Chip design for a time encoding A/D converter Dries Vercaemer · 2013. 9. 19. · H (s) ADC DAC U V...
Transcript of Chip design for a time encoding A/D converter Dries Vercaemer · 2013. 9. 19. · H (s) ADC DAC U V...
Dries Vercaemer
Chip design for a time encoding A/D converter
Academiejaar 2012-2013Faculteit Ingenieurswetenschappen en ArchitectuurVoorzitter: prof. dr. ir. Jan Van CampenhoutVakgroep Elektronica en Informatiesystemen
Master in de ingenieurswetenschappen: elektrotechniekMasterproef ingediend tot het behalen van de academische graad van
Begeleider: Amir Babaie FishaniPromotor: prof. dr. ir. Pieter Rombouts
Dries Vercaemer
Chip design for a time encoding A/D converter
Academiejaar 2012-2013Faculteit Ingenieurswetenschappen en ArchitectuurVoorzitter: prof. dr. ir. Jan Van CampenhoutVakgroep Elektronica en Informatiesystemen
Master in de ingenieurswetenschappen: elektrotechniekMasterproef ingediend tot het behalen van de academische graad van
Begeleider: Amir Babaie FishaniPromotor: prof. dr. ir. Pieter Rombouts
Dankwoord
Na vijf jaar studeren schrijf ik nu mijn masterproef als kroon op het werk. Het zou niet
tot stand gekomen zijn zonder de hulp van een aantal personen. Vooreerst wil ik mijn
promotor, Prof. Pieter Rombouts bedanken voor de tijd die hij genomen heeft om mij
hierbij te begeleiden. Met aanstekelijk enthousiasme verschafte hij uitleg en hielp hij deze
thesis tot een goed einde te brengen. Verder kan Amir Babaie Fishani ook niet ontbreken.
In geval van problemen was zijn deur snel gevonden. Ik bedank ook Maarten De Bock
voor het oplossen van een aantal praktische zaken. En ten slotte ook Johan Raman die
een belangrijke intellectuele bijdrage heeft geleverd.
Toelating tot bruikleen
De auteur geeft de toelating deze masterproef voor consultatie beschikbaar te stellen en de-
len van de masterproef te kopieren voor persoonlijk gebruik. Elk ander gebruik valt onder
de beperkingen van het auteursrecht, in het bijzonder met betrekking tot de verplichting
de bron uitdrukkelijk te vermelden bij het aanhalen van resultaten uit deze masterproef.
The author gives permission to make this master dissertation available for consultation
and to copy parts of this master dissertation for personal use. In the case of any other
use, the limitations of the copyright have to be respected, in particular with regard to the
obligation to state expressly the source when quoting results from this master dissertation.
Dries Vercaemer, juni 2013
Chip design for a time encoding A/D converter
Dries Vercaemer
Promotor:prof. dr. ir. Pieter Rombouts
Begeleider: Amir Babaie Fishani
Masterproef ingediend tot het behalen van de academische graad van
Master in de ingenieurswetenschappen: elektrotechniek
Vakgroep Elektronica en Informatiesystemen
Voorzitter: prof. dr. ir. Jan Van Campenhout
Faculteit Ingenieurswetenschappen en Architectuur
Academiejaar 2012-2013
Trefwoorden: Self-oscillating, Time encoding, Analog digital converter, Sigma delta
modulator
Extended Abstract
Inleiding
In het verloop van deze thesis is een zelf-oscillerende Σ∆ modulator ontworpen, zowel
op systeemniveau als circuit niveau. Het doel is een laag vermogen ADC met 20MHz
bandbreedte en een SNDR hoger dan 70dB te ontwerpen. Het circuit is geımplementeerd
in een 60nm technologie en de bemonsteringsfrequentie is 2GHz. Het vermogensverbruik
van de op-amps in de lus filter is ongeveer 1,2mW. Het voordeel van de zelf-oscillatie
is dat de lus slechts signalen aan moet kunnen met een frequentie maximaal gelijk aan
de oscillatie-frequentie. Aangezien deze een derde is van de bemonsteringsfrequentie kan
de bandbreedte van de op-amps in de lus gereduceerd worden wat vermogen bespaart.
Verder is een nieuwe aanpak toegepast om de slewing-vereisten verder te versoepelen.
Een laagdoorlaatfilter is in de lus geplaatst om de oscillatie te onderdrukken. Om de
invloed van dit filter op het lusfilter te elimineren is er een compensatiefilter toegevoegd in
parallel met de eerste integrator. Op circuit niveau is het lus filter geımplementeerd met
feedforward op-amps wegens hun laag vermogengebruik. Deze thesis bouwt onder andere
voort op de ontwerpen beschreven in [5] en [6].
PWM
Het systeem gaat uit van een Σ∆ modulator, afgebeeld in Figuur 1. Het idee is om puls-
breedtemodulatie(PWM) te gebruiken voor de ADC aan de uitgang van de modulator.
Deze modulatie techniek stelt de waarde van zijn ingangssignaal voor door de lengte van
een puls. Figuur 2 toont dit concept. De ingang is gesuperponeerd met een oscillatie en
dit signaal wordt aan een comparator aangelegd. Het resultaat is een blokgolf waarvan
de breedte een maat is voor het ingangssignaal. PWM heeft het voordeel dat een 1
bits DAC kan gebruikt worden in het terugkoppelpad, hetgeen betekent dat deze perfect
lineair is. Daarnaast speelt PWM perfect in op de eigenschappen van moderne halfgelei-
der technologieen namelijk een lage voedingsspanning en de capaciteit om nauwkeurig
tijdsintervallen te meten dankzij de hoge mogelijke snelheden. Belangrijker nog is het feit
dat het pulsbreedtegemodulleerd signaal slechts twee keer verandert per oscillatieperiode.
Dit betekent dat de lus de bemonsteringsfrequentie niet hoeft aan te kunnen. Het hoeft
ten snelste te werken aan de oscillatiefrequentie, fosc. Bovendien kan de bemonsteringsfre-
Σ H(s) ADC
DAC
U V
Figuur 1: Diagram van een algemeen sigma delta modulator
Figuur 2: Werking van PWM
quentie verhoogd worden zonder de lus aan te hoeven passen, zolang de oscillatiefrequentie
maar gelijk blijft.
Hoewel PWM slechts 1 bit quantisatie gebruikt is het uitgangssignaal toch een ver-
liesloze voorstelling van de ingang. Omdat de informatie is opgeslagen in de breedte van
de puls is het de bemonstering die een fout introduceert, zelfs als de bemonsteringsfre-
quentie hoger is dan de nyquistfrequentie. Deze fout kan voorgesteld worden door witte
quantisatieruis met volgende variantie:
σ2 =2fosc3fs
(1)
De quantisatieruis neemt dus toe met de oscillatiefrequentie en neemt af als de bemon-
steringsfrequentie verhoogt.
Zelf-oscillatie
De oscillatie die nodig is voor PWM wordt gecreeerd door de lus zelf. De lus wordt
onstabiel gemaakt door een vertraging toe te voegen in het terugkoppelpad. Figuur 3 toont
het systeem. Indien de DAC een non return to zero puls heeft moet de oscillatiefrequentie
aan volgende vergelijking voldoen:
− ωdTs −ωTs
2+ α(H(s)) = π (2)
De lusfilter van het ontworpen systeem zal een eerste orde integrator benaderen voor hoge
frequenties. We kunnen vergelijking 2 als volgt oplossen in dat geval: fosc = fs4(d+ 1
2).
Vervolgens wordt de comparator gelineariseerd. Zijn ingang bestaat uit een snel os-
cillerende component, en een tragere component gerelateerd aan het ingangssignaal van
Σ H(s)
DAC
U Vfs
z−d1 bit
Figuur 3: Diagram van een zelf-oscillerende sigma delta modulator
het systeem, U. Volgens [2] kan men de comparator voorstellen als een versterker op voor-
waarde dat de trage component een kleinere amplitude heeft dan de snelle oscillerende
component. De resulterende versterker heeft een versterking Gosc = 4πA voor de snelle
component, en een versterking Gs = 2πA voor de tragere en kleinere component. Hierin is
A de amplitude van de oscillatie aan de ingang van de comparator. Indien we de blokgolf
aan de uitgang van de comparator benaderen door de eerste term uit zijn fourierreeks, dan
vinden we dat A gelijk is aan 4π |H(jωosc)DAC(jωosc)| ≈ 4
π |H(jωosc)|. Nadat we Vergelijk-
ing 2 oplossen naar fosc kunnen we dus de comparator lineariseren en een transferfunctie
opstellen voor de quantisatieruis. Als we dit dan paren met Vergelijking 1 kunnen we de
prestatie van de modulator schatten.
Systeemniveau
Figuur 4 toont het diagram van het systeemniveau van de ontworpen modulator. Er zijn
twee lussen zichtbaar. Een buitenlus die drie integratoren bevat, en een binnenlus die er
slechts een bevat. De oscillatie is merendeels beperkt tot de binnenste lus. Ze wordt hard
onderdrukt door de drie integratoren in de buitenlus. Hierdoor hoeft de tweede integrator
de oscillatie niet te verwerken. De lokale terugkoppeling zorgt voor een nul in de ruistrans-
ferfunctie met hoekfrequentie√gc1c2. Door g aan te passen kan deze nul dicht tegen de
bandbreedte van de modulator geplaatst worden waar de quantisatieruis het hoogste is.
Dit verhoogt de SNR van het systeem. Er is een inkoppelpad van de ingang naar de derde
integrator. Dit maakt de signaaltransferfunctie vlakker en verlaagt de uitgangszwaai van
de eerste integrator. De vertraging in de buitenlus verlaagt de benodigde insteltijd voor
de eerste twee integratoren. Het systeem is in Simulink gesimuleerd en tevens is er een
ruistransferfunctie afgeleid met het lineaire model. het resultaat is te zien op Figuur 5.
De SNR is 75dB voor een -1.9dBFS ingangssignaal. Het lineaire model voorspelt een SNR
van 76dB.
Hoewel het niet nodig is voor de goede werking van de modulator dat de buitenste lus
de oscillatie verwerkt heeft de eerste integrator dit groot en snel signaal wel aan zijn ingang.
De integratoren zijn met op-amps geımplementeerd en hun uitgang kan voorgesteld worden
als een spanningsgestuurde stroombron die een capaciteit oplaadt. De stroom die deze bron
kan leveren is echter beperkt en door de wet van de condensator, is dus de afgeleide van de
Σ
DAC
U Vfs
z−d
1 bit
c1s
c2s
c3sΣ
g
DAC
z−12
1 bit
Figuur 4: Diagram van de derde orde SOSDM
10−4
10−3
10−2
10−1
−160
−140
−120
−100
−80
−60
−40
−20
0
f*Ts
FFT(dB)
Figuur 5: Spectrum van de uitgang na simulatie van de modulator samen met de lineaire be-
nadering
uitgangsspanning beperkt. Wanneer de op-amp deze maximale stroom levert wordt gezegd
dat ze slewt. Dit slewen is een niet lineair effect dat distorsie introduceert en het moet
voorkomen worden om de prestatie van het systeem te behouden. Dit kan gedaan worden
door de maximale leverbare stroom van de op-amp te verhogen maar dit kost vermogen.
In deze thesis is een nieuwe aanpak geımplementeerd om de oscillatie te onderdrukken in
de buitenste lus om slewing te voorkomen.
Er wordt een laagdoorlaatfilter in het terugkoppelpad geplaatst met een afsnijfrequen-
tie die lager ligt dan fosc. Dit filter beınvloedt echter de lusfilter en de stabiliteit van de
lus wat het ontwerp bemoeilijkt. Dit kan voorkomen worden door een compensatiefilter,
Hc, toe te voegen in parallel met de laagdoorlaatfilter en de eerste integrator. Indien
de laagdoorlaatfilter volgende vorm heeft 11+sτ moet Hc = c1τ
1+sτ . Gedurende simulatie
is het duidelijk geworden dat indien de afsnijfrequentie te dicht bij de doorlaatband van
de modulator ligt, de uitgangszwaai van de eerste integrator verhoogt. Er zal dus een
afweging moeten gemaakt worden tussen de maximaal toegelaten uitgangszwaai en de on-
derdrukking die het laagdoorlaatfilter biedt. In dit ontwerp is het oscillatiesignaal met een
factor hoger dan 6 onderdrukt.
Circuitniveau
Het integreren en someren wordt geımplementeerd met op-amps. Figuur 7 toont een
integrator met als ingang de som van twee signalen waarvan een door een laagdoorlaatfilter
gaat. In het totale circuit van de modulator zijn er drie op-amps en zij zijn de meest
cruciale componenten. De eerste op-amp in het bijzonder omdat die het leeuwendeel van
het vermogensverbruik voor zijn rekening neemt. De op-amps zijn geımplementeerd in
een feed forward architectuur met drie trappen. Het kleinsignaal circuit is afgebeeld op
Figuur 8. Volgende transferfunctie is gewenst:
A(s) = − 1
sτ1
(1 + sτ2
sτ2
)2
(3)
Vanwege stabiliteit moet τ1 kleiner zijn dan τ2. Het versterkingsbandbreedteprod-
uct(GBWP) is gegeven door 12πτ1
en wordt bepaald door de derde feedforward trap. Deze
trap verbruikt het meeste vermogen van de op-amp. De drie trappen van de eerste op-amp
staan afgebeeld op Figuren 5.3, 5.5, en 5.6 van de thesis. De tweede en derde op-amp
zijn heel gelijkaardig aan de eerste.
Het hoogste niveau van het circuit is afgebeeld op Figuur 4.3 van de thesis. De com-
parator met bemonstering is gerealiseerd door een voorversterker gevolgd door een flip
flop. De vertraging van een halve klok periode is geımplementeerd met een flip flop die
gestuurd wordt door een inverse klok. Alleen de invertoren komen niet op het systeem-
diagram voor. Ze opereren als buffers tussen het analoog en het digitaal deel van het
circuit. Ze hebben een hoge ingangsimpedantie en worden belast door de lage impedantie
van het analoge circuit. Door de lage resistieve belasting van de invertoren verbruiken zij
heel wat vermogen. Een ruwe schatting voor hun verbruik is 0,5mW. De drie op-amps
daarentegen verbruiken zo’n 1,2mW samen waarvan de eerste op-amp 0,9mW voor zijn
rekening neemt. Het verbruik in de flip-flops en de voorversterker is verwaarloosbaar.
