ripple allowed for the Core voltage, assuming
10mv ripple.
I IO
I IO Re gion = xJ (5.8)
K
• The total IO current is not equal to the IO current sourcing and
sinking defined in the DSP data sheet. The majority of the to-
tal IO current depends on the external loads, for example re-
Power Supply Decoupling 81
VIO
I IOTrans = (5.9)
Zo
• In Eq. (5.10), the total IO current for the region is equal to the
IO current of the DSP itself plus the IO current driving the
transmission lines.
VIO
I IOTotal = I IO Re gion + Jx (5.11)
Zo
• Calculate the total decoupling capacitance for the region by
applying Eq. (5.13) below.
dVIO
I IOTotal = C IO , (5.12)
dt
dt
C IO = I IOTotal , (5.13)
dVIO
where dt is the fastest rise-time in the region and dV is the
maximum ripple allowed for the IO voltage, assuming
50mv ripple.
tal decoupling capacitance [4]. Use one bulk capacitor per re-
gion to minimize the parasitic inductance between the bulk
and the decoupling capacitors.
AA1 AA21
REGION 2
REGION 1 REGION 3
(digital PLL) (analog PLL, Y21)
REGION 4
A1 A21
Since the Core and I/O voltage operate at different frequencies, they re-
quire separate decoupling calculations. The following shows the steps
needed to calculate and select the decoupling capacitors for both Core and
I/O supplies.
To find the decoupling capacitance, plug the peak current, the rise-time
and the maximum ripple voltage parameters into Eq. (5.7) below and solve
for C. It is acceptable to assume that the maximum ripple voltage is 10mV
for Core and 50mV for IO and the typical rise-time is 2nS.
dt
CCore = I C Re gion
dVCore
Use the capacitor Eq. (5.13) below to calculate the total capacitance for the
IO voltage decoupling:
dt
C IO = I IOTotal
dVIO
Now lets calculate the total capacitance required for each region.
Region 1: Total Core capacitance,
(2nS )
CCore = 157 mA = 0.03uF
(10mV )
There are 3 Core voltage pins operating at 150MHz (CPU frequency) and
8 I/O voltage pins operating at 40MHz (EMIFS frequency). It would be
desirable to use multiple capacitors for the multiple supply pins, but there
is a physical limitation due to the limited space available around the de-
vice. For the DSP [5] package, there is enough board space to place about
4 or 5 capacitors per region. In this case, select two capacitors with the to-
tal capacitance of around 0.03uF. At least one of the capacitors should
have a self-resonant frequency around 150MHz to decouple the Core volt-
age pins in Region 1. Then, select three capacitors with a total capacitance
of around 0.08uF with at least one of the capacitors having the self-
resonant frequency around 75MHz to decouple the I/O voltage pins in Re-
gion 1.
In summary, for Core voltage in Region 1, use two 0.022uF (0.044uF to-
tal) ceramic capacitors and, for the I/O voltage, use three 0.033uF
(0.099uF total) ceramic capacitors.
The next step is calculating the bulk capacitors for both Core and IO. Bulk
capacitor placement is not as critical as decoupling capacitor placement.
But bulk capacitors are needed to filter the low frequency ripple typically
generated by switching power supply and to recharge the decoupling ca-
pacitors.
A rule of thumb is to select bulk capacitors with at least ten times the total
decoupling capacitance. Lets use 40 times to be conservative. For the Core
voltage,
40 x total Core capacitance = 40 x (0.03uF) = 1.2uF for Region 1 of
the Core voltage
and 40 x total IO capacitance = 40 x 0.08uF = 3.2uF for Region 1
of the IO voltage
As mentioned earlier in this chapter, the best technique is adding 4 bulk
capacitors to 4 regions of the DSP and the smallest tantalum capacitor
available is 4.7uF. In this case, select 4.7uF tantalum bulk capacitors for
both IO and Core voltages in Region 1.
In summary, Figure 5.12 shows the complete schematic diagram for de-
coupling Region 1 of the DSP. Next is to repeat the same steps for Re-
gions 2, 3 and 4.
Power Supply Decoupling 87
There are 3 Core voltage pins operating at 150MHz (CPU frequency) and
4 I/O voltage pins operating at 40MHz (EMIFS frequency). For the DSP
[5] package, there is enough board space to place about 4 or 5 capacitors
per region. In this case, select two capacitors with the total capacitance of
around 0.03uF. At least one of the capacitors should have a self-resonant
frequency around 150MHz to decouple the Core voltage pins in Region 2.
Then, select three capacitors with a total capacitance of around 0.13uF
with at least one of the capacitors having the self-resonant frequency
around 75MHz to decouple the I/O voltage pins in Region 2.
In summary, for Core voltage in Region 2, use two 0.022uF (0.044uF to-
tal) ceramic capacitors and for the I/O voltage, use three 0.047uF (0.14uF
total) ceramic capacitors.
88 Chapter 5
The next step is calculating the bulk capacitors for both Core and IO. A
rule of thumb is to select bulk capacitors with at least ten times the total
decoupling capacitance. Lets use 40 times to be conservative. For the Core
voltage,
40 x total Core capacitance = 40 x (0.03uF) = 1.2uF for Region 2.
For the IO voltage,
40 x total IO capacitance = 40 x 0.13uF = 5.2uF for Region 2 of the
IO voltage
In this case, select 4.7uF tantalum capacitor for the Core voltage and 6.8uF
tantalum capacitor for the IO voltage in Region 2. Figure 5.13 shows the
complete decoupling schematic of Region 2.
6.8uF
4.7uF
0.047uF
0.022uF
0.047uF
0.022uF
0.047uF
There are 4 Core voltage pins operating at 150MHz (CPU frequency) and
6 I/O voltage pins operating at 40MHz (EMIFS frequency). For the DSP
[5] package, there is enough board space to place about 4 or 5 capacitors
per region. In this case, select two capacitors with the total capacitance of
around 0.042uF. At least one of the capacitors should have a self-resonant
frequency around 150MHz to decouple the Core voltage pins in Region 4.
Then, select three capacitors with a total capacitance of around 0.124uF
with at least one of the capacitors having the self-resonant frequency
around 75MHz to decouple the I/O voltage pins in Region 4.
In summary, for Core voltage in Region 4, use two 0.027uF (0.054uF to-
tal) ceramic capacitors and for the I/O voltage, use three 0.047uF (0.14uF
total) ceramic capacitors.
90 Chapter 5
The next step is calculating the bulk capacitors for both Core and IO. A
rule of thumb is to select bulk capacitors with at least ten times the total
decoupling capacitance. Lets use 40 times to be conservative. For the Core
voltage,
40 x total Core capacitance = 40 x (0.054uF) = 2.16uF for Region
4.
