picture.iczhiku.com...STANFORD UNIVERSITY Department of Electrical Engineering Prof. Boris Murmann...
Transcript of picture.iczhiku.com...STANFORD UNIVERSITY Department of Electrical Engineering Prof. Boris Murmann...
STANFORD UNIVERSITY Department of Electrical Engineering
Prof. Boris Murmann
EE214: Analog Integrated Circuit Design - Autumn 2006/07 -
http://eeclass.stanford.edu/ee214/
Table of Contents
Introduction 3Lecture 1 Long Channel MOS Model 10Lecture 2 Common Source Amplifier 19Lecture 3 Figures of Merit for Transistors 31Lecture 4 Subthreshold Operation, Short Channel Effects 41Lecture 5 fT and Intrinsic Gain, gm/ID-based Design 53Lecture 6 Extrinsic Capacitances 67Lecture 7 Miller Effect, Zero Value Time Constant Analysis 78Lecture 8 Electronic Noise 90Lecture 9 Electronic Noise (Continued) 103Lecture 10 Body Effect, Common Gate Stage 112Lecture 11 Common Drain Stage 124Lecture 12 Differential Pair 135Lecture 13 Offset Voltage, Current Mirrors 146Lecture 14 Process Variations, Feedback 159Lecture 15 Fully Differential Amplifiers 169Lecture 16 Stability, Feedback Circuit Analysis 178Lecture 17 Loop Gain Simulation 189Lecture 18 Two-Stage OTA 201Lecture 19 Compensation, Noise in Feedback OTAs 210Lecture 20 OTA Design Considerations 221Lecture 21 Step Response 226Lecture 22 Slewing 236Lecture 23 Feedback and Port Impedances, OTA Variants, CMFB
Implementation 246
Lecture 24 Single Ended OTAs, Output Stages 259Lecture 25 Supply Insensitive Biasing 268Lecture 26 Bandgap Reference 278Lecture 27 Bandgap Reference (Continued) 284Lecture 28 Technology Scaling 293Lecture 29 Class Summary 309
1
2
EE 214 IntroductionB. Murmann 1
EE214Analog Integrated Circuit Design
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 IntroductionB. Murmann 2
A Few Words About Your Instructor
• Assistant Professor in EE since 2004
• PhD, UC Berkeley 2003
– Digitally assisted A/D conversion
– Use sloppy analog circuits (low power, fast)
– Correct errors using digital post-processor
• ~ 4 years work experience in IC industry
– Mixed signal IC design, low power, high voltage
• Current research direction
– Digital correction techniques for data converters
– Sensor interfaces
– Circuit design in new technologies• Post-CMOS devices, organic devices
3
EE 214 IntroductionB. Murmann 3
EE214 Basics (1)
• Teaching assistants
– Mohammad Hekmat, Alireza Dastgheib, Abhishek Kamath
– Review Session: Fri 2:15-3:05 PM, Skilling 193 (Live on E4)
• Administrative support
– Ann Guerra, CIS 207
• Lectures are televised
– But please come to class to keep the discussion interactive!
• Web page: http://eeclass.stanford.edu/ee214
– Check regularly, especially FAQ section
– Register for online access to grades and solutions• Only students enrolled can register; we manually control the
access list based on Axess data
EE 214 IntroductionB. Murmann 4
EE214 Basics (2)
• Required text
– Analysis and Design of Analog Integrated Circuits, 4th Edition, Gray, Hurst, Lewis and Meyer, Wiley, 2001. (On reserve in Engineering Library)
• Useful reference
– The Designer’s Guide to Spice and Spectre, Ken Kundert, Kluwer Academic Publishers, 1995. (On reserve in Engineering Library)
• Course prerequisites
– EE101B or equivalent
– Basic device physics and models (PN junctions, MOSFETs, BJTs), basic linear systems (poles, zeros)
– Some exposure to a circuit simulator, basic Unix commands
– May consider concurrent enrollment in EE114X to brush up on the above (primarily for undergraduates)
4
EE 214 IntroductionB. Murmann 5
Assignments
• Homework (20%)
– Handed out on Mondays, due following Monday in class
– Late policy• Score drops 0.5 dB per hour after deadline
– Policy for off-campus students: Fax/email to SCPD before deadline stated on handout
• Midterm Exam (30%)
• Project (20%)
– Design of an amplifier using HSpice (no layout)
– Work in teams of two• OK to discuss with other teams, but no file exchange!