Σ
DAC
U Vfs
z−d
1 bit
c1s
c2s
c3sΣ
g
DAC
z−12
1 bit
11+sτ c1τ
1+sτ
Σ
Hc
x
y
Figuur 6: Diagram van de zelf-oscillerende sigma delta modulator met oscillatie onderdrukking
X2+ Y+
X2- Y-
+
--+
X1+
X1-
V+
V-
C
C
R1RLP
R2
CLP
RLP R1
R2
Figuur 7: Integratie van de som van twee signalen van welke een laagdoorlaatgefilterd is.
R1 R2 R3C1 C2 C3gm1Vin gm2V1 gm3V2
V1 V2 Vuit
gmFF2Vin gmFF3Vin
Figuur 8: Het kleinsignaalcircuit van een drie-traps feedforward op-amp.
Besluit
In het verloop van deze thesis is een laag vermogen ADC ontworpen met een bemonster-
ingsfrequentie van 2GHz. Er werd gestart vanuit een Σ ∆ modulator. Door PWM te
gebruiken en de lus te laten oscilleren aan 666,7MHz werd het vermogensverbruik gere-
duceerd. Deze werd verder teruggedrongen door een laagdoorlaatfilter toe te voegen. Een
compensatiefilter elimineerde de invloed van dit laagdoorlaatfilter op het lusfilter. Het
ontwerp is geımplementeerd op circuitniveau met feedforward op-amps in een 60nm tech-
nologie. De bandbreedte is 20MHz en de maximale SNR 74dB. De op-amps verbruiken
ongeveer 1,2mW
Extended Abstract
Introduction
In this thesis, a self-oscillating Σ∆ modulator is designed on system and circuit level.
The aim is to design a low power ADC with 20 MHz bandwidth and an SNDR higher
than 70dB. The sample frequency is 2GHz and the circuit is implemented in a 60nm
technology. The power consumption of the op-amps in the loop filter is only 1.2mW. The
advantage of the self-oscillation is that the highest frequency of the signals in the loop
is the oscillation frequency. As the loop oscillates at a third of the sample frequency,
this means the bandwidth of the op-amps in the loop is reduced, saving power. A novel
approach is applied to further reduce slewing requirements. A low pass filter is introduced
in the loop to suppress the oscilation, and a compensation filter is added, bypassing the
first integrator, to eliminate the influence of the low pass filter on the loop filter. On
circuit level the loop filter is implemented with feed forward op-amps as they have a low
power consumption. This thesis started from the designs described in [5] and [6].
PWM
The system starts from a sigma delta modulator, depicted on Figure 1. The idea is to
use pulse width modulation in the ADC at the output of the sigma delta modulator.
Pulse width modulation represents the value of its input voltage into a length of a pulse.
Figure 2 shows the operation. The input is summed with an oscillating signal and applied
at a comparator. The output is a square wave which width is proportional to the value of
the input. PWM has the advantage that a one bit DAC can be used in the feedback path,
which means it is perfectly linear. Also PWM makes perfect use of the characteristics of
modern semi-conductor technologies: low voltage headroom and accurate measuring of
Σ H(s) ADC
DAC
U V
Figure 1: Diagram of a general sigma delta modulator
Figure 2: Operation of PWM
time intervals due to high speeds. But more importantly, the PWM signal changes only
twice every oscillation period. This means the loop does not have to handle the sample
frequency but only the oscillation frequency, fosc. Furthermore, the sample frequency can
be increased without affecting the loop as long as the oscillation frequency remains the
same.
Although PWM uses one bit quantization, the output is a lossless representation of its
input. Because the information is stored in the width of a pulse, it is the sampling that
will introduce an error, even if the sample frequency is higher than the Nyquist frequency.
This error can be represented by white quantization noise with a variance of:
σ2 =2fosc3fs
(1)
This means the quantization noise decreases with increasing sample frequency and de-
creasing fosc.
Self-Oscillation
The oscillation needed to perform PWM is generated by the loop itself. A delay is added,
making the loop unstable. Figure 3 shows the system. If a non return to zero pulse is
assumed for the DAC, the oscillation frequency has to satisfy:
− ωdTs −ωTs
2+ α(H(s)) = π (2)
The loop filter for the designed system will approximate a first order integrator for high
frequencies. If that is the case, Equation 2 can be solved and the result is: fosc = fs4(d+ 1
2).
Σ H(s)
DAC
U Vfs
z−d1 bit
Figure 3: Diagram of the self-oscillating sigma delta modulator
Next the comparator is linearized. Its input consists of a oscillation component and a
slower component related to the system input U. From [2] it is learned that if this slow
component is smaller than the fast input component, the comparator can be linearized as
a gain Gosc = 4πA for the fast oscillating component, and a gain Gs = 2
πA for the slower
signal component. A is the amplitude of the oscillation at the input of the comparator. If
the square wave oscillating output of the comparator is approximated by the first term of
its Fourier series, A is found to be 4π |H(jωosc)DAC(jωosc)| ≈ 4
π |H(jωosc)|. After solving
Eq. 2 for the oscillation frequency, a transfer function can be derived for the quantization
noise by using the equivalent comparator gain Gs. After coupling this with Eq. 1 an
estimate can be made of the systems performance.
System level
Figure 4 shows the system level diagram of the designed modulator. There are two loops
visible. An outer loop containing three integrators and an inner loop with only one. The
oscillation is chiefly contained in the inner loop. It is very weak in the outer loop because
of the suppression with a third order integrator. This way, the second integrator does
not need to process the fast oscillation. The local feedback path introduces a zero in the
noise transfer function of the system at the angular frequency√gc1c2. By changing the
feedback gain g, this zero can be placed close to the bandwidth, where the output noise is
highest. This effectively increases the SNR. A feed forward path is also introduced from
the input to the third integrator. This flattens the signal transfer function and reduces
the output swing of the first integrator. The delay in the outer loop reduces the settle
time requirements of the first two integrators. The system is simulated in simulink and
the noise transfer function is calculated using the linear model of the system. Figure 5
shows the result. The SNR is 75dB for a -1.9dBFS input signal, while the linear model
predicts a SNR of 76dB. This is a good fit.
Although the outer loop does not need to process the oscillation for the good operation
of the modulator, the first op-amp still receives this large and fast signal at its input. The
Σ
DAC
U Vfs
z−d
1 bit
c1s
c2s
c3sΣ
g
DAC
z−12
1 bit
Figure 4: Diagram of a third order SOSDM
10−4
10−3
10−2
10−1
−160
−140
−120
−100
−80
−60
−40
−20
0
f*Ts
FFT(dB)
Figure 5: Spectrum of the output of the simulated system and the linear approximation
integrator is implemented with an op-amp. The output of this op-amp can be represented
as a voltage controlled current source loaded by a capacitor. The maximum current this
source can deliver is limited. As a result the maximum rate of change of the output voltage
is limited. When the op-amp delivers this maximum current, the op-amp is said to be
slewing. This is a non-linear effect that introduces distortion and it should be prevented to
safeguard the performance of the modulator. Slewing is avoided by increasing the current
the op-amp can deliver but this also increases the power consumption. In this thesis a
novel approach was tested to suppress the oscillation in the outer loop and hence reduce
the slewing specifications for the first op-amp.
A low pass filter is introduced with a cut off frequency lower than fosc. This filter
does however influence the loop filter and the stability of the loop. This makes the design
process more complex. However by introducing a compensation filter, Hc, that bypasses
the first integrator, the effect of the low pass filter on the system can be eliminated.
Figure 6 shows the system diagram. For a first order low pass filter 11+sτ , Hc is found
to be c1τ1+sτ . It has been observed that placing the cut off frequency too close to the pass
band of the modulator increases the output swing of the first integrator. This means a
trade off will have to be made between suppression and output swing. In this design the
oscillation amplitude was suppressed by a factor 6.
Circuit level
The integrating and summing is implemented using op-amps. Figure 7 shows an integrator
that acts on the sum of two signals, one of which is low pass filtered. There are three
op-amps in the modulator circuit and they are the most crucial components. Especially
the first op-amp as it determines most of the power consumption. All op-amps have been
implemented using a feed forward architecture with three stages. The small signal circuit
Σ
DAC
U Vfs
z−d
1 bit
c1s
c2s
c3sΣ
g
DAC
z−12
1 bit
11+sτ c1τ
1+sτ
Σ
Hc
x
y
Figure 6: Diagram of a third order SOSDM with oscillation suppression
is depicted in Figure 8. Following transfer function is wanted:
A(s) = − 1
sτ1
(1 + sτ2
sτ2
)2
(3)
For stability τ1 should be smaller than τ2. The GBWP equals 12πτ1
and is determined by
the third feed forward stage. This stage consumes most of the power of the first op-amp.
The three stages of the first op-amp are represented in Figures 5.3, 5.5, and 5.6 of the
thesis. The second and third op-amps are very similar to the first.
X2+ Y+
X2- Y-
+
--+
X1+
X1-
V+
V-
C
C
R1RLP
R2
CLP
RLP R1
R2
Figure 7: Integration of the sum of two signals, one of which is low pass filtered
R1 R2 R3C1 C2 C3gm1Vin gm2V1 gm3V2
V1 V2 Vuit
gmFF2Vin gmFF3Vin
Figure 8: Small signal circuit of a three stage feed forward op-amp
The top level circuit is depicted in Figure 4.3 of the thesis. The sampled comparator is
implemented with a preamp followed by a flip flop. The half clock cycle delay is realized
with a flip flop clocked at the inverse clock. Only one set of components has not been
discussed: the inverters. They act as buffers between the analog and digital parts of
the circuit. They offer a high input impedance and are loaded by the low impedance of
the analog part of the circuit. They also are responsible for a large part of the power
consumption because of their low resistive loading. A raw estimate of their consumption
is 0.5mW. An estimated consumption of the three op-amps together is 1.2mW, of which
0.9mW is consumed in the first op-amp. The consumption in the preamp and flip-flops is
negligible.
Conclusion
In this thesis a low power ADC was designed with a sample frequency of 2GHz. The
starting point was a Σ ∆ modulator. By using PWM and making the loop oscillate at
666,7MHz, the power consumption was reduced.It was further reduced by introducing a
low pass filter. A compensation filter eliminated the influence of the low pas filter on the
loop filter. The design was implemented on the circuit level using feed forward op-amps
in a 60nm technology. The bandwidth of the system was 20MHz and the maximum SNR
was 74dB. The op-amps consumed approximately 1.2mW.
Contents
Glossary
1 Introduction 1
2 Self-Oscillating Sigma Delta Modulator 3
2.1 The ADC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Sigma Delta Modulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 PWM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 Self-Oscillating Sigma Delta Modulator . . . . . . . . . . . . . . . . . . . . 8
3 System design 13
3.1 A third order SOSDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Results of the linear approximation . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.4 Slewing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.5 Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4 The circuit 23
4.1 Top level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2 Dimensioning the Passive components . . . . . . . . . . . . . . . . . . . . . 26
4.3 Preamp, Flip Flops and Buffers . . . . . . . . . . . . . . . . . . . . . . . . . 27
5 Op-Amp Design 31
5.1 Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1.1 GBWP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1.2 DC Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.1.3 Slewing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.1.4 Other Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.2 General Feed Forward Op-Amp Design Flow . . . . . . . . . . . . . . . . . 34
5.3 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.3.1 CMFB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.3.2 Outputswing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Contents
5.3.3 Cascode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3.4 Biassing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3.5 Other Stages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.3.6 Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3.7 Op-Amp1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.3.8 Op-Amp2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.3.9 Op-Amp3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.3.10 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6 Conclusion 53
Bibliography 55
Glossary
ADC analog to digital converter.
CMFB common mode feedback.
DAC digital to analog converter.
dbFS value relative to full scale in decibel.
FF op-amp feedforward op-amp.
GBWP gain bandwidth product.
IirT invariant impulse response transformation.
NRZ non return to zero (pulse).
NTF noise transfer function.
op-amp operational amplifier.
OSR oversampling ratio.
Preamp preamplifier.
PWM pulse width modulation.
SDM sigma delta modulator.
SNDR signal to noise and distortion ratio.
SNR signal to noise ratio.
SOSDM self-oscillating sigma delta modulator.
Glossary
Chapter 1
Introduction
Digital electronics have become an important part of our lives. Costs decreased and speed
rose due to Moore’s law, which enabled electronics to penetrate deeper into our world.
There are dozens of chips in modern factories and cars. Digital consumer products are
omnipresent. However, though it would be very convenient, the physical world stubbornly
refuses to go digital. High clock speeds and fast memories are useless without data. This
is where analog digital converters come in to play. They are the bridge between our digital
world and the analog reality. Manny micro controllers work in real time and if their digital
speed increases, the ADC’s must follow. On top of this, more and more systems become
wireless and have to rely on battery power. This creates a market for high bandwidth low
power ADC’s.
The aim of this thesis is to design a self-oscillating sigma delta modulator with low
power consumption and a bandwidth of 20MHz. It will have a sample frequency of 2GHz
and offer a SNR higher than 70dB.
The design will first take place on the system level. The operation of the SOSDM will
be explained, starting from the conventional sigma delta modulator. A linear model will
be proposed to predict the performance of the system. Afterwards the coefficients of the
system will be optimized for a high SNR. Still on the system level, a novel technique will
be used to further decrease the power consumption of the system.
After the system level design is complete, the design will shift to the circuit level. The
modulator will be implemented in a 65nm technology. Feed forward op-amps will be used
due to their power friendliness.
I used [5] and [6] as a starting point for this thesis.
1
Chapter 1. Introduction 2
Chapter 2
Self-Oscillating Sigma Delta
Modulator
2.1 The ADC
Signals in the analog world are continuous, both in time and value. To store them without
precision loss would require an infinite memory. The analog digital converter approximates
its analog input so that it can be stored in a digital memory with finite dimensions. The
ADC performs two operations. First it samples its input. Every sample period, it stores
the value of the input signal at that time. Nyquist proved that if the sample frequency
is higher than twice the maximum frequency of the input(the Nyquist frequency), this
sampling does not lead to precision loss. The second operation is quantization. Every
value of the sampled input is rounded to a quantization level. This operation does lead to
information loss. It is the quantization that limits the resolution of the ADC. The error
introduced by the quantization is often represented as white noise, called the quantization
noise.