For the IO voltage,
40 x total IO capacitance = 40 x 0.14uF = 5.64uF for Region 4 of
the IO voltage
In this case, select 4.7uF tantalum capacitor for the Core voltage and 6.8uF
tantalum capacitor for the IO voltage in Region 4 as shown in Figure 5.15.
Table 5.3 shows a summary of all the capacitors calculated for the 4 re-
gions of the DSP and Figure 5.16 shows the complete schematic.
In summary, there should be two bulk capacitors per region, one for Core
and one for IO, and as many decoupling capacitors as space allows. Figure
5.17 shows a very good example of the capacitors placement on the bottom
side of the PCB. The Core decoupling capacitors and four large bulk ca-
pacitors are placed on the interior of the BGA package in the open space
right under the DSP. The IO decoupling and bulk capacitors are placed on
the perimeter of the BGA package. This is possible because this particular
BGA package is not a full BGA package where all the balls are fully popu-
lated on the bottom of the package.
• Power and ground pins need to be closest to the lanes. This allows
the shortest connection paths to the capacitors.
Power Supply Decoupling 93
1
Ferrite Bead, Z Output
3
2
C1 C2
7
DSP
1
Ferrite Bead, Z1 Ferrite Bead, Z2 Output
3
2
C1
7
DSP
Ferrite Bead
VDD
Power TI DSPA
14
+ 1
3
C1 C2 2
-
Load Cap
7
Figure 5.21. Pi Filter Circuit for Power Supply Isolation
Lz
, where Lz is the inductance of the ferrite bead Z. (5.14)
2xC1
For this special Pi filter assuming C1 equal to C2, the corner frequency of
the 3-pole filter [6] is
1
fC = . (5.15)
π Lz C1
Figure 5.22 shows the frequency response of this special Pi filter.
96 Chapter 5
1
fC = = 200MHz.
π Lz C1
Lets select a ferrite bead with 100 ohms impedance at 100MHz and calcu-
late Lz. The impedance, Z, of the ferrite bead is
Z = 2πfL = 100 ohms,
100 100
Lz = = = 0.16uH .
2πf 2π (100 x106 )
Now, calculate C1 by substituting Lz and fc into the Eq. (5.15) and solve for
C1.
Power Supply Decoupling 97
1
fC = = 200MHz,
π Lz C1
1
200 x106 = ,
π (0.16 x10− 6 )C1
C1 = 15.8pF.
Therefore, the Pi filter has two 15.8pF capacitors and one 0.16uH inductor.
Now, lets use an analog circuit simulator [7] to verify the design.
VF1
VF2
R1 71 L1 160n
C2 15.8p
C1 15.8p
VG1
+
R2 71
To match the source and load impedance, use Eq. (5.14) and calculate R1
and R2 values. In this case, R1 = R2 = 71 ohms for C1 = 15.8pF and
Lz=0.16uH.
Figure 5.24 shows the simulation results of the circuit model shown in
Figure 5.23. In the pass-band from DC to 100MHz, the circuit shows a
-6dB attenuation. This is because the voltage divider formed by the 71
ohm source resistor and 71 ohm load resistor. In this case, the attenuation
is
VF 2
= 20 log10 , where (5.16)
VF 1
71
VF 2 = VF 1 . (5.17)
71 + 71
Now, substitute Eq. (5.17) into Eq. (5.16) and solve calculate the attenua-
tion.
71
Attenuation = 20 log10 = −6dB .
71 + 71
This correlates with the simulation results showing a -6dB signal attenua-
tion within the pass-band.
For the filter corner frequency, the simulation results show that the signal
starts rolling off at 100MHz with a slope of 60dB/decade. This is correct
as this is a 3-pole low-pass filter and each pole has a 20dB/decade slope.
Since the -3dB corner frequency of each pole is at 200MHz and the Pi fil-
ter has 3 poles (2 capacitors and 1 ferrite bead), the combined corner fre-
quency at 200MHz has a -9dB attenuation as shown in the simulation.
Lz
, where Lz is the inductance of the ferrite bead Z. (5.17)
2xC1
For this special T filter assuming L1 equal to L2, the corner frequency of
the 3-pole filter [6] is
1
fC = . (5.18)
π Lz C1
Figure 5.25 shows the frequency response of this special Pi filter.
1
fC = = 200MHz.
π Lz C1
Lets select a ferrite bead with 100 ohms impedance at 100MHz and calcu-
late Lz. The impedance, Z, of the ferrite bead is
Z = 2πfL = 100 ohms,
100 100
Lz = = = 0.16uH .
2πf 2π (100 x106 )
Now, calculate C1 by substituting Lz and fc into the Eq. (5.18) and solve for
C1 .
1
fC = = 200MHz,
π Lz C1
1
200 x106 = ,
π (0.16 x10− 6 )C1
C1 = 15.8pF.
Therefore, the T filter has two 0.16uH ferrite beads and one 15.8pF capaci-
tor. Now, lets use an analog circuit simulator [7] to verify the design.
VF1
VF2
R1 71 L1 160n L2 160n
C1 15.8p
VG1
+
R2 71
To match the source and load impedance, use Eq. (5.17) and calculate R1
and R2 values. In this case, R1 = R2 = 71 ohms for C1 = 15.8pF and
L1=L2=0.16uH.
Figure 5.27 shows the simulation results of the circuit model shown in
Figure 5.26. In the pass-band from DC to 100MHz, the circuit shows a -
6dB attenuation. This is because the voltage divider formed by the 71 ohm
source resistor and 71 ohm load resistor. In this case, the attenuation is
VF 2
= 20 log10 , where (5.19)
VF 1
71
VF 2 = VF 1 . (5.20)
71 + 71
Now, substitute Eq. (5.20) into Eq. (5.19) and solve calculate the attenua-
tion.
71
Attenuation = 20 log10 = −6dB .
71 + 71
This correlates with the simulation results showing a -6dB signal attenua-
tion within the pass-band.
102 Chapter 5
For the filter corner frequency, the simulation results show that the signal
starts rolling off at 100MHz with a slope of 60dB/decade. This is correct
as this is a 3-pole low-pass filter and each pole has a 20dB/decade slope.
Since the -3dB corner frequency of each pole is at 200MHz and the T filter
has 3 poles (2 ferrite beads and 1 capacitor), the combined corner fre-
quency at 200MHz has a -9dB attenuation as shown in the simulation.
Table 5.4 shows a comparison between T and Pi filters.
Pi Filter T Filter
1. Components Two capacitors and one ferrite Two ferrites and
Required one capacitor
5.3 SUMMARY
• Use an analog simulator [7] and simulate the design to verify all
the calculations before going into layout.