• Final Exam (30%)
EE 214 IntroductionB. Murmann 6
Honor Code
• Please remember you are bound by the honor code
– I will trust you not to cheat
– I will try not to tempt you
• But if you are found cheating it is very serious
– There is a formal hearing
– You can be thrown out of Stanford
• Save yourself and me a huge hassle and be honest
• For more info
– http://www.stanford.edu/dept/vpsa/judicialaffairs/guiding/pdf/honorcode.pdf
5
EE 214 IntroductionB. Murmann 7
Be Reasonable When Asking TAs
• The TAs will not give you "the answer times two"…
• They will also NOT debug your HSpice deck
– Figuring out what's wrong with your circuit is an essential component of this class
EE 214 IntroductionB. Murmann 8
Circuit Simulation
• Primary tools: HSpice/AWaves
– You can use other tools at "own risk"
– Tutorial docs and example simulation decks for HSpice are provided on course web site
• Great alternative to AWaves: Plot HSpice results using Matlab– Download "Hspice Toolbox" at:
http://www-mtl.mit.edu/research/perrottgroup/tools.html#hspice
– This toolbox is installed in course directory (see tutorial for path)
• EE214 Technology
– 0.35µm CMOS
– BSIM3v3 models provided on web site
• First review session (this week) will focus on simulation basics
6
EE 214 IntroductionB. Murmann 9
The Spice Monkey Problem (1)
• What most people know
– Even a very large number of monkeys randomly arranging characters will never manage to write an interesting book
• What some people tend to forget
– Even a very large number of "Spice Monkeys" randomly tweaking circuits will never manage to design a robust, optimized IC
[Courtesy Isaac Martinez]
EE 214 IntroductionB. Murmann 10
The Spice Monkey Problem (2)
• Simply put
– Spice is nothing but a "calculator" that lets you evaluate and test your ideas
– There is no need to simulate anything unless you already know the (approximate) answer!
– Must always be aware of modeling limitations
• Especially in the integrated circuits arena, uneducated, purely simulator driven design can be costly
– Mask sets cost up to $2 Million (90 nm production)
– Turnaround time is on the order of months
– If your chip doesn't work, you cannot simply send the customer a "patch"…
7
EE 214 IntroductionB. Murmann 11
Analysis versus Design
• Unlike common perception, analog circuit analysis and design is not "black magic"
• Circuit analysis
– The art of decomposing a circuit into manageable pieces
– Based on the simplest, but sufficiently accurate models
– One circuit ⇒ one solution
• Circuit design
– The art of synthesizing circuits based on experience from extensive analysis
– One set of specifications ⇒ Many solutions
– Design skills are best acquired through "learning by doing"• This is why we'll have a design project…
EE 214 IntroductionB. Murmann 12
Course Topics
• CMOS technology and device models
• Electronic noise
• Single-stage amplifiers
• Current mirrors, active loads
• Differential pairs
• Operational and transconductance amplifiers
• Feedback, stability and compensation
• Temperature and supply independent biasing
8
EE 214 IntroductionB. Murmann 13
Learning Goals
• Develop deeper understanding of MOS device behavior relevant to analog design
• Develop a feel for limits and tradeoffs in analog circuits (speed, noise, power dissipation)
• Learn to bridge the gap between complex device models/behavior and basic hand calculations
– Design using look-up tables, "gm/ID methodology"
• Develop a systematic, non-spice-monkey design style
• Solidify the above aspects in a hands-on design project
– Design and optimization of a high performance feedback amplifier used in many industrial circuits/applications
9
EE 214 Lecture 1B. Murmann 1
Lecture 1CMOS Technology
Long Channel MOS Model
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 1B. Murmann 2
Overview
• Reading– 2.8 (MOS fabrication), 2.9 (Active MOS devices)– 2.10.1 (Resistors), 2.10.2 (Capacitors)– 1.1, 1.5.0, 1.5.1, 1.5.2, 1.5.3 (Large signal MOS model)
• Introduction– In this first lecture, we will cover some of the background
that positions EE214 as an introductory course on circuit design using CMOS technology. In the lectures to come, we will focus on the problem of amplifier design as a vehicle to establish a set of considerations that apply to more complex circuits and also other technologies. At first, we will review the ‘long channel model’ of a MOS transistor. Driven by circuit examples, we will later augment this simple model to include additional effects that are relevant in practice.