There are several types of ADC’s. Flash converters are very fast but require a large
chip area. Successive approximation leads to high precision ADC’s with limited chip area
but it is rather slow. The design of this thesis is based on the sigma-delta modulator.
The SDM is able to offer high precision in reasonable time but it does require high sample
frequencies.
2.2 Sigma Delta Modulator
The basic schematic of a sigma delta modulator is depicted in figure 2.1. It is a feedback
loop. It uses a simple, low accuracy ADC and uses feedback to reduce its error. The
sampling frequency by which the loop operates is much higher than the Nyquist frequency.
This is called oversampling and is crucial to the operation of the sigma delta modulator.
3
Chapter 2. Self-Oscillating Sigma Delta Modulator 4
The oversampling ratio or OSR is given by fs2BB . The effect of the loop filter is that it
pushes the quantization noise, produced by the simple ADC, away from the baseband to
higher frequencies. If the output of the loop is low pass filtered, a lot of the quantization
noise will be filtered out, leading to much higher accuracy than the simple ADC could
deliver.
Σ H(s) ADC
DAC
U V
Figure 2.1: A sigma-delta modulator with a loop filter in continuous time
The accuracy of most ADC’s is limited to the accuracy of the components. Components
with high precision values are difficult to produce so performance is often limited. Some
integrating ADC can obtain high accuracy with low precision components but they achieve
this at the cost of low conversion rates. This shows the strength of the SDM. It can obtain
high precision in a limited time with a low influence of the accuracy of the components
on the performance. The trade off is the oversampling which necessitates a fast circuit.
The operation is illustrated for a modulator with a first order loop filter and a one
bit ADC. The ADC consists of a sampler and a one bit quantizer which is basically a
comparator. The comparator projects its input to +1 or -1 and hence produces an error.
It is well established [4] that this error, under certain conditions, closely resembles white
noise. So the ADC can be approximated as an amplifier followed by a sampler with a
white noise source at its output. Figure 2.2 shows this system.
Σa
sTs
UG Σ
E
DAC
Vfs
Figure 2.2: SDM with first order loop filter and one bit DAC.
If the sampling is neglected, the transfer function from the noise to the output issTs
sTs+aG . The bode plot is shown on figure 2.3. It is clear that the noise transfer function,
NTF, pushes the noise away from baseband. The loop filter ’shapes’ the noise. If the OSR
Chapter 2. Self-Oscillating Sigma Delta Modulator 5
is increased by decreasing the bandwidth or increasing the sample frequency (and the inte-
grator constant) less quantization noise would fall in the bandwidth and our performance
would increase. The signal transfer function is:
aGsTs
1 + aGsTs
(2.1)
which, for low frequencies approximates as 1.
10−2
10−1
100
101
102
−50
−45
−40
−35
−30
−25
−20
−15
−10
−5
0
Mag
nit
ud
e (
dB
)
Bode Diagram
Frequency (rad/sec)
Figure 2.3: NTF of first order SDM
Another way of explaining the SDM operation and the noise shaping is in the time
domain. Assume a zero input and a small charge on the integrator. The output of the
comparator would be one volt. This leads to a an negative input of the integrator and
at the next clock cycle the output of the integrator will have decreased. After a few
clock cycles it will be negative. The comparator switches to -1 and every few clock cycles
the output switches between +1 and -1. We assumed the output is low pass filtered or
averaged. In this case the average output will be zero which is a better representation of
the input than the simple 1 bit ADC could offer. This also works for a non zero input. If
the input is positive the output will be +1 for more clock cycles because the error signal
will have a positive offset. The averaged output will get a decent approximation of the
input. If the sample frequency is increased, there will be more values to average over and
the result will be more accurate.
This simple system can now be used to illustrate the problem of overloading, which
will limit the input range of the design. Assume an input U, higher than the output of the
comparator. The input of the integrator, the error signal, is U-1 which is always positive.
This means the integrated signal will increase without end. This desensitizes the system to
its input and will lead to non-linearities as the limit of the systems components is reached.
Chapter 2. Self-Oscillating Sigma Delta Modulator 6
It is called overloading the quantizer. For big signals the white noise approximation of
the quantization is no longer valid and the circuit becomes unstable. This instability is
sometimes hard to predict for higher order systems, which can become unstable even at
zero input.
This SDM has a continuous time loop filter, H(s). A discrete loop filter is also possible
by sampling before the loop filter. The advantage of a continuous loop filter is that the op-
amps can have a lower settling time, because no stable value has to be reached before the
next clock cycle as opposed to a switched C realization. Another advantage specifically for
the Self-Oscillating SDM is the scalability of the sampling frequency as will be illustrated
later.
2.3 PWM
To improve the accuracy of the ADC, The OSR can be increased. However, if a high
bandwidth is required, this leads to a high sampling frequency. Although the scaling of
the semi conductor technology allows for high frequencies, high enough OSR’s can often
not be achieved due to power limitations. Another way to increase the accuracy is to
use a multi bit quantizer. However this has some drawbacks. Firstly, a multi bit ADC
and DAC contain some non-linearities. The non-linearities of the ADC are not a limiting
factor since they are suppressed, like the quantization noise, by the loop filter. The non
linearities of the DAC in the feedback path are a bigger problem. They directly influence
the performance of the SDM. This problem does not pose itself for a one bit DAC. A one
bit DAC is always linear since two points define a straight line.
Another problem of the multi bit quantizer is the low supply voltage of new high speed
technologies. The newest semiconductor technologies barely achieve a supply of one volt.
A classical ADC, like the flash ADC, produces its bit by comparing its input with multiple
voltage levels. Because of the falling supply voltage these levels become very close to each
other which decreases precision.
A solution to these two problems is to use a PWM-quantizer. Instead of comparing
the input voltage to different voltage levels, a rapid oscillating signal is added to the input
and the result is compared to zero. This is illustrated in Figure 2.4.
PWM produces a square wave output signal. The width of each pulse is a measure
of the input signal. The quantization happens in time, not in the voltage of the signal.
Also the output is a one bit signal so the feedback DAC is inherently linear. Pulse width
modulation is adapted to new, fast speed technologies. It is able to work with low supply
voltages and makes use of the abilities to measure time with high accuracy. However the
main reason to use PWM is that it leads to lower power consumption. The square wave
output signal only changes twice every oscillation period. This means the fastest signal
the loop of the SDM has to handle changes at this oscillation frequency, not the sample
Chapter 2. Self-Oscillating Sigma Delta Modulator 7
Figure 2.4: Pulsewidth modulation
frequency. This leads to bandwidth reduction for the op-amps in the loop, and hence
power saving.
The signal at the output of the PWM-Quantizer is an exact representation of the
input signal. Unlike conventional quantizers, the error is not introduced by the quantizer.
Because the signal level is represented by the length of a pulse, the quantization error is
caused by sampling. If sampling doesn’t happen at the moments that the output square
wave makes a transition, an error is made in determining the pulse length and thus an
error in determining the signal level. The loop filter shapes the error signal at its input.
+2
-2
-1
-1
+1
+1
Figure 2.5: Output of the comparator before and after sampling together with error signals
This signal is the difference between the continuous input and the zero order hold version
of the output. Therefore a good representation of the error made by quantizing is the
Chapter 2. Self-Oscillating Sigma Delta Modulator 8
difference between the comparator output before sampling and the shifted zero order hold
version of the sampled comparator output. We have to shift the second signal because the
zero order hold pulse introduces a half clock cycle delay. This error pulse, E1, will have
a maximum length of Ts and a height of +/-2 and appears twice every oscillation period.
A sampled error signal, E2, is now constructed by demanding that its zero order hold
version has the same area as the error signal E1. E2 ranges from −8fosc/fs to 8fosc/fs
and represents the quantization noise. The variance of a stochastic variable, uniformly
distributed over [−∆,∆], is ∆12 . If the quantization noise is assumed to be uniformly
distributed, its variance is:
∆/12 =2
3
foscfs
(2.2)
Figure 2.5 shows the output signal before and after sampling together with the error
signals.
2.4 Self-Oscillating Sigma Delta Modulator
To implement a PWM ADC, an oscillation is needed. An easy way to implement this
oscillating signal is by adding some delay in the loop. Figure 2.6 shows the SOSDM. By
Σ H(s)
DAC
U Vfs
z−d1 bit
Figure 2.6: Diagram of SOSD-modulator
adding this delay, the loop becomes unstable thus an oscillation is created. The output
comparator changes at the oscillation frequency. This means the fastest frequency the loop
has to handle is the oscillation frequency not the sample frequency. This reduces the power
consumption, as the bandwidth of the components in the loop can be reduced. Earlier it
was mentioned that an advantage of a continuous loop filter is scalability. Equation 2.2
shows that the noise decreases with increasing sample frequency. However the loop filter
does not operate at the sample frequency. The sample frequency can be increased without
affecting the loop filter, as long as the delay is also increased to offer the same oscillation.
This eases the way to newer technologies. Old designs can easily be translated to newer
technologies, allowing the sample frequency to be increased. This means a better noise
performance without the need to completely redesign the loop filter.
The loop will start to oscillate at the frequency at which the total phase shift of the
Chapter 2. Self-Oscillating Sigma Delta Modulator 9
loop becomes 180. So the phase shift of the loop and the delay determine the frequency
of oscillation. As illustration the frequency of oscillation for a first order loop filter is
calculated.
If the DAC has a non return to zero pulse it’s transfer function is 1−esTssTs
. If the effect
of sampling is ignored, the total phase shift of the loop is:
− 90 − ωdTs − ωTs/2. (2.3)
This equals 180 for ωTS = π2(d+1/2) so
fosc =1
4(d+ 1/2)fs (2.4)
For example a delay of one gives an fosc of fs6 . This simple result is still relevant for the
design of the higher order system. Although this will be a third order system, the loop
filter will predominantly behave as a first order filter for high frequencies so Equation 2.4
will hold approximately.
In the previous analyses the effect of sampling was ignored. This assumption has little
influence on the result, as will be illustrated later, apart from the fact that the period
of oscillation has to be an integer multiple of the sample period. For higher order loop
filters, the frequency at which the phase shift of the loop becomes 180 is not necessarily
an integer division of the sample frequency. So how is the exact oscillation frequency to
be determined?
As Figure 2.7 shows, the sampler will add extra delay and thus an extra amount of
phase, ∆φ. This means that the oscillation frequency will be the frequency at which the
phase of the loop becomes 180 −∆φ. However ∆φ is unknown. Only an interval can be
determined. The delay added by the sampler, ∆T , ranges from 0 to Ts. This means ∆φ
lies in [0, ωoscTs]. If ωosc = 2 ∗ pi ∗ fs/6 for example, than ∆φ = π/3 which is quite a large
phase shift. Often multiple oscillation frequencies will be possible. If this is the case, it
is best to choose the highest possible oscillation frequency. This choice is supported by
simulation. Multiple possible oscillations will be closer examined later in this thesis.
To optimize the loop filter and the delay, it is necessary to be able to predict the SNR
of the system. An estimation of the SNR can be obtained through linearisation of the
system. There are two non-linear elements in the modulator: the comparator and the
sampling. The latter will be ignored for now. The input of the comparator consists of
ΔT Ts
Figure 2.7: Delay caused by sampling
Chapter 2. Self-Oscillating Sigma Delta Modulator 10
an oscillation component and a much slower signal related to the input of the system U.
Luckily a comparator with two sinusoids as input has been linearized in literature[2]. The
comparator can be considered as an amplifier with two different gains provided one input
has a much lower frequency than the other. The gain of the fast input can be approximated
as Gc = 4πA , with A the Amplitude of this signal, and the slow signal gain is:
Gs =2
πA(2.5)
Now the amplitude of the oscillating signal can easily be calculated. The output of the
comparator is a square wave with an amplitude of one which can be approximated by the
first term of its Fourier series 4πsin(ωosct). This means the amplitude of the oscillation at
the input of the comparator is 4π |H(ωoscj)| and Gc = 1
|H(ωoscj)| . If z is approximated as
esTs , the NTF of the system becomes:
NTF =1
1 +GsH(s)DAC(s)e−sdTs(2.6)
By combining this with the Expression for the variance of the quantization noise (Eq 2.2),
the SNR can be predicted. Figure 2.8 shows the linearized system.
Note that it was assumed that the oscillation was a sine while in reality it is more like a
triangle wave. The harmonics of the oscillation were neglected. Of course this linearizion
is an estimation of the behavior of the system.
Up until now the effect of sampling was neglected. To include this effect, a transfer
function is needed in the Z domain. Such a loop gain exists as can be seen by cutting
the loop before the DAC, where the signal is digital, and calculating the loop gain around
this cut. To calculate the transfer function of a mixed discrete continuous system, the
Invariant impulse response Transform[1] can be used. First the loop gain in the S domain
is calculated by neglecting the sampling. Afterwards this transfer function is transformed
to the Z domain: LGeq(z) = IirTLG(s).This transformation takes following steps. First the impulse response of the S domain
loop gain is calculated using the inverse Laplace transform. Then this impulse response is
Σ H(s)U V
G
e−dsTs1−esTs
sTs
Σ
E
Figure 2.8: Diagram of linearized first order SOSD-modulator
Chapter 2. Self-Oscillating Sigma Delta Modulator 11
sampled. Eventually the z transform of this discreet impulse response is produced. This
will be illustrated for a first order loop filter: H(s) = cs and a half clock cycle delay: d = 1
2 .
Assume the DAC has a NRZ pulse: DAC(s) = 1−e−sTs
s . It can easily be seen that full
clock cycle delays are transformed by the IirT into z−1. They will be neglected for now
and added later. The loop gain in the S domain equals
LG′(s) = Gsc
s2e−
sTs2 (2.7)
Where the extra s in the denominator is all that remains of the DAC for now. Next, the
inverse Laplace transform is calculated and the result sampled.
h(t) = L−1LG(s) (2.8)
= Gsc(t−Ts2
)u(t− Ts2
) (2.9)
h(n) = GscTs(n− 1/2)u(n− 1) (2.10)
Note that a half clock cycle delay in a unit step, u(t− Ts2 ), gives rise to a full clock cycle
delay in the sampled step. Finally the Z transform is calculated and the delays are added:
LGeq(z) = Zh(n)(1− z−1) (2.11)
= GscTs
(z−2
(1− z−1)2+
1
2
z−1
1− z−1
)(1− z−1) (2.12)
Figure 2.9 shows the magnitude and the phase for the S and Z domain loop gain. As you
can see, the error made by neglecting the sampling is only relevant for frequencies close
to the sample frequency.