104 Chapter 5
REFERENCES
[3] Murata Manufacturing Co. (2009) Murata EMI Filter Selection Simu-
lator.
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_6, 105
© Springer Science+Business Media, LLC 2010
106 Chapter 6
fREF fOUT
/N PFD CP VCO
R
C
C
/M
M
f OUT = f IN
N
The following table provides a brief description of each block shown Fig-
ure 6.1 for the APLL.
M
f out = fin , where M is the PLL
N (6.1)
• Use the input clock to trigger the scope and set the scope in
the Infinite Persistence mode.
• The deviation is measured from the first rising edge to the Nth
cycle. The “fuzz” shown on the scope in Figure 6.3 is the long
term jitter.
• Set the scope in the Infinite Persistence mode and trigger the
PLL clock output on the rising edge.
• The DPLL does not have analog filter components such as ca-
pacitors which can cause leakage current. This leads to lower
power consumption.
• The DPLL design can be optimized for low jitter. But it may
not be acceptable for jitter sensitive designs such as USB, au-
dio and video clocks.
The disadvantages of the DPLL are:
Figure 6.6 shows a typical DPLL architecture [6] and Table 6.2 describes
the function of each block in the architecture.
/M
M
f OUT = f IN
N
Designers need to be careful when injecting a signal onto the power supply
to do jitter measurements. The nature of the signal used for simulating a
noisy power supply condition can have a major impact on the PLL jitter. A
squarewave signal with a frequency less than the PLL bandwidth charac-
terizes the worst case PLL jitter. As far as the amplitude of the noise, the
peak-to-peak voltage has to be within the power supply limits. For exam-
ple, for a 1.6V +/-3% Core, the maximum acceptable peak-to-peak noise is
96mV (-48mV min and +48mV max).
114 Chapter 6
VDD_3V3 VDD_PLL
L
FERRITE BEAD
C1 C2
The Pi filter circuit consists of one ferrite bead, L and two capacitors, C1
and C2. This circuit provides both input and output isolation where noise
from the 3.3V supply is attenuated by the ferrite bead and the C2 capacitor
and noise generated by the PLL circuit is isolated by the ferrite bead and
the C1 capacitor. Refer to Chapter 5 for the filter design and simulation in-
formation.
Phase-Locked Loop (PLL) 115
VDD_3V3 VDD_PLL
L1 L2
C1
A T filter consists of two ferrite beads and one capacitor as shown in Fig-
ure 6.8. Just like in a Pi filter, 3.3V supply noise is attenuated by the L1
ferrite bead and the C1 capacitor and PLL noise is isolated by the L2 ferrite
bead and C1 capacitor. Refer to Chapter 5 for the filter design and simula-
tion information.
Both Pi and T circuits are good for filtering high frequency noise but they
are not as effective for low frequency filtering since ferrite beads have al-
most zero AC impedance at low frequency. The Pi circuit has an advantage
over the T circuit. Because this topology makes it possible to place the ca-
pacitor closer to the PLL voltage pin that ensures low impedance to ground
and also the smallest current loop area, which reduces noise and EMI.
For low frequency isolation, there are two common techniques, Pi filter
with large bulk capacitor and linear voltage regulator.
VDD_3V3 VDD_PLL
R
+
C1 C2 C3
10uf
One method for low frequency filtering is shown in Figure 6.9, where a re-
sistor R replaces the ferrite bead and a bulk capacitor C3 (10uF to 33uF) is
added to the circuit. Low frequency noise is attenuated by the resistor R
and the bulk capacitor C3. The resistor needs to be selected such that the
voltage drop across the resistor is negligible. The low frequency -3dB cor-
ner for this filter is approximated by Equation 6.2. Notice that C1 and C2
are negligible in this case, since its value is a lot lower than the bulk ca-
pacitor C3.
1
f−3dB =
2πRC3 (6.2)
1
f−3dB = = 1.5KHz,
2πRC3
RC3 = 1.06x10-4,
Let R = 10 ohms,
Phase-Locked Loop (PLL) 117
C3 = 10.6uF or 10uF.
VDD_3V3 R3 10 VDD_PLL
C2 10n
C1 10n
C3 10u
VDD_5V VDD_PLL
Linear
Reg
+
C1 C2 C3
10uf
One issue with using a linear regulator is that it does not reject high fre-
quency very well. As shown in Figure 6.13, the ripple rejection is ap-
proaching 0dB (no rejection at all) for noise that is higher than 1MHz.
This high frequency noise can cause more jitter in the PLL.
rite bead and capacitors so there is no IR drop across the resistor as shown
in the previous example. The final circuit is shown in Figure 6.14.
L1
VDD_5V LINEAR VDD_PLL
REGULATOR
+
C3 C4
C2 C1
6.4 SUMMARY
Because of low power consumption and fast response time, most of the
PLL designs integrated in the DSP today are based on digital PLL con-
cepts. As discussed, DPLL is very sensitive to power supply and input
noise, so proper design noise isolation filters are required to achieve the
lowest jitter possible. The best approach is using a combination of Pi filter
and linear regulator as shown in Section 6.3.2. This may not be possible
due to PCB space limitation so designers have to make design compro-
mises. If there is not enough room for the regulator circuit, then imple-
menting a Pi filter using a resistor instead of a ferrite bead is the second
best approach. This has low frequency and high frequency filtering char-
acteristics as demonstrated in Section 6.3.1.
120 Chapter 6
REFERENCES
[3] Wavecrest (2002) Examining Clock Signals And Measuring Jitter with
the WAVECREST SIA-300. Application Note No. 142.
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_7, 121
© Springer Science+Business Media, LLC 2010
122 Chapter 7
In Figure 7.2, 1 LSB is defined as one Least Significant Bit and at every
sampling point; the analog level can vary from ±1/2LSB from the center.
This is known as a quantization error. The LSB voltage, VLSB, is equal to
Vref
V LSB = , where Vref is the reference voltage (7.1)
2N
and N is the number of bits.
For the 3-bit ADC and Vref of 8V, the VLSB is equal to 1V. Table 7.1
shows an example of a 3-bit ADC sampling an 8V analog input signal.
Overall, the equation for calculating the voltages of the ADC [1] is as fol-
lows:
Vref(b12-1 + b22-2 + b32-3 + …… + bN2-N) = Vin ± Vx, (7.2)
where Vx is
1 1
− V LSB ≤ V X ≤ V LSB and b1 is the most significant
2 2
bit and bN is the least significant bit.
Figure 7.3 shows a practical ADC block diagram where the analog input
has to be band-limited before being converted to digital word. This is be-
cause the Nyquist sampling theory defined that the sampling clock has to
be at least two times the analog bandwidth to prevent aliasing. Aliasing is
the image of the analog signal folded back into the frequency of interest;
aliasing degrades the ADC performance. Therefore, an anti-aliasing filter
must be placed at the input of the ADC.