10
EE 214 Lecture 1B. Murmann 3
The Big Picture
• Most modern electronic information processing systems rely on amplification of "small" physical signals– E.g. signal from RF antenna, disk drive head, microphone, …
• EE214 uses amplifiers as a vehicle to teach you the basics of analog integrated circuit analysis and design– Material forms basis for other and/or more complicated circuits
EE 214 Lecture 1B. Murmann 4
Technological Progress
Vacuum Tube1906
ModernCMOS
Transistor1947
Modern DiscreteTransistors
Integrated Circuit1958
11
EE 214 Lecture 1B. Murmann 5
Economics
[European Nanotechnology
Roadmap]
EE 214 Lecture 1B. Murmann 6
Future Applications
12
EE 214 Lecture 1B. Murmann 7
Discrete vs. Integrated Circuits
• Minimize transistor count
• Devices usually don't match
• Arbitrary resistor values
• Capacitors 1pF…10mF
• "Unlimited" number of transistors
• Devices match well
• Keep resistors < 10…100k
• Keep capacitors < 10…50pF
Discrete Audio Amplifier Integrated CMOS Audio Amplifier
EE 214 Lecture 1B. Murmann 8
Modern Integrated Circuit Technologies
• Why use CMOS for analog integrated circuits?
– Low cost, driven by high volume digital ICs
– Integration with high density digital circuits• BiCMOS tends to be expensive
Parameter CMOS Si BJT SiGe BJT
Device Speed High High High
Noise Poor Good Good
Transconductance Poor Good Good
Intrinsic gain Poor Better Best
13
EE 214 Lecture 1B. Murmann 9
Basic MOS Operation (1)
• With zero voltage at the gate, device is "off"
– Back to back reverse biased pn junctions
0V VD (>0V)0V
0V
EE 214 Lecture 1B. Murmann 10
Basic MOS Operation (2)
• With a positive gate bias applied, electrons are pulled toward the positive gate electrode
• Given a large enough bias, the electrons start to "invert" the surface (p→n type), a conductive channel forms
– Magic "threshold voltage" Vt (more later)
>0
14
EE 214 Lecture 1B. Murmann 11
Basic Operation (3)
• If we now apply a positive drain voltage, current will flow
• How can we calculate this current as a function of VGS, VDS?
>0
VDS>0
ID=?
EE 214 Lecture 1B. Murmann 12
Assumptions
1) Current is controlled by the mobile charge in the channel. This is a very good approximation.