In Equation 2.5 it was assumed that the output of the comparator is plus or minus
one. If this output level is called D, Equation 2.5 changes to Gs = 2DπA . As A, the
oscillation amplitude, is proportional to D, the operation of the system does not depend
10−5
10−4
10−3
10−2
10−1
100
−350
−300
−250
−200
−150
−100
−50
0
50
100
f*Ts
Mag
nit
ud
e in
dB
10−5
10−4
10−3
10−2
10−1
100
−280
−260
−240
−220
−200
−180
−160
−140
−120
−100
−80
f*Ts
Ph
ase in
deg
rees
Figure 2.9: Plot of magnitude and phase of LG in Z domain(blue) and in the S domain(red)
Chapter 2. Self-Oscillating Sigma Delta Modulator 12
on D. However simulation shows, that halving D doubles the SNR. The more exact formula
given by [2] is:
Gs =2D
πBsin−1
(B
A
)(2.13)
This equation holds for a comparator with two inputs with amplitudes A and B, with A
larger than B, and if signal B is much slower than signal A. In equation 2.5 sin−1(BA
)was
approximated as BA , but in the actual system, A is not necessarily much larger than B. In
this case, Gs and thus the noise suppression does depend on D. A smaller D also causes
the system to become unstable at smaller inputs since overloading happens if the input is
larger than D. Considering this, D is chosen as large as allowed, which is about half the
supply voltage of the 65nm technology that will be used or about 0.5V.
Chapter 3
System design
3.1 A third order SOSDM
Figure 3.1 shows the system diagram of a third order modulator. As you can see the loop
Σ
DAC
U Vfs
z−d
1 bit
c1s
c2s
c3sΣ
g
DAC
z−12
1 bit
Figure 3.1: Diagram of the third order SOSD-modulator
filter has been realized using two feedback paths. This forms two loops: The inner loop
containing the integrator c3/s, and the outer loop containing all three integrators. Because
of the triple integration, the oscillation will be very weak in the outer loop provided the in-
tegration coefficients are smaller than the oscillation frequency. This means the first order
inner loop chiefly determines the oscillation frequency and amplitude and Equation 2.4
still holds approximately. The reason for this lay-out is the power consumption of the
circuit. The integrators are realized using op-amps. By containing the oscillation to the
inner loop, the bandwidth of the first two op-amps can be reduced which leads to lower
power consumption.
There is also a half clock cycle delay in the outer loop. This lowers the settling time
requirements of the two first op-amps without affecting the loop filter and oscillation to a
great extent.
To linearize this system the comparator is replaced by a gain G and the sampling is
13
Chapter 3. System design 14
neglected. z is again approximated by esTs and a non return to zero pulse is assumed for
the DAC’s. This leads to the following loop gain:
LG =GC3
s
c1(c2 + s)e−sTs/2 + (s2 + c1c2g)e−dsTs
s2 + c1c2g
1− e−sTss
(3.1)
This illustrates that the local feedback with gain g causes a pole at ωz =√c1c2g. Because
of the noise shaping (see Figure 2.3), the quantization noise is highest at high frequencies.
By placing the pole in the LG close to the bandwidth, the total amount of quantization
noise in the signal band is strongly reduced. The trade-off is a lower noise suppression at
low frequencies. Because of the pole, the low frequent noise is only suppressed with one
integrator, which is still enough.
The feed forward from the input U to the second adder leads to a STF closer to one
and hence reduces the output swing of the first integrator. Figure 5.32 shows this output
with and without the feed forward path. It is clear to see that the first op-amp has to be
able to produce a very rapidly changing output. This means there could be slewing which
will be discussed later.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x 10−6
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15Output OA1
t0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x 10−6
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15Output OA1
t
Figure 3.2: Output of first integrator with(left) and without(right) FF
To determine the error made by neglecting the sampling, the invariant impulse response
transformation is used to determine the loop gain in the Z domain. The magnitude and
phase of the loop gain in the S and Z domain are depicted in Figure 3.3. As you can see,
the mistake made by neglecting the sampling is very small and only significant for high
frequencies. However it is best to use the Z domain expression to determine the oscillation
frequency since a small error in the phase can lead to a different frequency.
Now that the system is linearized, it is possible to predict a SNR which will be verified
using a simulink simulation. However one problem still needs to be solved. As can be seen
on Figure 3.3, there are three different frequencies at which the loop gain becomes 180.
Figure 3.4 depicts a part of the Nyquist plot. Because the magnitude of the loop gain
varies if the oscillation frequency changes, the Nyquist curve can be stretched, depending
Chapter 3. System design 15
10−5
10−4
10−3
10−2
10−1
100
−100
−50
0
50
100
150
200
f*Ts
Mag
nit
ud
e in
dB
10−5
10−4
10−3
10−2
10−1
100
−400
−350
−300
−250
−200
−150
−100
−50
f*Ts
Ph
ase in
deg
rees
Figure 3.3: Bode plots of the loop gain in the Z domain (blue) and S domain(red)
Figure 3.4: Unscaled sketch of a part of the Nyquist plot
on the gain of the comparator.There are three possible oscillation frequencies. If the
system oscillates at one of these frequencies, that crossing point will shift to the -1 point,
depicted by the red dot, to make the loop gain equal to minus one.
Not all three points represent stable oscillations. Assume that the gain of the com-
parator is such that the -1 point is at the left of point 2, as it is depicted now. In that case
the minus one point lies at the right of the curve, which means there is a pole in the right
half of the complex plane. This is an unstable system that oscillates with rising amplitude.
As the amplitude rises, the gain of the comparator decreases together with the magnitude
of the loop gain, so the Nyquist plot shrinks and point two moves away from the minus
one point. Now assume that point two lies at the left of the red dot. Now the system is
stable because minus one is at the left of the Nyquist curve. The oscillation-amplitude
decreases and the comparator gain increases which makes the Nyquist plot expand. Again
point two moves away from the red dot. Using a similar analysis, it can be concluded
that the other two points move towards minus one if they are disturbed. So point one
and three represent stable oscillations. It is clear however that point three oscillates at a
smaller amplitude than point one as it needs more gain. So if the system starts from a
Chapter 3. System design 16
zero condition, an oscillation will start up with a rising amplitude until the amplitude is
high enough to support oscillation three. As this is a stable oscillation, the amplitude will
not rise further and point one will never be reached.
3.2 Results of the linear approximation
Figure 3.5 shows the simulated spectrum together with the predicted NTF in black. the
red line depicts the NTF of a corrected linear model where the fosc and the oscillation
amplitude are determined after simulation. both lines lie close to the simulated spectrum.
The SNR of the linear model is about 1dB higher than the actual SNR and after correction
with the actual fosc and amplitude the SNR is 0.4dB lower. The model predicts a Gs of
9 and a fosc of fs/5 while the simulation shows a Gs of 13 and a fosc of fs/3. The low Gs
is a result of the wrong prediction of fosc. The simulated system without an input shows
a fosc of fs/4 which is closer to the result of the linear model. fosc is clearly influenced by
the input signal which was not predicted. But although the model has it shortcomings, it
still predicts the SNR fairly well.
10−4
10−3
10−2
10−1
−160
−140
−120
−100
−80
−60
−40
−20
0
f*Ts
FFT(dB)
Figure 3.5: Simulated spectrum together with linear model(red) and adjusted linear model(black)
If the comparator is overloaded with a high input signal the system starts to oscillate
at a much lower frequency ( fs128 =
√c1c2g2π ). This corresponds to point one on Figure 3.4.
The maximum input signal lies around 80% of the comparator output level. This can
easily be explained by examining the Nyquist curve. If the amplitude of oscillation is too
big, the minus point will be at the left of the unstable point two. As was already discussed
the amplitude will build up until it reaches point one. point three and two are very close to
each other, which means a big input signal can easily push the system over the edge, even
if the amplitude of oscillation of the first point is very high. If the integrating coefficients
c1 and c2 are increased, points two and three will come closer to each other and the system
will more easily jump past point two.
Chapter 3. System design 17
3.3 Optimization
At first it was tried to optimize the system using the linear model as it should be accurate
enough to at least form a starting point for further optimizations. There are five param-
eters that can be considered for optimization. The integrating coefficients: c1, c2 and c3,
the delay d and the feedback gain g.
For ease of implementation, the delay is lower bounded to half a clock cycle. A higher
delay means a lower fosc and thus less quantization noise. However it also leads to a higher
amplitude of oscillation which lowers the comparator gain and thus noise suppression.
Simulation showed that small delay gives better performance, so d was chosen to be 1/2.
The coefficient c3 does not have much influence. It only scales the input of the com-
parator but it does not change the zero crossings, which is the only thing that matters for
the comparator. So c3 does not have to be included in the optimization. It does however
have an influence on circuit level. A large c3 leads to a high output swing and bandwidth
of the third op-amp. If c3 is too small, errors could arise because of a non-perfect com-
parator. Therefore a value of c3 is chosen through simulation, that offers a reasonable
output swing.
The feedback gain g determines, together with c1 and c2, the position of the zero in
the NTF. Although an unstable system oscillates at the frequency of the pole in the loop
gain determined by g, g does not influence the stability of the system. Figure 3.6 shows
the Nyquist plot for a system with g = 0. This system does not have a pole but it still has
a point of unstable oscillation and if the gain of the comparator is too small, the system
will still become unstable, with or without the pole. Furthermore it was assumed that the
optimal values of c1 and c2 do not depend on g. This greatly simplifies the optimization
and seems a valid assumption.
Figure 3.6: Nyquist plot of the loop gain with g equal to 0
In the end only c1 and c2 needed to be optimized and afterwards g was chosen to place
the pole at an optimal position. As stated before the linear approximation of the system
performs reasonably well. However when the coefficients were optimized, the resulting
system was unstable, even without an input. Here as before unstable means it started
Chapter 3. System design 18
to oscillate at a much lower frequency than expected. As stated before, increasing the
coefficients moves points two and three on the Nyquist plot (3.4) closer to each other.
Because of nonlinear effects, the amplitude of oscillation is able to become bigger than
the amplitude of the highest frequency oscillation and the unstable point two pushes it
towards the low frequency oscillation.
To save the linear model for optimization purposes it could be demanded that point
two and three remain at a certain distance of each other. However because of limited
time this possibility has not been fully explored and a brute force approach was used to
optimize the design. c1, c2 and c3 where chosen to be 2π0.015fs, 2π0.045fs and 2π0.0225fs
respectively. For this values, the SNR is about 75dB for a -8dB input at 10MHz. If full
scale is defined as D, although we can only go up to 0.8D, this is a -2dBFS input.
3.4 Slewing
An op-amp forming an integrator can be represented as a voltage controlled current source
loaded by a capacitor C. This means the rate of change of the output voltage is equal to
the delivered current divided by the capacitance. A rapidly changing output voltage can
only be delivered by a large current. However the op-amp is not able to produce an infinite
amount of current and the maximum current it is able to deliver is directly related to the
amount of power it consumes. This means the rate of change of the output is limited, and
when the output is changing at this maximum rate, the op-amp is slewing. Slewing is a
non-linear effect which introduces distortion in the SDM and degrades performance.
The feedback signal at the input of the first op-amp is the pulse width modulated
version of the input U. This means it contains two components: the slowly varying input
signal and the oscillation. The oscillation signal is the source of the slewing, but as this is
a non-linear effect, superposition no longer holds and the integrator is also non linear for
the slowly varying component. If the current runs out, it runs out for all inputs.
But even when the current remains under its maximum level, performance is still
degraded. Figure 3.7 illustrates this. The output current of the op-amp is only a linear
magnification of the input voltage as long as it remains small. The higher the current, the
more non-linear the integrator becomes and the more the performance is degraded. This
could be called soft slewing.
This problem mainly presents itself at the first op-amp because the input of the second
integrator is already softened by the first integrator and the third op-amp is loaded with
a smaller capacitance. Non-linearities are also less suppressed by the loop filter if they are
introduced close to the input. So the problem is caused by the oscillating component of
the feedback signal at the first op-amp. It is not necessary for the outer loop to process
this oscillation. It would be better if the first op-amp processes only the slow signal U
because of slewing and bandwidth limitations.
Chapter 3. System design 19
Figure 3.7: Slewing
To reduce the oscillation, a low pass filter was introduced in the outer feedback path.
However this filter changes the loop gain leading to a more complicated design. It changes
the loop filter into a fourth order instead of a third order filter. This is compensated by
adding a second filter that bypasses the first integrator as can be seen on Figure 3.8.
Σ
DAC
U Vfs
z−d
1 bit
c1s
c2s
c3sΣ
g
DAC
z−12
1 bit
11+sτ c1τ
1+sτ
Σ
Hc
x
y
Figure 3.8: diagram of the third order SOSD-modulator with Low-Pass filter and compensation
If a first order low pass filter is used, 11+sτ , the compensation filter, Hc, can be cal-
culated by demanding that the transfer function of the path between x and y remains
unchanged. The system with the filters satisfies:
y = −c1
s(
1
1 + sτx+ g
c2
sy)− xHc (3.2)
y
x= −
c1s
11+sτ +Hc
1 + g c2c2s2
(3.3)
In the system without filters Hc and τ are zero which means that Hc is subject to:
Chapter 3. System design 20
Hc +c1
s
1
1 + sτ=c1
s(3.4)
so
Hc =c1
s(1− 1
1 + sτ) (3.5)
=c1τ
1 + sτ(3.6)
If the compensation filter is dimensioned right, the oscillation is suppressed without
changing the operation of the system. There is however a trade off. If τ becomes too
large, the low pass filter affects the input component of the PWM signal. Normally the
error signal at the input of the first integrator does not contain a large contribution of U.
However if the feedback signal is low pass filtered, the phase of the input component of
the feedback signal changes a little, and the input is not completely eliminated. The first
op-amp receives a larger input and its output swing increases. Because this SOSDM will
be implemented in a 65nM technology with a supply of only 1.2V, this could become a
problem. Therefore a trade off will have to be made. The cut off frequency was chosen
to be five times larger than the bandwidth as simulation showed this offered a reasonable
output swing.