ANALOG ANALOG-TO-
ANTI-ALIASING DIGITAL
INPUT DIGITAL
FILTER OUTPUT
CONVERTER
Sampling
Clock, fS
7.2.1 Sampling
An ADC utilizes the sampling clock, fS, to sample the analog input and
represents the level in a digital word as shown in the previous section.
Sampling is equivalent to amplitude modulating the signal, fA, into a car-
rier equal to the sampling frequency and generates a frequency spectrum
shown in Figure 7.4.
Data Converter Overview 125
Now assume that the quantization error voltage between the quantized lev-
els and the sampled voltage is uniformly distributed between –q/2 and +q/2
where q is equal to 1 LSB voltage. In this case, the Probability Density
Function, fQ(x), is shown in Figure 7.8 where the area under the curve is
equal to one.
1/2
⎡1 T
⎤
∫0 V ( t ) dt ⎥⎦
2
VIN(rms) = ⎢ , where V(t) = Acos(2πfct)
⎣T
and A is the zero-to-peak voltage as shown in Figure 7.9.
Therefore,
1/2
⎡1 T
⎤
∫A ( 2 π f c t ) dt ⎥
2 2
VIN(rms) = ⎢ cos . Since
⎣T 0 ⎦
1 + cos( 4 π f c t )
cos 2
( 2π f ct ) = ,
2
1/ 2
⎡ A2 T
1 + cos( 4 π f c t ) ⎤
VIN(rms) = ⎢
⎣ T
∫0 2
dt ⎥
⎦
1/ 2
⎡ A2 T ⎤ A
= ⎢ ⎥ = . (7.5)
⎣ T 2⎦ 2
A is equal 1/2VREF since VREF is equal to the peak-to-peak voltage of a si-
nusoidal wave as shown in Figure 7.10 and A is the zero-to-peak value.
Data Converter Overview 129
V REF
VIN(rms)= . (7.6)
2 2
Or
VREF = Vpeak-to-peak = 2.828VIN(rms).
The signal-to-noise or SNR is defined as the log of the ratio of the input
RMS voltage over the quantization noise.
V IN ( rms )
SNR = 20log10 . (7.7)
V Q ( rms )
Substitute Eqs. (7.6) and (7.4) into Eq. (7.7),
⎛ V REF 12 ⎞
SNR = 20log10 ⎜⎜ ⎟ .
⎟
⎝2 2 q ⎠
Since q = VLSB and from Eq. (7.1)
Vref
V LSB = ,
2N
130 Chapter 7
⎛ 3 N ⎞
SNR = 20log10 ⎜⎜ 2 ⎟
⎟
⎝ 2 ⎠
= 6.02N + 1.76 dB. (7.8)
Eq. (7.8) indicates that the performance of an ADC depends on the number
of bits used to quantize the analog signal. It is roughly 6dB per bit. This
equation was derived assuming that the only error in the system is quanti-
zation error. In the real world, other factors such as power supply noise
and clock jitter generate additional errors that degrade the signal-to-noise
significantly. So, another equation to measure the overall performance of
the ADC is
SNR − 1 . 76
ENOB, Effective Number of Bits = . (7.9)
6 . 02
For example, if a 16-bit ADC has an SNR specification of 86dB, the Effec-
tive Number of Bits is
86 − 1 . 76
ENOB = = 14 .
6 . 02
What this means is that the 16-bit ADC only performs at a 14-bit level due
to quantization noise and other system related noise, such as power sup-
plies, clocks and others degrading its performance.
Similarly to ADC, for the 3-bit DAC and Vref of 8V, the V LSB is equal to
1V. Table 2 shows an example of a 3-bit DAC taking a digital word and
converting it to an equivalent analog level assuming that the sampling er-
ror is ±0.5LSB.
000 0V 0.5V
001 1V ±0.5V
010 2V ±0.5V
011 3V ±0.5V
100 4V ±0.5V
101 5V ±0.5V
110 6V ±0.5V
111 7V ±0.5V
132 Chapter 7
Figure 7.12 shows a practical DAC block diagram where the digital input
is being converted to analog and the Reconstruction Filter eliminates the
sampling noise modulated on the analog waveform. The design of this fil-
ter depends on which DAC is being used and how much noise suppression
is necessary to achieve a certain signal-to-noise specification. The filter
topologies and design methodologies are covered in Chapter 8 of this
book.
The filter requirements for the DAC are analog bandwidth, fA , samples per
second input, fs , and stop band attenuation. As shown in Figure 7.4, the
sampled data has an image closest to the band of interest at
Image = fs – fA.
For example, a video signal has a bandwidth of 6MHz and the DAC input
is 27 MSPS (Mega Samples Per Second). The image of this video signal is
Image = 27MHz – 6MHz = 21MHz.
If the system needs a 60dB signal-to-noise performance, then the image
needs to be attenuated at least 60dB at 21MHz. The details of the filter de-
sign are in Chapter 8.
• Sampling Frequency
• Differential Non-linearity
• Integral Non-linearity
S S
SNR − 1 . 76 55 − 1 . 76
ENOB= = ≅ 9 .
6 . 02 6 . 02
ENOB indicates that the performance of this 10-bit ADC is equivalent to
the performance of an ideal 9-bit ADC. Achieving the theoretical resolu-
tion is very difficult so getting a 9-bit performance out of a 10-bit ADC is
considered as a very good ADC.
134 Chapter 7
The gain stages of 2 and 0.5 can be implemented using op amps as shown
in Chapter 8. Another option to design a gain of 0.5 is using a resistor di-
vider to divide the input voltage by half. Here is an example.
For the voltage divider shown in Figure 7.15, the output voltage is
R2
VO= V IN . (7.30)
R1 + R 2
If R2 = R1, then Vo is half of VIN.
Regarding the gain circuit of 2, refer to Chapter 8 for more details.
Similarly, for the DAC output voltage range, if the DAC being used has a
low level voltage output that is not compatible with the next stage or not
compliant to some input and output standards, then a gain circuit at the
DAC output is required. Refer to Chapter 8 to design this amplifier circuit.
Figure 7.17 [4] shows a DAC DNL error where the input 010 code has an
analog range larger than 1 LSB voltage. In this case, the error is +1/4LSB.
Figure 7.19 [4] shows an INL error caused by the DAC. In this case, the
maximum deviation from the ideal curve happens at the 011 input digital
code and the error is equal to 1/2LSB.