2) "Gradual Channel Assumption" - The vertical field sets channel charge, so we can approximate the available mobile charge through the voltage difference between the gate and the channel
3) The last and worst assumption (we will fix it later) is that the carrier velocity is proportional to lateral field (ν = µE). This is equivalent to Ohm's law: velocity (current) is proportional to E-field (voltage)
>0
VDS>0
15
EE 214 Lecture 1B. Murmann 13
First Order IV Characteristics (1)
• What we know:
[ ]tGSoxn VyVVCyQ −−= )()(
WvQI nD ⋅⋅=
Ev ⋅= µ
[ ] WEVyVVCI tGSoxD ⋅⋅⋅−−=∴ µ)(
EE 214 Lecture 1B. Murmann 14
First Order IV Characteristics (2)
dyydVE )(=[ ] WEVyVVCI tGSoxD ⋅⋅⋅−−= µ)(
[ ] dVVyVVCWdyI tGSoxD ⋅−−= )(µ
[ ]∫ ⋅−−=∫DSV
tGSox
L
D dVVyVVCWdyI00
)(µ
( ) DSDS
tGSoxD VV
VVLWCI ⋅
⎥⎥⎦
⎤
⎢⎢⎣
⎡−−=
2µ
• For VDS/2 << VGS-Vt, this looks a lot like a linear resistor: I=1/R × V
• Lets plot this IV relationship...
16
EE 214 Lecture 1B. Murmann 15
Plot of First Order IV Curves
• Something is wrong here...
– Current should never decrease with increasing VDS
• What happens when VDS>VGS-Vt?
– VGD = VGS-VDS becomes less than Vt, i.e. no more channel or "pinch off"
VDSID
VGS-Vt
EE 214 Lecture 1B. Murmann 16
Pinch-Off
• Effective voltage across channel is VGS - Vt
– After channel charge goes to 0, there is a high lateral field that ‘sweeps’ the carriers to the drain, and drops the extra voltage (this is a depletion region of the drain junction)
• To first order, current becomes independent of VDS
N N
– V G S +
+ V DS –
y
y = 0 y = L
Q ( y ) , V ( y ) n
Voltage at the end of channelIs fixed at VGS-Vt
17
EE 214 Lecture 1B. Murmann 17
Modified Plot and Equations
VDS
IDVGS-Vt
Triode Region
Forward Active Region
( ) DSDS
tGSoxD VV
VVLWCI ⋅
⎥⎥⎦
⎤
⎢⎢⎣
⎡−−=
2µTriode Region:
Forward Active: ( ) 2)(21)(
2)(
tGSoxtGStGS
tGSoxD VVLWCVVVVVV
LWCI −=−⋅⎥⎦
⎤⎢⎣⎡ −−−= µµ
EE 214 Lecture 1B. Murmann 18
Model Accuracy
• The above equations constitute the most basic MOS IV model
– "Long channel model", "Quadratic model", "Low field model"
• Unfortunately it doesn't describe modern CMOS devices accurately
– Pushing towards extremely small geometries has resulted in very high electric fields
• Some of the assumptions on slide 12 become invalid
• Other second order dependencies arise
• Nevertheless, we will use this simple model in the first few lectures to develop some basic circuit intuition
– Will fix and refine as we go…
18
EE 214 Lecture 2B. Murmann 1
Lecture 2Common Source Amplifier
Small-Signal Model
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 2B. Murmann 2
Overview
• Reading– 3.0 (Amplifier basics), 3.1 (Model selection)
– 3.3.2 (Common source amplifer)
– 1.6.0 - 1.6.5 (Small signal MOS model)
• Introduction
– In this lecture, we will use the simple long channel MOS model to construct our first amplifier - a common source stage. Looking at its transfer function, we'll find that treating signals as "small" with respect to the bias conditions allows us to linearize the circuit. Next, we generalize this approach and develop a more universal "small signal model" for MOS devices that are biased in the forward active region.
19
EE 214 Lecture 2B. Murmann 3
First-Order MOS Model Summary
( )221
tGSoxD VVL
WCI −≅ µ
Sub-Threshold
(more la
ter...)