3.5 Noise
A major factor in determining the power consumption of the system is the maximum
allowed noise level. The circuit will consist of several resistors and transistors that will
create thermal and other noise. To determine the noise contribution to the output, several
noise sources are added to the model(Figure 3.9). One at each integrator input, one in
the local feedback loop before g and one before the Low pass filter. This is an accurate
representation as most resistors are connected to the input of the op-amps forming the
integrators and the noise of the op-amp themselves can be referenced to their inputs. Since
the SNR is around 75dB with a -8dB input, the amount of quantization noise is around
-83dB. This means that to not degrade the performance, the noise contribution should be
lower than this level. Therefore it will be demanded that the noise contribution in the
pass band is lower than -90dB. It is assumed that all three noise-sources deliver white
noise, so shot noise and 1/f noise will be neglected.
The noise source at the first integrator, N1, can be added to the input signal. This
means that the NTF1 is equal to the STF if there is no feed forward path from the input
to the third integrator. It was already determined that this STF approximates one, so the
noise of N1 is not suppressed. This means that the noise contribution of this source has
to be lower than -90dB. If Ng is substituted by a noise source after the gain g, this noise
contribution can also be added to the input. The transfer function NTFg is approximately
Chapter 3. System design 21
ΣU V
fsc1s
c2s
c3sΣ
g
DAC z−121 bit
Σ
N1 N2 N3
11+sτ
c1τ1+sτ
Σ
Ng
Σ
NLP
Σ
NLP
Figure 3.9: Model with noise sources
equal to g which leads to around 20dB suppression. This means the contribution of Ng
has to be smaller than -70dB. The source before the low pass filter can be replaced by
three sources. One at the input of the system, one at the input of the second integrator
and one at the input of the third integrator. Since the low pass filter offers almost no
attenuation in the pass band, the first restriction on NLP is that it has to be lower than
-90dB. With this restriction, the attribution of NLP two the input of the second and third
integrator is negligible.
The transfer function from N2 to V is given by:
N2
V=
1
1 + LG
Gsc3
s
c2s− gc1c2
s2 + gc1c2(3.7)
where LG is the loop gain. Figure 3.10a shows the transfer function from N2 to the output.
This NTF2 suppresses the noise with 20dB, this means the noise power in the pass band
of the source N2 should be smaller than -70dB. To easily calculate the contribution of N3
to the output, this source is substituted by a source after the comparator. To deliver the
same contribution, this source has to be of the form:Gs∗N3c3s . As this new source is in
exact the same position as the quantization noise source its NTF is the same: 11+LG so
NTF3 is:N3
V=
1
1 + LG
Gsc3
s(3.8)
NTF3 is depicted at Figure 3.10b. NTF3 suppresses the noise with 40dB, the result is that
the noise delivered by N3 should be smaller than -50 dB. During simulation it became clear
that too much noise could make the loop unstable. However this was only a problem at
Chapter 3. System design 22
10−4
10−3
10−2
10−1
−25
−20
−15
−10
−5
0
5
10
15
20
f*Ts
NTF2(dB)
(a) Transfer function from N2 to V
10−4
10−3
10−2
10−1
−80
−70
−60
−50
−40
−30
−20
−10
0
10
20
30
f*Ts
NTF3(dB)
(b) Transfer function from N3 to V
unreasonable high noise levels. To conclude this section, Table 3.1 shows the parameters
of the designed system.
Table 3.1: Parameters of the optimized system
Parameter system value circuit value
c1 2π0.015fs 30MHz
c2 2π0.045fs 90MHz
c3 2π0.0225fs 45MHz
fs 1 2 GHz
foscfs3 667MHz
fbfs100 20MHz
fLPfs20 100MHz
d 12 0.25ns
g 0.0901
D 0.5V
Chapter 4
The circuit
4.1 Top level
As mentioned before, the integrators are implemented using op-amps. Figure 4.1 shows
the realization of an integrator and the low pass filter. Because of negative feedback, the
op-amp will act as a nullator and offer a virtual ground at its input for differential signals.
This means x1 and x2 do not influence each other. If the gain is not high enough to offer
a perfect ground, which it never is, the inputs will interact with each other and there will
be some distortion. The output of this simple circuit is:
y = − v
R1Cs− x2
R2Cs(4.1)
Because a non-inverting integrator is needed, the wires will be crossed at the output. The
Voltage v is given by
v = x1
R1R1+RLP
1 + s2CLPRLPR1
R1+RLP
(4.2)
X2+ Y+
X2- Y-
+
--+
X1+
X1-
V+
V-
C
C
R1RLP
R2
CLP
RLP R1
R2
Figure 4.1: Circuit of integral of a sum of two signals, one of which is low pass filtered.
23
Chapter 4. The circuit 24
To find the response of this filter to differential signals, the capacitor had to be split up
into two capacitors in series. That is why the capacitance is double in Equation 4.2.
With this basic integrator and summing circuit, the total circuit can now be built.
However an adjustment has to be made. In the diagram of Figure 3.7, the output of Hc
is added to the output of the first integrator. This is not possible as the used integrator
circuit is only able to add signals at its input, not at the output. Also, in order to simplify
the system, the low pass filter and Hc will be merged since they are very similar. Figure 4.2
shows the final diagram of the system. Figure 4.3 shows the circuit. The three integrators
ΣU V
fsc1s
c2s
c3sΣ
g
DAC z−121 bit
11+sτ
Σ
c1τ
Figure 4.2: Diagram of the SOSDM as it will be implemented
and the comparator are clearly visible. The comparator is followed by a flip flop to realize
the sampling. The half clock cycle delay is implemented with a flip flop clocked at the
inverse clock. This second flip flop also inverts the signal. There are inverters between the
flip flops and the resistors. They serve a double purpose. Firstly they drive the resistor
with a voltage Vref . Vref is the equivalent of D, the output voltage of the comparator
on the system level. This voltage does not necessarily correspond to the digital voltage
supplied by the flip flops. Secondly they act as buffers between the analog and the digital
parts of the circuit. If the flip flops are too heavily loaded, their output voltage lowers
and they become slow. The buffer offers a lower capacitance and a high resistance. If
the capacitive impedance difference between the analog and digital part is too big, it is
possible that multiple inverters are necessary to minimize the delay. Because the two
feedback paths have a different loading, they have different buffers.
Chapter 4. The circuit 25
CLKN
VDD!
CM
Vref+
Vref-
Vref+
Vref-
Vref-
Vref+
Vref+
Vref-
CM
CM
CLK
VDD!
Bias
in-
in+
out-
out+
in+
in-
ref
out+
out-
in
Vref-
Vref+
out
in
Vref-
Vref+
out
in+
in-
ref
out+
out-
in
Vref-
Vref+
out
in
Vref-
Vref+
out
D
Dn
vdd
clk
out+
out-
in+
in-
ref
out+
out-
Vbias
in+
in-
vdd
out+
out-
D Dn
vdd
clk
out+
out-
RLP_a
CFB3_a
FF2
FF1
RLP_b
CFB3_b
Rin_a
Rin_b
RLP1_aCLP
RLP1_b
Rg_b
Rg_a
CFB1_b
CFB1_a
R12_b
RLP2_a
R12_a
RLP2_b
CFB2_a
CFB2_b
RLP3_a
RLP3_b
R13_a
R23_b
R23_a
RFB_a
R13_b RFB_b
RinFF_a
RinFF_bFigure
4.3:
Th
eto
ple
vel
of
the
circ
uit
.
Chapter 4. The circuit 26
4.2 Dimensioning the Passive components
The next step is to dimension the resistors and capacitors. To start all equations that
have to be satisfied are given. Rpar is the equivalent parallel resistance of RLP1, RLP2 and
RLP3. Lets first look at the circuit around the first op-amp:
1
Rg ∗ CFB1= g ∗ c1 (4.3)
1
RinCFB1= c1 (4.4)
2CLPRLPRparRLPRpar
= τ (4.5)
Here τ is the time constant of the low pass filter
RparRLP +Rpar
1
CFB1RLP1= c1 (4.6)
The equations for the circuit around the second op-amp are:
1
R12CFB2= c2 (4.7)
RparRLP +Rpar
1
CFB2RLP2= c1τc2 (4.8)
In Equation 4.8 c1τ is the gain of Hc. And the the third op-amp:
1
R23CFB3= c3 (4.9)
1
RinFFCFB3= c3 (4.10)
1
R13CFB3= c3 (4.11)
RparRLP +Rpar
1
CFB3RLP3= c1τc3 (4.12)
1
RFBCFB3= c3 (4.13)
Before these equations are solved, lets first discuss the influence of the values on the
performance of the circuit. The first constraint is the noise generated by the resistors.
This sets a maximum value on each resistance. Lowering the resistance however will
increase the capacitance of the feedback capacitors. This increases the output capacitance
of the op-amps which is bad for bandwidth and slewing and thus for power consumption.
Another incentive for increasing the resistor values is the resistive loading of the op-amps.
The lower this is, the more current the op-amp has to deliver to obtain a certain output
voltage. If the maximum deliverable current is too small, the output voltage will saturate
and this will degrade performance. To prevent this the current and the power consumption
will have to be increased. Apart from noise, the only drawback to large resistors is that
Chapter 4. The circuit 27
they require a large surface on the chip. A 100kΩ resistor is more than 45µm long, and this
will be taken as a maximum. So the resistors will be made as large as is allowed for noise
performance while keeping them smaller than 100kΩ. This is of course an arbitrary choice
and increasing this maximum resistance will decrease the power consumption as it leads
to lower capacitive loading and smaller resistive dissipation. The size of the capacitors is
also quite large and this should also be considered when choosing the resistor values.
Resistors generate white noise with a root mean square voltage of√
4 ∗ k ∗ T ∗R V√Hz
.
Here k is the Boltzmann constant, k = 1.38∗10−23 JK , and T is the absolute temperature in
Kelvin, which will be assumed to be 300K. Only the noise in the pass band is considered.
If the mean square of the noise voltage has to be smaller than a value V 2max in dB, then
the resistor will have to satisfy:
R <10
V 2max
10
4 ∗ k ∗R ∗ fB= Rmax (4.14)
here fB is the bandwidth of the modulator. RLP1 and Rin both correspond to N1 on
Figure 3.9 so their contribution has to be smaller than -90dB which leads to a Rmax of
3kΩ. RLP corresponds to NLP , which means it also has the same Rmax. Equations 4.4
and 4.5 show that Rg = Ring . However its noise contribution is also scaled with g so Rg
has a Rmax ofRin,max
g , which means that the noise requirement for Rg is satisfied, if it is
satisfied for Rin.
R12 corresponds to N2 and its Rmax equals 302kΩ This is limited to 100kΩ. RLP2
also corresponds to N2 so its Rmax is also 100kΩ. In fact, every resistor that is not yet
discussed is bounded by 100kΩ as they are allowed to be noisier than N2.
As there are 11 equations and 15 unknowns, 4 values have to be assigned to solve the
equations. Table 4.1 shows all the resistors and capacitors together with their values and
the maximum values for the resistors. Rin, R12, R23 and CLP were assigned before solving
the equations.
Due to a miscalculation the maximal value of R12 was mistakenly defined as 31kΩ
during the design of the modulator. This error was corrected too late to redesign. This
led to a slightly increased Power consumption as the minimal noise and load capacitance
of the second op-amp were too big. More importantly the load resistor of the first op-amp
could have been larger. The result will be a somewhat larger power consumption of the
system.
4.3 Preamp, Flip Flops and Buffers
The system has been simulated using a verilog model for the comparator as it has not been
designed. The comparator is not a critical component for the operation and its design
is rather simple. A possible implementation of the comparator could be a preamplifier
followed by a flip flop. This flip flop is also used to implement the half clock cycle delay.
Chapter 4. The circuit 28
Table 4.1: Values of the resistors and capacitors in the system.
Resistor Value Rmax
Rin 3kΩ 3kΩ
Rg 33.3kΩ 33.3kΩ
RLP1 913.7Ω 3kΩ
RLP 2kΩ 3kΩ
R12 31kΩ 100kΩ
RLP2 31.5kΩ 100kΩ
R23 100kΩ 100kΩ
RinFF 100kΩ 100kΩ
R13 100kΩ 100kΩ
RFB 100kΩ 100kΩ
RLP3 101.5kΩ 100kΩ
Capacitor Value
CLP 1.3pF
CFB1 1.77pF
CFB2 57fF
CFB3 35.4fF
A design was handed to me for the preamplifier and the flip flop. They will be discussed
for completeness and to estimate a total power consumption of the modulator. Figure 4.4
shows the preamplifier. It is a standard differential pair loaded by diodes. The power
consumption will mainly be determined by its GBWP. A good GBWP would be ten
times the oscillation frequency. This comparator is loaded by the flip flops. This digital
component has a very small input capacitance in the order of a fF. Since the preamp
is loaded with a high impedance and the input signal is already integrated, the power
consumption will mainly be determined by the GBWP and not by slewing or the voltage
across the output resistance. The current is determined by I = gmgmID
and gm is the
product of the GBWP and the load capacitance, CL. With gmID
= 15 and CL = 1fF , ID
becomes 2.8µA and the dissipated power 3.4µW . As will become evident later, this is
negligible compared to the power dissipated in the op-amps.
The flip-flops are depicted in Figure 4.5. It is very important for this component to
have identical rise and fall times. Simulation shows that if the rise and fall time differ
too much, this has a dramatic impact on the SNR. A difference of only two pico seconds
reduced the SNR with 7dB. Figure 4.6 shows an equivalent schematic. The first two
vertical inverters(4) have to force the rise and fall times to be equal. The two transistors
force one input low if the other is high. This increases the fall time. The first two horizontal
inverters(1) only react while the clock is high. The second pair(2) only reacts while the
clock is low. This means this flip flop is triggered on the falling slope of the clock. The last
two inverters(3) preserve the output voltage when the clock is high. They are the memory
of the system. If the clock is low, they offer a high impedance, and the second pair of
inverters can easily change the internal state. The second pair of vertical inverters(5) act
as a buffer, offering a clean digital output.
Chapter 4. The circuit 29
vddvdd
gnd!
vddvdd
gnd!
vddvdd
vddvdd
gnd!
Vbias
out-
in+
vdd
in-
out+
M9
M7
M3
M4
M8
M5
Figure 4.4: A possible circuit for the preamp
The flip flop circuit has no static dissipation. If the input and clock do not change,
the power consumption will be next to nothing. The only dissipation results from loading
and unloading the capacitors. The amount of energy stored on a capacitor is equal toCV 2
2 . During unloading of the capacitor, this energy is dissipated in resistive elements. If
the voltage on a capacitor is switched from V to 0 with a frequency of f , the dissipated
power is CV 2
2 f . There are several capacitors in the flip flops. Most are very small. The
biggest capacitors are the gate capacitors and they have a value of around 0.5 fF. The
voltage on these capacitors changes twice per oscillation period. If the input of the flip
flops rises , half the capacitors will load from 0 to V volts, and the other half will unload.