138 Chapter 7
7.5 SUMMARY
• DNL and INL errors occurred in ADC and DAC can not be elimi-
nated using system design techniques. These are inherent in the
data converter itself. Choose the converters with the lowest DNL
and INL specifications if possible.
Data Converter Overview 139
Overall, the design goal for the data converter system is to match the per-
formance specified in the converter data sheet. For example, if the con-
verter has an 80dB SNR specification, the system performance target
should be 80dB SNR. This is the best performance that can be achieved
because all other components (anti-aliasing filters, amplifiers, power sup-
plies, and reconstruction filters) around the converter tend to generate ad-
ditional noise and errors to the system.
140 Chapter 7
REFERENCES
This chapter presents passive and active filter topologies and design tech-
niques, including practical design examples and system simulations. In
DSP systems, there are analog filters required for signal conditioning and
limiting the bandwidth before sampling. To design these filters, designers
need to be knowledgeable of operational amplifiers, DC biasing circuits,
AC coupling techniques and traditional passive components like inductors,
capacitors and resistors.
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_8, 141
© Springer Science+Business Media, LLC 2010
142 Chapter 8
In summary, if gain is not required and source and load impedances are
known, then it is better to go with passive filters. Now, if impedance isola-
tion and gain are required, then active filters would be better.
Vin R1 Vout
C1 Rload
Assuming Rload is much higher than R1, the -3dB corner frequency for the
filter in Figure 8.1 is
1
f − 3 dB = . (8.1)
2 π R 1C 1
Vin Rs L1 Vout
C1 RL
The frequency response of the 2nd order filter circuit has amplitude peaking
at the -3dB corner and the amount of this peaking depends on the ratio of
RL and RS. This peaking typically does not affect the circuit performance
as long as the noise at the corner frequency is very low, so it is crucial for
designers to verify that the frequency response and signal-to-noise over the
passband are as expected. It is very common to fine tune RL and RS to get
the frequency response needed for the application.
For RL >> RS, the -3dB corner frequency for the circuit in Figure 8.3 is
R + R
f − 3 dB = L S
. (8.2)
2π L 1C 1 R L
7 K + 70 .
20 KHz =
2π ( 7 K ) L 1C 1
C1 1n RL 7k
• Virtual ground: The voltages at the negative and positive inputs are
equal.
U1 LM318
VN, Inverting Input
- VO, Output
+
+
VP, Non Inverting Input
R2
I2
R1
- VO
I1 VN
+
+
VP
U1 LM318
+
VIN
V N
I1 = ,
R1
Substitute VIN into Eq. (8.7) and solve for the gain,
VO/VIN.
VO R2
Gain = = + 1 . (8.8)
V IN R1
B. Inverting Amplifier
R2
I2
R1
- VO
I1 VN
+
+
+
VIN
U1 LM318
V − V
I1 = IN N ,
R1
Replace VN with zero in Eq. (8.9) and solve for the gain,
VO/VIN.
VO R2
Gain = = − . (8.11)
V IN R1
R2
-VDD
VIN R1
- VO
VN
+
+ U1 LM318
+VDD
The rule is to always bias the positive terminal of the op amp as shown in
Figures 8.10 and 8.12 and the bias voltage must be set at the level where
the output has the maximum swing as shown in Figure 8.11.
For the op amp with a single rail power supply, the bias voltage must be
half of the power supply to guarantee maximum symmetrical swing as
demonstrated in Figure 8.11. The circuit shown in Figure 8.12 has a volt-
age divider formed by two equal resistors, R3 and R4, to generate a bias
voltage at half of the power supply rail, +VDD. For an ideal op amp, the
positive and negative input voltages are equal. But this is not the case in
the real word where there is always some small offset voltage between the
two inputs. This offset voltage is in the range of micro volts and can be an
issue for small signal detection and processing applications. In general,
this offset voltage is not a concern for video, audio and communication de-
signs, but it is good to keep it as low as possible. To minimize the offset
voltage, set the parallel combination of R3 and R4 equal to R2.
R3R4
R2 = R3//R4 = ,
R3 + R4
2
R3
and R3 = R4 so R2 = .
2 R3
R2
V IN R1
- VO
VN
+
+ U1 LM318
R3 +VDD
+VDD/2
R4
Again, only the positive terminal of the op amp needs to be biased at half
of the power supply. Due to virtual ground rule, the negative DC voltage
is at +VDD/2 and this also sets the output at +VDD/2. Now the whole cir-
cuit is DC balanced, which enables the signal to swing symmetrically
around +VDD/2.
Another important rule to remember is that if a point in the circuit is con-
nected to a DC voltage, the connection point becomes an AC or signal
ground. So for a resistor divider circuit in Figure 8.12, it is good to add a
bypass capacitor C1 in parallel with R4 to provide a good AC ground as in
Figure 8.13. This capacitor does not affect the signal path at all, since the
capacitor is on the positive input of the op amp which only has a DC bias
voltage.
R2
VIN R1
- VO
VN
+
+ U1 LM318
R3 +VDD
+VDD/2
R4
C1
In Eq. (8.13), it is preferable to select C1 such that the -3dB corner fre-
quency is low enough to filter out the power supply noise. Now, let’s bias
a non-inverting amplifier circuit. The rule is the same as in the inverting
case where only the positive input of the op amp needs to be biased. Fig-
ure 8.14 shows the biasing circuit where R3 is connected to ground which
is in the middle +VDD and –VDD rails. Again, to minimize the offset
voltage, set R3 equal to the parallel combination of R1 and R2. So,
R1R 2
R3 = R1//R2 = , (8.14)
R1 + R 2
R2
-VDD
R1
- VO
VN
VIN
+
+ U1 LM318
VP
R3
+VDD
Similarly, Figure 8.15 shows a biasing circuit for the single rail supply
non-inverting amplifier. The resistors R3 and R4 bias the positive input at
half of the supply voltage. To minimize the offset voltage, set the parallel
combination of R3 and R4 equal to R2, since DC current does not flow
through R1 because of the DC blocking capacitor C3.
R3//R4 = R2,
R3R4
= R2. Since R3 = R4,
R3 + R4
R2
C3 R1
R4
- VO
VN
VIN
+
+ U1 LM318
VP
Rload
R3
VDD
R2
C3 R1
R4
- C2 VO
C1 VN
VIN
+
+ U1 LM318
VP
Rload
R3
VDD
From the input VIN looking into the circuit, the C1 capacitor and resistors
R3 and R4 in Figure 8.17 form a highpass filter and the corner frequency is
1
f − 3 dB = ,
2π (R 2 // R 3 ) C 1
R2 + R3
f − 3 dB = . (8.16)
2π ( R 2 R 3 )C 1
For the output AC-coupling capacitor C2, looking out from the op amp
output, the capacitor C2 and the resistor Rload form a highpass filter and its -
3dB corner frequency is
1
f − 3 dB = . (8.18)
2π R load C 2
So, R2 = R1.