Vt VGS
VDS
VGS-Vt
FORWARD ACTIVE
TRIODE
( ) DSDS
tGSoxD VVVVL
WCI ⋅⎥⎦⎤
⎢⎣⎡ −−≅
2µ
≅
≅
"VCCS"
EE 214 Lecture 2B. Murmann 4
One Way to Amplify
• Convert input voltage to current using voltage controlled current source (VCCS)
• Convert back to voltage using a resistor (R)
• "Voltage gain" = ∆Vout/∆Vin
– Product of the V-I and I-V conversion factors
20
EE 214 Lecture 2B. Murmann 5
Common Source Amplifier
• MOS device acts as VCCS
( )221
tioxD VVL
WCI −= µ ( ) RVVL
WCVV tioxDDo ⋅−−= 221 µ
EE 214 Lecture 2B. Murmann 6
Biasing
• Need some sort of "battery" that brings input voltage into useful operating region
• Define VOV=VI-Vt, "quiescent point gate overdrive"– VOV=VGS-Vt with no input signal applied
"Bias"
"Signal"
VI
∆Vo
∆Vi
VO
VOV
21
EE 214 Lecture 2B. Murmann 7
Relationship Between Incremental Voltages
• What is ∆Vo as a function of ∆Vi?
( )
( )[ ][ ]
⎥⎦
⎤⎢⎣
⎡ ∆+∆⋅⋅−=
∆+∆⋅−=
−∆+⋅−=∆
⋅∆+−=∆+
OV
ii
OV
D
iiOVox
OViOVoxo
iOVoxDDoO
VVVR
VI
VVVRL
WC
VVVRL
WCV
RVVL
WCVVV
212
22121
21
2
22
2
µ
µ
µ
• As expected, this is a nonlinear relationship
• Nobody likes nonlinear equations, we need a simpler model
– Fortunately, a linear approximation to the above expression is sufficient for 90% of all analog circuit analysis
EE 214 Lecture 2B. Murmann 8
Small Signal Approximation (1)
• Assuming ∆Vi << 2VOV, we have
⎥⎦
⎤⎢⎣
⎡ ∆+∆⋅⋅−=∆OV
ii
OV
Do V
VVRV
IV2
12
iOV
Do VR
VIV ∆⋅⋅−≅∆ 2
• If we further pretend that the input voltage increment is infinitely small, we can find this result directly by taking the derivative of the large signal transfer function at the "operating point" VI
RV
IdVdV
OV
D
VVi
o
Ii
⋅−==
2
22
EE 214 Lecture 2B. Murmann 9
Small Signal Approximation (2)
• Graphical illustration:
VI
VO
VOV
dVo/dVi
• The slope of the above tangent is the so called "small signal gain" of our amplifier
EE 214 Lecture 2B. Murmann 10
Small Signal MOS Model
• Fortunately we don't have to repeat this analysis for every single circuit we build
• Instead, we derive a linearized circuit model for the MOS transistor and plug it into arbitrary circuits
23
EE 214 Lecture 2B. Murmann 11
Transconductance
• The parameter that relates small signal gate voltage to drain current is called transconductance (gm), or y21 in two-port nomenclature
• The transconductance is found by differentiating the large signal I-V characteristic of the transistor in its operating point
( )221
tGSoxD VVL
WCI −= µ
( ) OVoxtGSoxGS
D
gs
dm V
LWCVV
LWC
VI
vig µµ =−=
∂∂==
OV
Dm V
Ig 2=
EE 214 Lecture 2B. Murmann 12
Additional Model Components
• Now that we've decided to move on using "small signal" approximations, it also becomes easier to refine our model and make it more realistic
• Let's first take a look at "intrinsic gate capacitance"– Intrinsic means that these capacitances are unavoidable and
required for the operation of the device– Note that there are plenty of extrinsic, technology related
capacitances• We'll talk about some of those later
• When talking about gate capacitance, we must distinguish several operating regions– Transistor on
• Triode and forward active regions
– Transistor "off"• Subthreshold operation
24
EE 214 Lecture 2B. Murmann 13
Transistor in Triode Region
• Gate terminal and conductive channel form a parallel plate capacitor across gate oxide CGC= WLεox/tox= WLCox
– We can approximately model this using lumped capacitors of size ½ CGC each from gate-source and gate-drain