There are about 22 transistors whose gate capacitor is loaded once per oscillation period.
There are also 16 transistors that are driven by the clock. This means the dissipation
can be estimated by 22CV2
2 fosc + 16CV2
2 fs = 17µW . Most of this power comes from the
faster switching, clock driven transistors. There are two flip flops but even 34µW is small
compared to the power dissipation of the op-amps.
Most of the power consumption of the system will be in the op-amps. The flip flops
and preamp account for a small part of the power budget. The power consumption of the
buffers however could be quite large. Their output can be considered the superposition
of a differential and common mode part. The common mode voltage is constant and the
same at every node. This means no common mode current flows in the circuit. The
dissipation of the differential part is determined by the load resistance. Most resistors are
quite large and most differential signals rather small. However this is not true in the first
part. Due to noise limitations the resistors are small there. The series resistance of RLP
and RLP1 is approximately 3kΩ. They are by far the smallest resistors. The input of the
first op-amp offers a virtual ground for differential signals. The buffers switch between
Chapter 4. The circuit 30
+0.6 and -0.6V. For DC this gives rise to a current of 0.2 mA. This current is delivered
half the time by the first op-amp and half the time by the buffer. This gives rise to a total
power consumption of 0.48mW because there are two buffers. This consumption is quite
a large part of the total power budget. The other resistors are at least ten times higher
than these resistors so the other buffers have no significant contribution. We only included
static consumption in this analysis. The extra consumption caused by the loading of the
capacitors is given by CV 2
2 fosc, Where C is the load capacitance and V the difference in
voltage before and after loading. C is approximately CLP=1.3pF. Using Equation 4.2
we find that V is 0.06. So the dynamic dissipation is 2 µW which is small enough to be
ignored. This also means that the low pass filter did not introduce a significant amount
of extra dynamic dissipation because of higher capacitive loading of the buffers.
clk-
clk+
clk-
clk+
vddvdd
vddvdd
vddvdd
vddvdd
gnd!
gnd!
gnd!
vdd
gnd!
gnd!
gnd!
vdd
gnd!
gnd!
vddvdd
vddvdd
gnd!
gnd!
clk-
vdd
vdd
vdd
vdd
vdd
vdd
vdd
vdd
clk+
gnd!
gnd!
vddvdd
vddvdd
gnd!
gnd!
clk-
clk+
clk-
clk+
vddvdd
vddvdd
vddvdd
vddvdd
gnd!
gnd!
gnd!
vdd
gnd!
gnd!
gnd!
vdd
vddvdd
vddvdd
gnd!
gnd!
gnd!
gnd!
gnd!gnd!
gnd!gnd!
vdd
vddvdd
vdd
clk+
clk-
clk+
clk-
vddvdd
vddvdd
vddvdd
vddvdd
vdd
gnd!
gnd!
gnd!
vdd
gnd!
gnd!
gnd!
D
Dn
vdd
clk
out-out+
M23
M24
M25
M26
M29
M30
M27
M28
M31
M4
M20
M32
M21
M22
M34
M33
M41
M42
M35
M40
M38
M39
M36
M37
M52
M51 M55 M57 M56
M58
M43
M44
M45
M50
M46
M47
M48
M49
Figure 4.5: A flip flop circuit
CLK
CLK
CLK
CLK
CLK
CLK
D
Out
D
Out
1 2 3
4
5
Figure 4.6: The equivalent diagram of the flip flop circuit
Chapter 5
Op-Amp Design
The most crucial parts of the Sigma Delta Self Oscillating modulator are the op-amps.
They implement the integration and summation of the signals. If the integration coeffi-
cients are changed, the operation of the system will change and the loop could even become
unstable. They are also the components that consume the lion’s share of the power. Most
components are standard building blocks that only require a rescaling for this circuit. The
op-amps however were completely designed for this modulator.
5.1 Specifications
The op-amps have to comply with several specifications which will be discussed here.
5.1.1 GBWP
When the transfer function of the integrator circuit was calculated, it was assumed that the
op-amps had an infinite speed. It was taken for granted that the differential input signal
of the op-amp was forced to zero by the negative feedback, even for unlimited frequencies.
A more exact expression is found by assuming that the op-amp acts as an integrator with
gain A = 1τas
. This means it has a gain bandwidth product of 12πτa
. Figure 5.1 again
shows the simple integrator circuit but with only one input. The following is derived:
Y = −Aε (5.1)
ε = YsRC
1 + sRC+X
1
1 +RSC(5.2)
With τi = RC this becomes
Y
X=
−A1 + (A+ 1)sτi
(5.3)
Which for this op-amp reduces to
Y
X=
−1
s(τi + τa)
1
1 + s τiτaτi+τa
(5.4)
31
Chapter 5. Op-Amp Design 32
X+ Y+
X- Y-
+
--+
ε
Figure 5.1: The integrator
This means that instead of an integrator coefficient equal to 1τi
, the coefficient becomes1
(τi+τa) . This effectively lowers the coefficients which in turn degrades the performance of
the system. To prevent this, τa is made ten times smaller than τi. This is a specification on
the Gain Bandwidth Product of the op-Amp. Even with this specification, the integrator
coefficient will still be changed slightly. This could be compensated for by changing the
feedback around the op-amp so τi changes to τ ′i . If τ ′i equals τi−τa, the unity gain frequency
of the integrator will be exactly correct. The GBWP of the op-amp could even be lowered
to save power. However it is to be taken into account that changing the feedback also
changes the loading of the op-amp which influences its GBWP. Furthermore τa is less
precise than τi so this compensation technique will not be relied on too hard.
Not only will the integrator have a lower unity gain frequency, for higher frequencies
it behaves as a second order integrator. If τa is much smaller than τi, this happens at the
GBWP of the op-amp. This should not be a problem for the first and second op-amp.
For those high frequencies the signals are already very small and the extra phase shift has
not a great influence. However this could be a problem for the third op-amp as the phase
shift of the third integrator is a determining factor for the oscillation frequency. During
the design of the third op-amp measures should be taken to prevent the phase of the third
integrator from drifting too far from π2 .
5.1.2 DC Gain
It was assumed that the op-amps behaved as integrators, which is a good assumption for
high frequencies. However it requires an infinite gain at DC which is clearly not realistic.
In reality the op-amp will behave more like a low pass filter with a finite DC gain ADC .
To determine the effect of a finite gain A is replaced by ADC in Equation 5.3.
Y
X≈ −ADC
1 +ADCτis(5.5)
Here ADC was assumed to be bigger than one. It is clear that the integrator is in fact a
low pass filter. There will be a finite suppression of the low frequent noise instead of an
Chapter 5. Op-Amp Design 33
infinite one. This is not a major problem as long as ADC is large enough. The effect will
be a lower bound on the transfer function of the quantization noise. The zero placed near
the bandwidth will no longer be a perfect zero. To determine the needed ADC the system
is integrated with low pass filters instead of integrators. If all op-amps have an ADC equal
to 200, the resulting SNR is degraded by less than 1dB. There is another effect of finite
ADC (and GBWP). If the ADC is too small, there will be distortion as multiple inputs of
an integrator start to interact. Therefore it is better to realize a higher ADC. With this
in mind the number of stages of the op-amps where chosen to be three. This offers more
than enough gain at almost no extra cost.
5.1.3 Slewing
DC-gain and GBWP are both linear specifications in that they both specify the linear
operation of the op-amp. To prevent the op-amp from operating as a non-linear component
a minimum current Imin will have to be delivered to prevent slewing. An approximate
result will now be derived for this Imin. Assume a sine input with frequency fin and
amplitude Ain. The amplitude at the output of an integrator cs , will be:
Aout = Ainc
2πfin(5.6)
so
Imin = CL
(dV
dt
)max
(5.7)
= CL2πAoutfin (5.8)
= CLAinc (5.9)
The result is that the current needed to prevent slewing does not depend on the rate of
change of the input. Although a faster changing signal will need more current to load the
capacitor to a certain voltage, this voltage will be lower because of harder suppression by
the integrator. Of course it was assumed that the circuit still operates as an integrator
at high frequencies, which would assume an infinite GBWP. However Equation 5.4 shows
that if the input is too fast for the integrator circuit, it will act as a double integrator
which would even lower the needed current further. It can be concluded that Equation 5.9
is a good upper boundary on the amount of current needed to prevent slewing.
5.1.4 Other Specifications
Another non-linear specification is the deliverable DC-current. Apart from CL, the op-amp
is also loaded by a resistor. This means the op-amp has to deliver a current to generate a
voltage. This leads to the following demand IDC >VDD2RL
where VDD is the supply voltage
which is 1.2V
Linked to the output swing is the common mode voltage. This voltage is taken to be
600mV both for the output as the input common mode voltage for all three op-amps and
Chapter 5. Op-Amp Design 34
the comparator input. This voltage is half the supply. For simplicity the input and output
common mode voltage were chosen the same. A lower input common mode voltage and
a higher output common mode voltage would have been better for output swing but it
would make the circuit more complicated.
Not only the resistors generate noise, also the transistors. Although every transistor
is noisy, most of the noise is determined by the first stage. This noise is amplified by all
following stages. This noise will be represented as a voltage source at the input of the
op-amps. If only white noise is included, the mean square voltage of this source is:
2
3
4kTα
gm(5.10)
α is the equivalent amount of noisy transistors in the first stage. For this design this
will be smaller than four for all op-amps. α will be determined more accurately during
the design of the first op-amp. It will be demanded that this noise is less than the noise
generated by the resistor at the input of the op-amp.
A last important specification is of course stability. A phase margin of 60 will be
demanded for each op-amp, which is a much used number.
The first and second op-amp have been designed and simulated. The design of the
third op-amp is very similar to the design of the other two op-amps but has not been
implemented.
5.2 General Feed Forward Op-Amp Design Flow
This design flow is largely based on an internal paper by Professor Rombouts. Because
the purpose of this thesis is to design a low power ADC, the Feed Forward architecture is
a good choice. To stabilize, it puts parasitics at frequencies lower than the dominant pole
instead of higher. As faster circuits use more power, this is a low power solution. The slow
impulse response of the FF op-amp is not a problem for this modulator. The behavior in
the frequency domain is what determines the performance of our modulator. As stated
before a three stage op-amp will be used because of DC gain considerations. A general
design flow will now be outlined for such op-amps. The goal is to determine the transistor
dimensions and the current source of the differential pairs. There are five differential pairs
to dimension: three stages and two feed forward stages.
Figure 5.2 shows the small signal circuit of the op-amp. The R’s and C’s are the total
output resistances and capacitances of the stages. This means the DC gain is given by
−gm1R1gm2R2gm3R3−gmFF2R2gm3R3−gmFF3R3. However the DC gain will not be used
as a specification anymore because it already determined the amount of stages. Therefore
from now on it will be assumed that the output resistances is infinite. To further simplify
the design, the two zeros are placed at the the same frequency. This means the following
Chapter 5. Op-Amp Design 35
R1 R2 R3C1 C2 C3gm1Vin gm2V1 gm3V2
V1 V2 Vuit
gmFF2Vin gmFF3Vin
Figure 5.2: The small signal circuit of a three stage FF op-amp
transfer function is targeted:
H = − 1
sτ1
(1 + sτ2
sτ2
)2
(5.11)
Clearly the op-amp behaves as a third order integrator at low frequencies, but as a first
order integrator near the unity gain frequency. The phase margin of this system rises if
this switch over point between third and first order lies further from the GBWP, 12πτ1
. To
have a phase margin of 60, τ2 should at least be 4 times higher than τ1. It will not be
made smaller either to offer enough gain. First the equations linking the time constants
to the circuit are written down :
−H =1
sτ1
(1 + sτ2
sτ2
)2
(5.12)
=1
sτ1+
2
s2τ1τ2+
1
s3τ1τ22
(5.13)
And from Figure 5.2:
=gmFF3
sC3+gmFF2gm3
s2C2C3+gm1gm2gm3
s3C1C2C3(5.14)
So:gmFF3
sC3=
1
sτ1(5.15)
gmFF2gm3
s2C2C3=
2
s2τ1τ2(5.16)
gm1gm2gm3
s3C1C2C3=
1
s2τ1(τ2)2(5.17)
The first and second stage can’t be dimensioned yet because C1 and C2 depend on the
dimensions of the third stage. That is why the design starts at the last stage and works
its way back.
step1: gmFF3
There are several restrictions on gmFF3. Firstly the GBWP specification, which together
with Equation 5.15 gives:gmFF3
2πC3> GBWP (5.18)
Note that C3 is not known exactly as it also contains some drain gate and drain source
capacitance from the transistors that still need to be dimensioned. But these capacitances
Chapter 5. Op-Amp Design 36
are lower than CL, the load capacitance of the op-amp. So CL is a good estimate for C3.
If capacitive sensing is used for the common mode feedback, than this contribution to C3
should not be forgotten.
To prevent slewing, the third stage has to be able to deliver a certain current Imin.
This gives us a second restriction:
gmFF3gmID
> Imin (5.19)
The value of gmID
is a design choice. A high gmID
offers low power in exchange for large
dimensions. A good value of gmID
depends on the technology that is used. Linearity of
the transistor should also be considered while choosing gmID
. Note that different stages can
have a different gmID
. Now gmFF3 is determined by the most demanding restriction.
step2: gm3
gm3 is chosen as follows:
gm3 =gmFF3
SF3(5.20)
Here SF3 is an arbitrary scale factor. It is better for power consumption to make SF3 big.
However gm3 must not become too small as not to violate the following specification:
gm3gmID
> Imin =VDD2R3
(5.21)
It must not be forgotten that R3 does not only include the external load resistance, but
also the resistors used for the common mode feedback. Note that it was assumed that the
third feed forward stage does not contribute to the DC current. Because of stability the
none-feed forward stages are made slower than the feed forward stages. However at low
frequency they have a much higher gain than the feed forward stages. This means that
because of feedback, the input voltage of the op-amp will be very low at DC and almost
no current will be delivered by the feed forward stages. That is why the third stage has
to deliver all the DC current.
gm3 is determined by the most stringent constraint and SF3 is updated. It is possible
that gm3 becomes bigger than gmFF3. This is not desired. If this is the case, gmFF3 will be
increased to the value of gm3. Now the input capacitance of the third stage is calculated.