Let R2 = R1 = 20KΩ, reasonable value for audio design.
From Eq. (8.15),
R3 = R4 = 2R2= 2(20K) = 40KΩ.
The input impedance is R3 in parallel with R4 which is
( 40 K )( 40 K )
= 20KΩ.
40 K + 40 K
40 K + 40 K
C = = 0 . 4 uF
2 π ( 20 )( 40 K )( 40 K )
1
158 Chapter 8
1
C = = 0 . 4 uF
2 π ( 20 )( 20 K )
2
1
C = = 0 . 4 uF .
2 π ( 20 )( 20 K )
3
C3 470n R1 20k
R4 40k
- C2 470n VO
C1 470n VN
VIN
+
+ U1 LM318
VP
Rload 20k
+
R3 40k
VG1
VDD 12
Now for the inverting circuit shown in Figure 8.20, from the input VIN
looking into the circuit, the gain of the circuit is modified by the imped-
ance of the capacitor and is equal to R2 divided by R1 plus the impedance
of the capacitor C2.
VO R2
= − , where Z2 is the (8.19)
V IN R1 + Z 2
impedance of C2 and is equal to
1
and f is frequency.
2 π fC 2
Substitute Z2 into Eq. (8.19). The magnitude of the gain, ignoring the
negative sign as the sign only indicates the output is inverted, is
VO 2 π fC 2 R 2
= . (8.20)
V IN 2 π fC 2 R 1 + 1
1
f corner = , where C2 and R1 (8.21)
2 π R 1C 2
is the input impedance looking into the circuit. Substitute the corner fre-
quency into Eq. (8.20) and the magnitude of the gain at this frequency be-
comes
VO R2
= . (8.22)
V IN 2 R1
R2
VIN C2 R1
- C3 VO
VN
+
+ U1 LM318
R5
R3 +VDD
+VDD/2
R4
C1
VO R2
Gain = = = 2 (neglect minus sign as it only indicates
V IN R1
the phase),
So, R2 =2R1.
Let R1 = 20K so R2 = 2(20K) = 40K.
From Eq. (8.12),
R3 = R4 = 2R2= 2(40K) = 80K.
The input impedance is equal to R1 which is 20K since the positive
and negative inputs of the op amp are equal, AC ground.
From Eq. (8.21), the input corner frequency (-6dB) is
1
20 Hz = ,
2 π R 1C 2
1
C = = 0 . 2 uF
2 π ( 20 )( 40 K )
2
1
C = = 0 . 4 uF
2 π ( 20 )( 20 K )
3
1
C = = 0 . 1 uF .
2 π ( 20 )( 80 K )
1
VG1
R5 20k
R3 80k
V1 12
C1 100n
R4 80k
All other design parameters and methodologies are the same as those
shown in the previous op amp design sections.
C4
R2
VIN C2 R1
- C3 VO
VN
+
+ U1 LM318
+
VG1
R5
R3
V1 12
C1
R4
Solution:
Use the circuit in Figure 8.24. From Eq. (8.23), the upper -3dB frequency
is
1
20 KHz = .
2π R 2C 4
From Design Example 8.2, R2 is 40K ohms. So,
1
C4 = = 200 pF .
2 π ( 40 K )( 20 KHz )
The final circuit and simulation are shown in Figure 8.25 and Figure 8.26
respectively. It is verified that the circuit frequency response has an upper
frequency limitation at 20KHz. This is the 1st order filter circuit with the
-3dB frequency at 20KHz and the slope decaying at 20dB per decade.
As mentioned earlier, the output of the inverting lowpass filter has a 180
degrees phase-shift as compared to the input. Figure 8.27 shows the simu-
lated gain of 2 and phase relationship of the input and output waveforms.
C4 200p
R2 40k
VG1
R5 20k
R3 80k
V1 12
C1 100n
R4 80k
Now to realize the non-inverting 1st order lowpass filter circuit, take the
non-inverting amplifier circuit and add a capacitor in parallel to R2 to limit
the op amp bandwidth and an RC filter at the positive input of the op amp
as shown in Figure 8.28. Op amp typically has a high gain bandwidth
product and can go unstable if the bandwidth is not limited to the operating
frequency range. Doing stability analysis is beyond the scope of this book
but designers can learn more in [2]. This new circuit is shown in Figure
8.28 and the -3dB corner is dominated by the resistor R5 and the capacitor
C5 ,
1
f − 3 dB = . (8.24)
2π R 5C 5
Analog Filter Design 167
1
f − 3 dB = . (8.25)
2π R 2C 4
Also, in Figure 8.28, the filter resistor R5 needs to be one tenth of the par-
allel combination of the resistors R3 and R4. This is to minimize the ef-
fects of the voltage divider formed by R5 and the parallel combination of
R3 and R4.
C4
R2
C3 R1
R4
- C2 VO
C1 R5 VN
VIN
+
+ U1 LM318
VP
Rload
R3
+
VG1 C5
VDD 12
1
20 KHz = .
2π R 5C 5
168 Chapter 8
1 1 R3R4
R5 = ( R 3 // R 4 ) = = 2K ,
10 10 ( R 3 + R 4 )
1
C = = 0 . 00398 uF .
2 π ( 20 K )( 2 KHz
5
)
Now, since the maximum upper frequency is 20KHz, lets limit the op amp
at 40KHz which is two times the signal bandwidth. This provides plenty
of bandwidth margins.
From Eq. (8.25),
1
40 KHz = ,
2π R 2C 4
1
C = = 200 pF .
2 π ( 20 K )( 40 KHz )
4
The final circuit and simulation are shown in Figure 8.29 and Figure 8.30
respectively. It is verified that the circuit frequency response has an upper
frequency limitation at 20KHz. This is the 1st order non-inverting filter
circuit with the -3dB frequency of 20KHz and the slope of -20dB/decade.
Figure 8.31 shows the simulated gain of 2 and phase relationship of the in-
put and output waveforms. The simulation results are correct and corre-
lated with the calculations very well.
C4 200p
R2 20k
C3 470n R1 20k
R4 40k
- C2 470n VO
C1 470n R5 2k VN
VIN
+
+ U1 LM318
VP
Rload 20k
+
R3 40k
VG1 C5 3.98n
VDD 12
C1
2Q
- VO
VIN R1 R2
+
+ U1 LM318
1 1
+
VIN C2 1/2Q
To calculate the values of the capacitors and resistors, use the following
equations.
Rnew = KmRold , where Km is the new resistance (8.27)
and Rold is the normalized resistance shown in Figure 8.32.