• Changing either voltage will change the channel charge
• The depletion capacitance CCB adds extra capacitance from drain and source to substrate
– Usually negligible
L
S D W
G
C GC
C CB
EE 214 Lecture 2B. Murmann 14
Transistor in Forward Active Region
• Assuming a long channel model, if we change the the source voltage in the forward active region
– The voltage difference between the gate and channel at the drain end remains at Vt, but the voltage at the source end changes
– This means that the "bottom plate" of the capacitor does not change uniformly
• Detailed analysis shows that in this case Cgs=2/3WLCox
– See text, section1.6.2
• In the long channel model for forward active operation, the drain voltage does not affect the channel charge
– This means Cgd=0 in the forward active region!• Neglecting second order effects and extrinsic caps, of course
25
EE 214 Lecture 2B. Murmann 15
Transistor Off
• There is no conductive channel
– Gate sees a capacitor to substrate, equivalent to the series combination of the gate oxide capacitor and the depletion capacitance
• If the gate voltage is taken negative, the depletion region shrinks, and the gate-substrate capacitance grows
– With large negative bias, the capacitance approaches CGC
L
S D W
G
C GC
C CB
EE 214 Lecture 2B. Murmann 16
Intrinsic MOS Capacitor Summary
Subthreshold TriodeForward Active
Cgs 0 ½ WLCox2/3 WLCox
Cgd 0 ½ WLCox 0
Cgb 0 01
11−
⎟⎟⎠
⎞⎜⎜⎝
⎛+
oxCB WLCC
WLx
Cd
SiCB
ε=
26
EE 214 Lecture 2B. Murmann 17
Dependence of ID on VDS
• Admittedly, our simple first order MOS I-V equation for the forward active region looks pretty unrealistic
– Ideal current source!
• In reality, the drain current has a weak dependence on VDS
VDS
ID
VGS-Vt
Triode Region
Forward Active Region
Finite dID/dVDS
EE 214 Lecture 2B. Murmann 18
Channel Length Modulation
• "Channel length modulation" is outdated nomenclature for a combination of several physical effects that cause finite dID/dVDS
• The precise dependence of ID on VDS is very hard to model
– You can convince yourself by looking at the BSIM3 manual
• One simple modeling approach was to assume that the large signal current ID increases linearly with VDS ("fudge factor" λ)
)1()(21 2
DStGSoxD VVVL
WCI λµ +−=
• Unfortunately, this simple modified equation has become too inaccurate to predict large signal behavior
– Must rely on simulation for the "final tweak"
– As we will see, this is actually not a big issue…
27
EE 214 Lecture 2B. Murmann 19
Small Signal Output Resistance
• Yet, we must somehow take the finite dID/dVDS into account, since it quantifies how much our forward active device deviates form an ideal current source
• Again, this becomes somewhat easier by limiting ourselves to "small signals"
• Define "small signal output resistance" ro
DS
D
o dVdI
r=1
VDS
ID
Operating Point
Slope = 1/ro
EE 214 Lecture 2B. Murmann 20
Dependence of ro on VDS
• In the above example, it seemed like ro is independent of the VDS operating point (as long as the device is forward active)
• Unfortunately, modern MOS I-V curves can look like this
VDS
ID
OP1
Slope = 1/ro1OP2
Slope = 1/ro2
• Only experimental data or simulation will help us in practice totake VDS dependence into account
28
EE 214 Lecture 2B. Murmann 21
Another ro Caveat
• Is vds really a "small signal"?