This approximates C2.
step3: gmFF2
Equation 5.16 is rewritten using Equation 5.15 and Equation 5.20:
gmFF2 =2SF3C2
τ2(5.22)
step4: gm2
As before gm2 is a scaled version of gmFF2:
gm2 =gmFF2
SF2(5.23)
Chapter 5. Op-Amp Design 37
If the common mode feedback resistors are not negligible, enough current will have to be
delivered to produce a high enough output voltage. But in this design the common mode
feedback of the second stage is not implemented with resistors as this would eat away
the gain and ask a lot of space. Now the input capacitance of the second stage can be
determined. This will be the major part of C1.
step5: gm1
Equation 5.17 is rewritten :
gm1 =C1SF2
τ2(5.24)
step6: Noise
Use Equation 5.10 to check if gm1 is big enough for noise constraints. If gm1 is too small,
add an explicit capacitor C1e at the output of the first stage. Set gm1 to its minimal value
and size C1e as follows:
C1e =gm1τ2
SF2− C ′1 (5.25)
Where C1 = C ′1 + C1e
step 7: gm2 and gmFF2
If gm1 is very different from gm2, an explicit capacitor can be added to C2. Recalculate
gm2, gmFF2 and C1e.
During several steps approximations were made. This means this is an iterative pro-
cedure. All three output capacitors contain parasitics that are not known up front. If C3
contains a large parasitic contribution, the GBWP and the phase margin will be reduced.
An underestimation of C1 and C2 increases stability at the cost of some gain. Even if
the estimate of the output capacitances is correct, the phase margin will still be higher
than 60. It was assumed that the phase of the op-amp was −270 for low frequencies but
this phase is in fact higher because the output resistors were neglected. If the resulting
phase margin is much more than desired, it can be traded for a higher gain. Slewing was
only discussed at the output of the op-amp. But if the output stage is very large, C2 is
very big and there could be slewing at the second stage. Therefore it should be checked
through simulation that at no point during operation, the current delivered by any stage
approaches its maximum value too close.
5.3 Design
All three op-amps will have a very similar architecture: A three stage feed forward op-
amp. There will be differences concerning the common mode feedback of the last stage
and whether or not there is cascode in the first stage. The architecture will be discussed
while going over the first op-amp.
It was already decided to use three stages, however to boost the gain, the first stage will
be implemented with a telescopic cascode. Figure 5.3 shows a cadence plot of the first stage
Chapter 5. Op-Amp Design 38
together with the bias circuits and the common mode feedback. Names of components of
the circuit will be written in italics. The core of this stage is the differential pair: T1n a,
T1n b with cascode transistors: TCn a, TCn b. This is loaded with a current source:
T1p a, T1p b which is also cascoded: Tcp a, Tcp b.
5.3.1 CMFB
T1CM is a triode transistor. Its gate voltage is the output voltage. It implements the
common mode feedback. The gate voltage of T1p is given by VDD − RT1CMID1 where
ID1 is the current flowing through T1CM. If the output common mode voltage rises, the
resistance of T1CM becomes bigger and the gate voltage of T1p in turn falls. This lowers
ID1 and since TCp is a current buffer, this also lowers the output current which in turn
lowers the output common mode voltage. The bias circuit generates Vbias such that T1p
injects a current equal to half the current of Istaart1 if the output common mode voltage
is equal to the voltage generated by refintern.
A big advantage to this kind of common mode voltage feedback is that it doesn’t load
the output of this stage with a low resistance or a high capacitance. It only loads the
output with the gate capacitance of T1CM. This capacitive loading is not an issue as a
big capacitor already has been placed to preserve the stability. Resistive loading of this
stage however would lower the gain. The high output impedance created by the cascode
would be in parallel with the much lower resistance of the common mode feedback. This
common mode feedback resistance can’t be made very large because its dimensions would
be unreasonably high. The disadvantage of this implementation is that it is only accurate
for a low output swing. Since this is only the first stage, the output swing at this point
is rather small. For a reference voltage of 600mV, the output common mode becomes
643mV. As expected this is not very accurate but it will do.
5.3.2 Outputswing
The output swing of this stage is given by:
VCM,in − VGS,T1n + VDsat,T1n + VDsat,TCn < Vout < VDD − VDsat,T1p − VDsat,TCp (5.26)
This means a lower VCM,in offers a higher output swing. For minimal VCM,in this becomes:
VDsat,Tstaart1uit + VDsat,T1n + VDsat,TCn < Vout < VDD − VDsat,T1p − VDsat,TCp (5.27)
The voltage over T1CM was neglected. Since this transistor is in triode its VDS is rather
small. After the first stage is dimensioned the maximum output range can be put in
numbers: 0.2690 < Vout < 0.9540. This has a midpoint of 0.6115 and a total swing 0.6850.
However with an input common mode of 600mV this becomes: 0.3972 < Vout < 0.9540.
The middle point is now 0.6756 and the swing 0.5568. However simulation showed that
Chapter 5. Op-Amp Design 39
the needed output swing for this stage is not even 50mV so there is more than enough
swing and the input common mode voltage is not critical.
5.3.3 Cascode
Figure 5.4 shows the small signal equivalent for differential signals of the first stage. For
simplicity it is assumed that the n and corresponding p transistors have the same small
signal parameters. Lets first calculate the output impedance of the upper cascode:
R′out =Y
IL(5.28)
= (1 + gmcR1)Rc +R1 (5.29)
≈ gmcR1Rc (5.30)
This impedance is in parallel with the output impedance of the lower cascode. The cascode
transistor acts as a current buffer which means Iout approximates gm1Vin. This makes for
a DC gain of gm1R′out/2 = gm1Rout. There are two poles: a dominant pole caused by
the load capacitance of the stage, and a pole caused by the capacitance on point X. The
resistance at X is approximately 1gmc
because the cascode transistor offers a low input
resistance due to its operation as a current buffer. With these poles the total transfer
function becomes:Y
X=
gm1Rout
(1 + sC1Rout)(1 + s CXgmc
)(5.31)
The effect of C1 was included in the previous stability analyses, however CX was not
brought into account. This pole has to be at a higher frequency than the GBWP of the
op-amp to ensure stability. This should not be a problem because CX is much smaller
than CL but if it is a problem it can be solved by increasing gmc.
5.3.4 Biassing
Figure 5.3 shows four bias circuits. The biasing of the tail transistor is rather straight-
forward. It is more difficult to produce the biasing voltages of the cascode transistors.
VCBiasn has to be large enough to prevent the two lower transistors from going into
triode mode. It has to be equal to 2VDsat + VGS + margin. From this bias circuit it is
derived that VCBiasn = VDS,TCBiasnstaart +VDS,TCnBias triode +VGS,TCBiasnin. To ensure
that this is large enough, TCnBias triode is a downsized version of TCn. The smaller it
is made, the larger the voltage across it will be and the higher the margin will be. The
channel width of TCnBias triode is made five times smaller than the width of TCn. The
other bias circuits are similar. To reduce the power dissipated in the bias circuits, the
injected current and the width of the transistors are also scaled. Preferably the amount
of fingers is reduced so the width per finger remains the same. It is important that the
bias circuit of TCn is scaled. Not only because of power consumption, but because it is
Chapter 5. Op-Amp Design 40
out1+
gnd!
VDD!
VDD!
VDD!
Vref
VBias2
gnd!
gnd!
VBias2
VDD!
VDD!
VDD!
VDD!
Vref
Vbias
out1-
out1-
gnd!
VDD!
VDD!
Vbiasstaart1
gnd!
out1-
VDD!
Vbias
VCBiasn
gnd!
VBias2 VDD!
VDD!
gnd!
out1+
out1+
Vbiasstaart1
gnd!
gnd!
VDD!
gnd!
VCBiasn
in+ in-
TCp_scsc
TCpBias_triode
T1CM_scsc
Tstaart1in Tstaart1uitTCBiasnstaart
ICM1_sc
refintern
T1CM_sc
T1p_sc
Istaart1
ICM1
TCp_sc
TCn_a
TCp_a
T1p_a
Cint1_b
T1n_a
T1CM_a
T1n_b
T1CM_b
Cint1_a
T1p_b
TCp_b
TCn_b
ICBiasn
TCBiasnin
TCnBias_triode
Figure 5.3: The first stage of the first op-amp
Y
gm1Vin
-gmcX
gmcZ
R1
Rc
Rc
R1
X
Z
Figure 5.4: The small signal circuit of the first stage of the first op-amp
Chapter 5. Op-Amp Design 41
directly connected to the first stage and can introduce an error in the biasing current of
this stage.
5.3.5 Other Stages
The other stages and the feed forward stages are standard differential amplifiers without
cascode and are depicted in Figure 5.5 and Figure 5.6.
The second stage uses the same kind of common mode feedback as the first stage. The
output common mode voltage is 746.8mV. This is a large error because the reference was
also 600mV. It needs an output swing around 120mV. This stage doesn’t use cascode so
the output range is:
VCM,in − VGS,T2n + VDsat,T2n < Vout < VDD − VDsat,T2p (5.32)
0.3430 < Vout < 1.04 (5.33)
The swing is 0.697 and the midpoint 0.6915. Clearly the output swing is high enough.
The feed forward stage does not restrict the output swing as it has a lower input VCM,in.
Only the third stage uses common mode feedback with capacitive and resistive sensing.
The reason is that the output is already loaded with a large capacitor and a low resistor.
The extra loading from the common mode feedback doesn’t make a large difference here.
Furthermore, the output swing at this point is higher so the conventional common mode
feedback is more accurate here. The circuit is depicted at Figure 5.7.
The gain is gmn/gmp which is close to one. The total gain of the common mode
feedback is:gmngmp
gmT3pR3 (5.34)
Because the third stage needs a large bandwidth, it has large transistors.This means the
common mode feedback has a large load capacitance and so stability could become a
problem. The time constant of the output pole is CL,CM/gmp where CL,CM is mainly the
gate source capacitance of T3p. For the common mode feedback loop to be stable it is
needed that:CL,CMgmp
<C3
gmT3p(5.35)
with C3 the outpout capacitance of the third stage. This kind of common mode feedback
has clearly a high gain as the output common mode voltage is 602.3mV. This is a deviation
of only 2.3mV. The current source of the CMFB is a scaled version of the current flowing
through T3p. The pmos transistors of the CMFB are also scaled versions of this transistor.
Connecting them like diodes ensures that they are in saturation which is where T3p needs
to be. The same form of biasing is used for all the non-cascode transistors.
The third stage needs an output swing of 286mV. The expression of the output range
is the same as Equation 5.32. In numbers this is:
0.419 < Vout < 1.042 (5.36)
Chapter 5. Op-Amp Design 42
Figure 5.5: The second stage of the first op-amp
Figure 5.6: The third stage of the first op-amp
Chapter 5. Op-Amp Design 43
VDD!
gnd!in-
in+CMFB
CMref
R1
R0
C0
C1
Tn_a
Tp_a
I0
Tp_b
Tn_b
Figure 5.7: The common mode feedback circuit of the last stage of the first op-amp
which offers a swing of 623mV and the center point is 730.5mV. The output feed forward
stage also does not restrict the swing because of its lower VCM,in. There is enough swing
for this stage but the output voltage comes close to its minimal value during simulation
(466mV). So in further design it should be considered to lower the reference voltage of the
second stage so its output common mode is lower.
5.3.6 Noise
The equivalent noise source at the input of the op-amp needs to be determined. As
previously mentioned, only the noise of the first stage is significant. The mean square of
this voltage is given by Equation 5.10 so α, the equivalent amount of noisy transistors,
has to be found. It is clear that the transistors of the differential pair contribute directly
to the noise. The noise current of T1p, InT1p, is buffered by the cascode transistor TCp.
This means it contributes directly to the output voltage as Vout,nT1p = InT1p ∗Rout. If this
is divided by the gain the equivalent input noise is derived:
Vin,nT1P =Vout,nT1p
gm1Rout(5.37)
=InT1p
gm1(5.38)
The mean square of InT1p is 234kTgmT1p. This leads to a total contribution from T1p and
T1n to α as:
α = 2 ∗ (1 +gmT1p
gmT1n) (5.39)
However the cascode transistors are noisy too. Figure 5.8 shows the small signal
equivalent of the first stage. The loading of the cascoded differential pair is represented
as a resistance=R′out. The noise of the biasing and common mode feedback circuits is not
Chapter 5. Op-Amp Design 44
included. Their noise will only contribute to the common mode voltages of the circuit
which is not a problem for performance. There is one noise-current source representing
the noise produce by TCn. The analysis for TCp is similar. The input voltage is put to
zero as it is not relevant for this analysis. The contribution of In to I1 needs to be found
. The input voltage corresponding to I1 is I1gm1
. From this circuit the following equations
are derived:
X = I1R1 (5.40)
I1 = In − gmCX +Y −XRC
(5.41)
Y = −I1R′out (5.42)
After solving this becomes:
I1
In=
RCRC + gmCR1RC +R′out +R1
(5.43)
BecauseR′out and gmCR1RC are very large, the noise contribution of the cascode transistors
is negligible and Equation 5.39 holds. During the design α will still be assumed to be four
as a worst case scenario.
Y
-gmcX Rc
R1
X
R'out
In
I1
Figure 5.8: The small signal equivalent of the first stage of the first op-amp with added noise
source
5.3.7 Op-Amp1
The specifications of the first op-amp are summarized in table 5.1. CL and RL already
include the loading by the common mode feedback. To calculate Imin,slewing, Ain needs to
be known. The oscillation is suppressed by the low pass filter with a cut off frequency atfs20 = 100MHz. So a sine input with frequency fs
3 = 667MHz and an amplitude of 0.6V,
is suppressed to an amplitude of 90mV. This leads to a minimal needed current of 32µA.
Without the low pass filter, Imin would be 215µA. The maximum deliverable current of the
Chapter 5. Op-Amp Design 45
designed op-amp delivers 422µA. This means that in this design slewing was not the main
factor determining the power consumption. The low pass filter just increased linearity due
to soft slewing. Now that all the specifications are lined up the first op-amp can finally be
dimensioned. During the dimensioning the previously outlined steps were followed, but
because this is an iterative procedure, only the results will be given. Note that whenever
a certain current is given, this is the current flowing through one transistor. If the power
consumption needs to be calculated, these currents will have to be doubled.