1
Cnew = C old , where Kf is the (8.28)
K fK m
corner frequency in Rad/s, K
f = 2πf; f is the corner fre-
quency in Hertz.
C1
R4
1
1
1
- VO
VIN R1 R2 R3
+
+ U1 LM318
1 Q
+
VIN C2 1/Q
R4 2 −1/Q 1
1+ = 1+ = 3 − . (8.28)
R3 1 Q
172 Chapter 8
All the values in the circuit are calculated by the same methods demon-
strated in the Sallen-Key Circuit with Gain = 1 section.
C1
R4
1
2-1/Q
1
- VO
VIN R1 R2 R3
+
+ U1 LM318
1 1
+
VIN C2 1
Let Km = 2x104.
For gain peaking = 4dB,
2
2Q
4 dB = 20 log 10 ,
4Q 2
−1
1 1
C1 new = C old =
K fK m (125663.7 )(2x104 )
= 398pF.
C2 old = 1/Q = 0.667,
1 0.667
C2 new = C old =
K fK m (125663.7 )(2x104 )
= 265pF.
The final circuit is shown in Figure 8.36 where C3, C5 and C4 came from
the Design Example 8.3. R6 and R7 form a voltage divider to bias the op
amp at half of the power supply voltage. The parallel combination of the
resistors R6 and R7 is selected to be 10 times larger than the total resistance
of R1 and R2. This guarantees that the bias resistors will not affect the
overall gain.
The simulations in Figures 8.36 and 8.37 show the following results:
o Gain peaking = 4dB
o Corner frequency = 20KHz
o Slope = -40dB/decade, 2nd order lowpass filter
174 Chapter 8
o Passband signal gain = 1.85 instead of 2. This is due to the bias re-
sistors R6 and R7 loading the signal down. Increasing the total R6
and R7 resistance reduces the effect but will cause problems with
not having adequate bias current required for the op amp. Another
option is reducing the total resistance of R1 and R2. This is a better
option but it requires a total redesign of the filter.
C1 398p
R4 20k
C5 470n
- C4 470n VO
VIN C3 470n R1 20k R2 30k R3 20k
+
+ U1 LM318
C2 265p R5 20k
+
VIN
R6 1M
V1 12
R7 1M
8.2 SUMMARY
As demonstrated throughout this chapter, filter designs are very compli-
cated and require to do thorough system analysis using a circuit simulator
such as [3]. A filter topology is selected base on the following criteria:
REFERENCES
[2] Franco Sergio (2002) Design with Operational Amplifiers and Analog
Integrated Circuits. McGraw-Hill, New York.
[4] Valkenburg M.E. Van (1982) Analog Filter Design. Holt, Rinehart and
Winston, New York.
Memory Sub-System
Design Considerations 9
The most critical bus in a DSP system today is the memory bus where a
large amount of ultra high speed data is being transferred from the DSP to
the physical memory devices and vice versa. The data on this bus are
switching very fast. The rise and fall times of the data, memory clocks,
control signals are approaching sub-nanosecond range. These fast tran-
sients generate noise, radiation, power supply droops, signal integrity, and
memory timing issues. This chapter covers memory sub-system design
techniques to minimize the effects of the high speed data propagating.
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_9, 177
© Springer Science+Business Media, LLC 2010
178 Chapter 9
Figure 9.1 shows the basic DSP and DDR interface and the signal defini-
tions are in Table 2.
DQS
1 0 1 0 1 0 1 0
D0
DQS
1 0 1 0 1 0 1 0
D0
For the ugly case in Figure 9.7, the overshoots and undershoots are so ex-
cessive that the peak of the overshoot crosses over the minimum input high
voltage, Vih, and causes false clocking. Also, these overshoots and under-
shoots generate a lot of noise and radiation. The rule-of-thumb is to fine
tune the signal until all the overshoots are much lower than the threshold
voltage, Vih.
• Add termination resistors on the data bus and the control signals.
Where to place the resistors on the bi-directional bus depends on
which device has higher drive strength. Place the resistors nearby
the device with the higher drive strength. For example, DDR
memory devices typically have strong buffers to allow for non-
embedded designs such as PC. In this case, always put the termi-
nation resistors right by the memory devices. Ideally, add termina-
tion resistors at the output of the device driving the bi-directional
bus. See the design examples in Figures 9.8 [3] and 9.9.
184 Chapter 9
• Isolate and decouple the DLL power supply and the VREF voltage
pins. For the DLL, follow the rules described in the PLL chapter,
Chapter 6. To generate VREF, use a resistor divider and divide the
memory power supply voltage, VDD, required for both DSP and
memory devices. See the design example in Figure 9.9 [1].
VREF
GENERATOR
TERMINATION
RESISTORS ON
THE DATA BUS
NEAR DDR
MEMORY
TERMINATION RESISTORS
DQS’s NEAR DDR MEMORY
REFERENCES
[1] Micron (2009) DDR2 SDRAM 1Gb: x4, x8, x16 DDR SDRAM.
http://download.micron.com/pdf/datasheets/dram/ddr2/1GbDDR2.pdf.
Once all the circuits have been designed, the next step is board layout. This
is a very critical step in the development process because the effectiveness
of the filtering circuits depends on where the components are placed rela-
tive to the DSP pins. Also, the board layout has a big effect on noise,
crosstalk and transmission line effects so optimizing the layout can mini-
mize these effects. This chapter covers the printed circuit board stackup
and layout techniques for low noise and EMI.
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_10, 187
© Springer Science+Business Media, LLC 2010
188 Chapter 10
• Are there buses, such as USB, Ethernet, and RapidIO, that re-
quire a tight differential impedance specification? If so, de-
signers need to follow the industry guidelines to control the
differential impedance of these buses.
The parallel capacitance, Cpp, between the power and ground planes is
calculated as
ε rA
C pp = k , (10.1)
d
where k is 0.2249 inches or 0.884cm,
εr, dielectric constant = 4.1 to 4.7 for FR4 type PCB,
A is area of the power and ground planes, and
d is the distance between the power and ground planes.
When using this topology, designers need to consider these points:
• The adjacent power and ground topology are not useable for
DSP systems that require many layers to route the signals out
from the DSP and interface with other circuits.
Figure 10.2 [1] shows a typical PCB stackup for the non-adjacent power
and ground topology. The power and ground planes are placed in Layer 5
and Layer 2 respectively. Layer 3 is best for routing high-speed traces
while Layer 1, Layer 4 and Layer 6 are acceptable. As shown in the figure,
190 Chapter 10
each of the routing layers is next to either a ground or power plane. Layer
3 is best because it is not only next to a ground plane but is also guarded
by a power plane below it. This scheme is best for difficult-to-route DSP
systems that don’t operate at very high frequencies. One thing to keep in
mind is that board capacitance becomes important for systems operating
above 300MHz [2].