• The small signal approximation for ro becomes somewhat inappropriate when the vds swing spans a large fraction of a nonlinear ID-VDS characteristic
• Luckily, in most practical situations, other (well understood) sources of nonlinearity dominate (e.g. transconductance)
OP1
VDS
IDOP
EE 214 Lecture 2B. Murmann 22
Output Resistance - Summary
• The above conclusions on output resistance seem pretty devastating
– Don't really know how to calculate ro, have to trust technology data
– ro depends on operating point, which also means that there is an additional source of nonlinearity
• In practice this has only mild implications
– We'll never design a circuit that relies on a precise value or particular dependence of ro
– In order to be able to reason about the impact of ro in the circuit design process, we will work with typical values and lower bounds from Spice simulations
29
EE 214 Lecture 2B. Murmann 23
1st Order Small Signal Model (Forward Active)
entsn/measurem simulatiofrom followsr
WLC32C
VI2g
o
oxgs
OV
Dm
−
=
=
30
EE 214 Lecture 3B. Murmann 1
Lecture 3Common Source Amplifier Performance
Figures of Merit for Transistors
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 3B. Murmann 2
Overview
• Reading
– 1.6.8 (Transit Frequency)
• Introduction
– Having established some basic modeling tools, we will now look at the achievable performance of our common source stage. We'll derive first order expressions for the amplifier's bandwidth and power dissipation. As we will see, these metrics are proportionally related to more fundamental performance measures of the MOS device itself: Transit Frequency and Transconductor Efficiency.
31
EE 214 Lecture 3B. Murmann 3
Common Source Amplifier Revisited
• Consider a common source stage driven by a "transducer"
• Ideally, what we want to achieve is
– A certain voltage gain
– Large bandwidth, low power dissipation
EE 214 Lecture 3B. Murmann 4
Transfer Function
gsiLom
i
oCsR
RrgsvsvsH
+⋅==1
1)||()()()(
• For simplicity, let's assume that RL is fixed, and RL<< ro
• The desired DC gain (ADC) then dictates our choice for gm
DC gain FrequencyDependence
OVmD
OV
Dm
VgI
VIg
21
2
=
=• Now recall that
L
DCm R
Ag =
32
EE 214 Lecture 3B. Murmann 5
Power Dissipation
• Substituting ID and gm using the expressions on the previous slide yields
OVDCL
DD VAR
VP ⋅⋅=21
• Clearly, minimizing VOV will also minimize power dissipation
– Actually, how about making VOV ≅ 0?• More later…
• Let's now look at the bandwidth of this circuit
• Power dissipation is given by
DDD IVP ⋅=
EE 214 Lecture 3B. Murmann 6
Bandwidth
• The –3dB bandwidth of our circuit is
oxgsgsi
dB WLCCCR 3
2:with13 ==−ω
OVm VLWCoxg µ=
• In order to make this expression more insightful, we can eliminate Cox using
OVDCi
LdB V
LARR ⋅⋅⋅=− 23
123 µω
• It then follows that
33
EE 214 Lecture 3B. Murmann 7
Discussion
• With the assumed fixed design specifications, and a given technology (µ, Lmin) , both power and bandwidth of our circuit are completely determined by the choice of VOV
• Bad news: Making VOV small to save power also means that we lose bandwidth
– This makes intuitive sense since
OVDCL
DD VAR
VP ⋅⋅=21
OVDCi
LdB V
LARR ⋅⋅⋅=− 23
123 µω
OVox
mVC
gLW
µ=
• With gm and L fixed, smaller VOV translates into a bigger (wider) device, and thus larger Cgs
EE 214 Lecture 3B. Murmann 8
Generalization (1)
• What we really want from our MOS transistor
– Lots of gm without investing much current
– Lots of gm without having large Cgs
• To quantify how good of a job our transistor does, we can therefore define the following "figures of merit"
gs
m
D
mCg
Ig and
• Using long channel MOS equations, we find
223and2
LV
Cg
VIg OV
gs
m
OVD
m µ==
• As expected, these equations reflect the same tradeoff that we've observed in the above circuit example