The current flowing through the third feed forward stage is 211µA. This is much larger
than Imin,slewing so the power consumption of this stage is determined by the GBWP. In
fact the power consumption is really determined by noise specifications. Due to noise, Rin
needed to be small which in turn made CL large. there is a clear trade off here between
performance and power consumption. If more noise is allowed, the power consumption
will be lowered. A gmID
of 17 was used, and it will be used as a default without always
mentioning it. The length of the transistor is 180nm. This length was chosen larger than
the minimum allowed length of 60nm because it offers a larger gain.
The third stage has a current of 101µA. This is determined by the resistive loading of
the stage. This is also one of the largest power consumers of the circuit. The gmID
forT3p
is 10. This reduces the dimensions of this big transistor. The trade off is a smaller CMFB
gain. This also improves the stability of the common mode feedback as it reduces the
capacitive loading and shifts the dominant pole to lower frequencies.
IDFF2 is 6µA. There is an explicit capacitor of 22.5fF at the output of this stage. This
is the value of the differential capacitor. It is implemented with two capacitors in parallel
with each a fourth of this value. The reason for this capacitor is the high gm of the first
stage. The channel length of TFF2 is 240nm. This increases gain while this transistor
remains rather small.
The second stage has the same current as the second feed forward stage, SF2 = 1,
because this stage uses very little power and to keep it from getting much smaller than
the first stage. Even with this explicit capacitor and SF2 = 1 the ratio of gm1 over gm2 is
still 8. The length of the p and n transistors is 360nm. The gmID
of T2p is 10 to reduce its
surface-size.
Table 5.1: Specifications of the first op-amp
GBWP 300MHz
PM 60
RL 23.7kΩ
CL 1.87pF
Rin 3kΩ
Imin,slewing 32µA
Chapter 5. Op-Amp Design 46
The first stage has a current of 52µA. This is determined by noise requirements. It was
demanded that the equivalent input noise voltage was less than the noise voltage created
by the input resistor. α was assumed to be four. For T1p gmID
also was 10. This means the
actual α is 3.2. So we have some margin on our noise. Because of the noise requirement
an explicit capacitor of 1.88 pF had to be placed to keep the op-amp stable. This means
the first stage is as heavily loaded as the third one. The lengths of T1n and T1p were also
240nm. The cascode transistors were dimensioned with the same length and gmID
as T1n.
Simulation of the op-amp showed a GBWP of 288 MHz and a DC gain of 83dB. The
phase margin was 67 and the unity gain frequency of the integrator was 30.27MHz. This
could be compensated for by increasing CFB1 a little. Figure 5.9 shows the bode plot
of the op-amp and Figure 5.10 shows the bode plot of the integrator formed with this
op-amp. Table 5.2 shows the channel widths and lengths of the transistors of the first
op-amp together with the dimensions of the CMFB resistor. The biasing transistors were
not included since they are just scaled versions of the transistors they are biasing. Also
the tail transistors were not included. They have the same length and the double width
of the TXn transistors. The power consumption of this op-amp is 0.9mW. This does not
include the consumption of the biasing circuits and the common mode feedback circuit.
5.3.8 Op-Amp2
The architecture of the second op-amp is the same as the first one with one difference.
Because the output resistance of the second op-amp is higher than the first one, the same
kind of CMFB will be used as in the first and second stage of the first op-amp. If resistive
sensing of the common mode voltage would be used here, the gain would be halved.
Table 5.3 shows the specifications on the second op-amp. This op-amp needs a higher
GBWP than the first one, but its load capacitance is much smaller. Figure 4.2 shows that
the second op-amp has two inputs. One input is low pass filtered and integrated, the other
is only integrated. The amplitude of this second input signal is:
Ain = 0.6 ∗ c1
2πfosc= 27mV (5.44)
From Equation 5.9, with c= 0.0452πfs, Imin is 0.87µA. This means slewing will not be a
problem.
The design of this op-amp will not be discussed with as much detail as the the design
of the first op-amp. Table 5.4 shows the needed and available output swing for every stage.
It is clear that this is never a problem. Table 5.5 shows the dimensions of the transistors.
Figure 5.11 shows the bode plot of the op-amp and Figure 5.12 shows the bode plot of the
second integrator. The GBWP of the second op-amp is 854MHz, the DC gain 80dB, the
unity gain frequency of the integrator is 90.68MHz and the phase margin 73. The total
power consumption is 0.1mW. For this design, a gmID
of 15 was used. Because this op-amps
power consumption is almost a tenth of the previous one, a lower gmID
was used because
Chapter 5. Op-Amp Design 47
50.0
100.0
150.0
0.0
V (deg)
200.0
-50.0
-25.0
0.0
V (dB)
100
50.0
75.0
25.0
freq (Hz)10
210
810
610
710
110
-110
010
910
1010
410
510
3
Figure 5.9: The bode plot of the first op-amp
Phase (deg) 100.0
50.0
-50.0
0.0
150.0
-100.0
200.0-75.0
-50.0
-25.0
0.0
25.0
50.0
75.0
100
(dB)
103
freq (Hz)10
010
110
210
410
510
610
710
810
910
1010
11
Figure 5.10: The bode plot of the integrator formed by the first op-amp
Chapter 5. Op-Amp Design 48
Table 5.2: The channel widths and lengths of the transistors of the first op-amp
Name Width(µm) Length(um)
T1n 9.69 0.24
T1p 7.52 0.24
TCn 7.75 0.24
TCp 31.56 0.24
T1CM 3 0.06
T2n 1.45 0.36
T2p 2.3 0.36
TFF2n 0.93 0.24
T3n 11.55 0.18
T3p 28.51 0.18
TFF3n 24 0.18
RCM 0.4 45.2
power consumption was less critical. However, one could just as easily increase gmID
and
trade larger dimensions for lower power. The reason why the second op-amp uses less
power is mainly its low load capacitance and high load resistance. IDFF3 = 22.4µA is
determined by bandwidth requirements. ID3 = 12.1µA is determined by resistive loading.
There is no explicit capacitor at the second stage. To comply with noise requirements and
keep the system stable, there is an explicit capacitor of 51 fF at the output of the first
stage. The first stage dissipates approximately 11.3% of the total power of this op-amp.
However, as was already mentioned, the noise requirements for the second op-amp could
be relaxed.
5.3.9 Op-Amp3
The specifications of the third op-amp are similar to those of the previous two op-amps
with one exception. The phase shift of the third integrator determines the oscillation
frequency. On Figure 5.12 it is clear that for this op-amp the phase is much lower than
Table 5.3: Specifications of the second op-amp
GBWP 900MHz
PM 60
RL 100kΩ
CL 58fF
Rin 31kΩ
Imin,slewing 0.87µA
Chapter 5. Op-Amp Design 49
Table 5.4: The needed and available output swing of the stages of the second op-amp
Stage Vout,CM (V ) minimum(V) center(V) maximum(V) swing(mV) needed swing(V)
1 0.751 0.386 0.711 1.035 0.649 0.01
2 0.741 0.445 0.743 1.04 0.595 0.081
3 0.579 0.392 0.7134 1.035 0.643 0.167
Table 5.5: The channel widths and lengths of the transistors of the second op-amp
Name Width(µm) Length(nm)
T1n 0.94 240
T1p 3.98 240
TCn 0.75 240
TCp 3.05 240
T1CM 3 60
T2n 0.27 360
T2p 0.63 360
TFF2n 0.17 240
T3n 0.94 180
T3p 3.33 180
TFF3n 1.88 180
M26: 854.4241MHz, 0.0dB
0.0
V (dB)
-25.0
100
25.0
75.0
50.0
M27: 854.424MHz, 73.4045dM28: 854.424MHz, 73.4045d
0.0
50.0
150.0
100.0
V (deg)
200.0
106
freq (Hz)10
1010
110
510
810
-110
910
710
010
210
410
3
Figure 5.11: The bode plot of the second op-amp
Chapter 5. Op-Amp Design 50
100.0
150.0
200.0
Phase (deg)
-100.0
-50.0
0.0
50.0
50.0
75.0
100 (dB)
-50.0
-25.0
0.0
25.0
104
105
106
107
108
109
1010
freq (Hz)10
010
110
210
3
Figure 5.12: The bode plot of the integrator formed by the second op-amp
90 at the oscillation frequency. It was already determined that for frequencies higher
than the GBWP, the integrator is actually a second order integrator(Equation 5.4). This
means the bandwidth might need to be increased to even higher frequencies. A GBWP
of ten times the oscillation frequency is probably needed to guarantee that the oscillation
frequency doesn’t change. An estimate of the power consumption can be made for a
GBWP of 6GHz. The op-amp is loaded by the comparator which has a very high input
impedance. It can be assumed that SF3 is high and ID3 can be neglected. The noise
requirements for this op-amp are less stringent than they were for op-amp2 which means
that ID1 can also be neglected. This means only IDFF3 has to be taken in account. For
CL of 35.4fF and a bandwidth of 6GHz, gmFF3 has to be 1.3mΩ−1. If gmID is 17, this means
IDFF3 is 78.4µA. Equation 5.9 is used and the result is that Imin,slewing equals 6µA. This
means the power consumption is determined by the GBWP requirement and is 0.19 mW.
For this op-amp, it is probably best not to use cascode, as it introduces a parasitic pole
at a high frequency. It was already concluded that a low gain at this point does not make
much difference.
5.3.10 Results
Table 5.6 shows the power consumption of the three op-amps together with the total con-
sumption. This consumption could be reduced to 1mW if R 12 is increased to 100kΩ. It is
clear that most of the power is dissipated in the first op-amp due to GBWP specifications.
Chapter 5. Op-Amp Design 51
If the system is simulated with ideal components the SNDR is 76.5 for a 10MHz input
signal with an amplitude of 0.4V. An ideal buffer was used with an output voltage between
0 and 1.2v. If the first and second op-amp are replaced with the designed op-amps the
SNDR is 74dB. This loss in performance could be due to increased delay or the small error
on the integrator coefficients caused by the finite GBWP. The cascode is not really nec-
essary. Even 60dB DC gain would probably be enough. For A 1MHz input the SNDR is
71dB. Now several harmonics lie in the pass band. The even harmonics are negligible, the
third harmonic is 77dB lower than the fundamental signal. The seventh harmonic is 80dB
lower than the fundamental. The other harmonics are at least 10dB lower. Figure 5.13
shows the output spectrum for this input.
Table 5.6: The power consumption of the three op-amps
op-amp Power(mW)
op-amp 1 0.9
op-amp 2 0.1
op-amp 3 0.2
Total 1.2
10−5
10−4
10−3
10−2
10−1
100
−150
−100
−50
0
f*Ts
FFT(dB)
Figure 5.13: The outputspectrum for a 1MHz input with 0.4V amplitude
Chapter 5. Op-Amp Design 52
Chapter 6
Conclusion
In this thesis I designed a low power ADC with a bandwidth of 20 MHz and a sample
frequency of 2 GHz. The design is similar to a third order sigma delta modulator, however
it was modified to implement pulse width modulation. This was achieved by making the
sigma delta loop oscillate at 666MHz and using a 1 bit quantizer. The result is that the
loop filter does not have to operate at the sample frequency. The fastest signal it has to
process is the oscillation. This reduces the power consumption of the system.
The power consumption can be reduced even further by adding a low pass filter in the
loop. This effectively reduces the slew rate requirements of the first op-amp. By adding
a compensation filter bypassing the first integrator, the influence of this low pass filter on
the loop filter was eliminated. This new technique reduces power consumption without
introducing the need to redesign the loop filter on the system level. While implementing
this loop filter it has been found that the cut off frequency of this low pass filter should
not be too close to the bandwidth. If the cut off frequency is too low, the output swing
of the first integrator is increased. One should make a careful trade off between output
swing and oscillation suppression. The system was designed first on the system level, and
later implemented on the circuit level.
To optimize the performance, I derived a linear model of the modulator. This model
proved to be accurate enough if the integrator coefficients did not become to large. If the
coefficients do become too large, the system starts to oscillate at a very low frequency,
which is not adequately predicted by the linear model. It is possible that the linear model
can be used for optimization purposes by demanding that the difference in magnitude of
the loop filter at the two highest frequencies were the loop filter becomes 180, remains
large enough. This corresponds to demanding that point two and three on the Nyquist
plot(Fig 3.4) remain at a certain distance from each other. However for this design, the
integrator coefficients were determined using a brute force approach.
On the circuit level, the integrators of the loop filter were implemented using feed
forward op-amps with three stages. The first and second op-amp were designed. The
53
Chapter 6. Conclusion 54
most critical element of the system is the first op-amp as it consumes most of the power.
The third op-amp is very similar to the other two and its power consumption can be
estimated by extrapolation. The total estimated power consumption of the three op-amps
is 1.2 mW. This power consumption is mainly caused by the gain bandwidth specifications,
not by the slewing.
An existing design was reused for the comparator and the flip flops. During simulation
it became clear that the flip flops were the limiting factor on the performance and that
they should be rescaled to offer more equal rise and fall times. A simulation of the system
with ideal flip flops showed a SNR of 74dB.
Bibliography
[1] J. De Maeyer (2005-2006). Efficiente architecturen voor A/D-convertoren in discrete
en continue tijd. Ph.D. thesis, Universiteit Gent.
[2] A. Gelb & W. E. Vander Velde (1968). Multiple-input Describing Functions And
Nonlinear System Design. McGraw-Hill Book Company.
[3] P. Rombouts (2011-2012). Geavanceerd analoog ontwerp. Course notes.
[4] R. Schreier & G. C. Temes (2005). Understanding Delta-Sigma Data Converters. John
Wiley & Sons.
[5] B. D. Vuyst & P. Rombouts (2011). A 5-mhz 11-bit self oscilating Σ∆ modulator with
a delay-based phase shifter in 0.025 mm2. IEEE Journal of Solid-State Circuits, 46(8).
[6] P. Woestyn, P. Rombouts, X. Xing & G. Gielen (2012). A selectable-bandwidth
3.5 mw, 0.03 mm2 self-oscillating sigma delta modulator with 71 db dynamic range
at 5 mhz and 65 db at 10 mhz bandwidth. ANALOG INTEGRATED CIRCUITS
AND SIGNAL PROCESSING, 72(1):55–63. URL http://dx.doi.org/10.1007/
s10470-012-9835-6.
55