10.4. SUMMARY
PCB routing and board stackup are major contributors to EMI so designers
need to apply best practices to reduce radiated emissions. An example is
the use of an Image Plane, a ground plane located next to the routing layer
that provides low inductance current return paths for high-speed signals.
An Image plane helps reduce the current loop areas and minimizes the po-
tential differences on the ground plane. Experiments conducted in [3]
compare the EMI for PCBs with and without an image plane. They dem-
onstrate that a PCB with an image plane shows around 15dB reduction in
EMI across the frequency spectrum.
194 Chapter 10
REFERENCES
[1] Montrose Mark (2000) Printed Circuit Board Design Techniques for
EMC Compliance. The Institute of Electrical and Electronics Engi-
neers, New York.
T.T. Tran, High-Speed DSP and Analog System Design, DOI 10.1007/978-1-4419-6309-3_11, 195
© Springer Science+Business Media, LLC 2010
196 Chapter 11
55
Field Strength (dBuV/m)
50
Class A
48
45
43
Class B
40
37
35
34 960
216
30
88
25
The following lists some of the most common sources of EMI in high
speed DSPs:
• Transmission lines
High-Speed Current, I
DSP
Current, I
SDRAM
the board or chassis. In theory, 100% of the source current returns back to
the source but a small portion of the current spreads over the entire plane
before finding its way back to the source. This current creates an imbal-
ance in the ground potential and causes common mode radiation as shown
in Figure 11.3 [4].
Cable
System
ground
plane Vgnd
dB C om m on M ode
D if f e r e n tia l M o d e
Emission
e
ad
ec
/D
dB 20
20 dB
/D
ec
ad
e
fo fre q u e n c y f
1
f = , w h e r e T r is r is e tim e
π Tr
Tr Pw
T
1
20dB/decade
π Pw 1
plitude
π T r
Am
40dB/decade
In Figure 11.5 assuming a 50% duty cycle signal where only odd harmon-
ics are present, the amplitudes of the harmonics decay slowly as frequency
increases. The first pole frequency is at
1
f −3dB = , (11.3)
πPW
and a second pole is at
1
f − 3dB = ⋅ (11.4)
πTr
Pw and Tr are the width and the risetime of the signal respectively. There-
fore, increasing the risetime increases attenuation of the harmonics which
leads to lower radiation. This method is not always practical because the
slower risetime reduces the timing margin and may violate electrical re-
quirements such as setup and hold times.
The best technique to minimize EMI generated by digital signals is keep-
ing the high-speed signal traces as short as possible. It is a good practice
for engineers to go through a design and analyze the traces to see if they
are effective antennas or not. A good rule-of-thumb is keeping the length
of the trace less than the wavelength (λ) divided by 20. Here is the equa-
tion.
λ c
max_ trace _ length = = , (11.5)
20 20 f
A B
Load
Signal Low speed current
return (purple)
As shown in Figure 11.7 [6], current return creates a loop area that is di-
rectly related to the radiated electric field, so reducing the loop area lowers
radiation. Skin effect modifies the current distribution within a conductor
and increases resistance, so the high speed current return is right under-
neath the signal. Skin effect is negligible at lower frequencies but increases
as frequency rises. For a typical conductor used in DSP systems, a 10MHz
or higher trace is considered to be a high-speed signal. Providing a con-
tinuous ground plane right underneath a high-speed signal is the most ef-
fective way to achieve the lowest current loop area.
202 Chapter 11
If the ground plane is not continuous underneath the high speed signal, all
crosstalk, reflections, and EMI will increase due to the impedance mis-
match and larger current loop return area as shown in Figure 11.8 [6].
U19A U19A
14
14
1 1
+ 3 + 100nF 3
2 2
- -
7
7
14081/SO 14081/SO
I I
ΔV 3 . 3V
I ( peak ) = = = 48 . 5 mA
Zo 68
Since the package inductance is 1.44nH for 1nS risetime signal, the inter-
nal voltage droop is
dI 48 . 5 mA
V ( droop ) = L = (1 . 44 nH ) = 70 mV
dt 1 nS
Typically, one DSP power supply pin is shared by many output buffers.
This creates larger droop and leads to higher radiation. This helps explain
why good power supply decoupling is required for low EMI design.
Table 1 [5] shows the source current for different values of the series ter-
mination resistor. Changing the value from 10 ohms to 39 ohms does not
have much effect on the waveform, [5] showing about 1nS degradation,
but dramatically reduces the source current which greatly lowers the radi-
ate emissions. Figure 11.11 shows a DSP board with a 47 ohm series re-
sistor added to the memory clock, reducing the radiated emissions 3dB
compared to the emissions of the signal without termination.
Overall, if slower risetime signals are acceptable and do not violate AC
timing specifications, designers should use the largest resistor value to
terminate high speed signals to optimize the design from an EMI stand-
point.
10 ohms ~ 40mA
22 ohms ~ 10mA
25 ohms ~ 5mA
30 ohms ~ 10mA
33 ohms ~ 9mA
39 ohms ~ 8mA
206 Chapter 11
• Add an EMI pi filter on all the signals exiting the box or enter-
ing the box.
• If the system fails EMI tests, find the source by tracing the
failed frequencies to their source. For example, assume the de-
sign fails at 300MHz but there is nothing on the board running
at that frequency. The source is likely a 3rd harmonic of a
100HMz signal.
REFERENCES
[2] Montrose Mark (2000) Printed Circuit Board Design Techniques for
EMC Compliance. The Institute of Electrical and Electronics Engi-
neers, New York.
[6] Renolds J (2003) DDR PCB Routing Tutorial. Texas Instruments Inc
Tutorial.
Glossary
AC Alternating current
ADC Analog-to-Digital Converter
APLL Analog phase-locked loop
BER Bit Error Rate
BGA Ball Grid Array
CODEC COmpression/DECompression
CP Charge pump
CPU Central Processing Unit
DAC Digital-to-Analog Converter
DC Direct Current
DCO Digital controlled oscillator
DDR Dual Data Rate
DLL Delay-Locked Loop
DNL Differential Non-Linearity
DP Display Port
DPLL Digital phase-locked loop
DSP Digital Signal Processing
DVI Digital Video Interface
ENOB Effective number of bits
EMC Electromagnetic Compatibility
EMI Electromagnetic Interference
EMIFF External memory interface fast
EMIFS External memory interface slow
ESL Equivalent Series Inductance
ESR Equivalent Series Resistance
FCC Federal Communication Commission
GP Gain peaking
GPS Global Positioning System
211
212 Glossary
213
214 Index