34
EE 214 Lecture 3B. Murmann 9
Generalization (2)
• Part of your job as a designer is to choose VOV such that– You get sufficient bandwidth– And use as little power as possible to accomplish this
• Even though we've come to this graph using a very simple example, the observed tradeoff tends to hold in general– Of course, second order considerations will factor in as you
learn more about circuit design…
VOV
gm/ID
gm/Cgs
EE 214 Lecture 3B. Murmann 10
Product
• In cases where we want to get the "best of both worlds", it is interesting to look at the product of our two figures of merit
VOV
gm/ID
gm/Cgs
gm/ID*gm/Cgs
23LC
gIg
gs
m
D
m µ=⋅
• While this result looks boring, it shows that using smaller channel lengths improves circuit performance
– Either or both speed and power efficiency
35
EE 214 Lecture 3B. Murmann 11
Scaling Impact
• Thanks to "Moore's Law" feature sizes and thus the available minimum channel lengths have been shrinking continuously
– Lmin decreases roughly 2x every 5 years
– Lmin=10µm in 1970, Lmin=90nm in 2004
• From the above discussion, it is clear that we can exploit technology scaling in different ways
– Build faster circuits (gm/Cgs), while keeping power efficiency constant (gm/ID)
• E.g. A/D converter for a disk drive - want to maximize bandwidth/throughput
– Build more power efficient circuits (gm/ID), while keeping the bandwidth constant (gm/Cgs)
• E.g. A/D converter for video signals - bandwidth fixed by a certain standard
EE 214 Lecture 3B. Murmann 12
Transit Frequency (ωT)
• The transit frequency of a transistor is defined as the frequency where the magnitude of the common source current gain (|io/ii|) falls to unity
• Ignoring extrinsic capacitance, it follows that
223
LV
Cg OV
gs
mT
µω ==
• Incidentally, this metric is identical to the figure of merit weconsidered earlier in the context of a CS amplifier…
• Note that we can also express transit frequency in units of Hz
πω2
TTf =
36
EE 214 Lecture 3B. Murmann 13
Transit Frequency Interpretation
• The transit frequency is only useful as a figure of merit in thesense that it quantifies gm/Cgs
• It does not accurately predict up to which frequency you can usethe device
– At high frequencies, many assumptions in our "lumped" transistor model become invalid
– Rule of thumb: Lumped model is good up to about ωT/10
• At higher frequencies, device modeling becomes more challenging and many effects depend on how exactly you layout and connect the device
– These effects are covered in more detail in EE314
– In EE214, we will assume that we operate below ωT/10
EE 214 Lecture 3B. Murmann 14
ωmax
• Can show that
gdgate
TCr
ωω21
max =
• A step into the right direction for quantifying the high frequency capability of a MOSFET is to look at its power gain with gate sheet resistance effects included
– The quantity ωmax is defined as the frequency at which the magnitude of the common source power gain falls to unity
– Also known as "maximum frequency of oscillation"
37
EE 214 Lecture 3B. Murmann 15
Gain
• Let's now return to our common source amplifier example to gain some insight into another performance metric of interest: Small signal DC voltage gain
gsiLom
i
oCsR
RrgsvsvsH
+⋅==1
1)||()()()(
DC gain FrequencyDependence
EE 214 Lecture 3B. Murmann 16
Case 1: RL<< ro
• In this case, the DC gain is
OV
LDLmDC V
RIRgA 2==
• For operation in the forward active region, we definitely need
DDLD VRI <
• Therefore, we have
OV
DDDC V
VA 2<
• Note that the above expression is an upper bound, and cannot quite be achieved in practice
– Nevertheless, this is a useful equation for back of the envelope limit considerations
38
EE 214 Lecture 3B. Murmann 17
Case 2: RL>> ro
• This corresponds to the case where RL is essentially replaced by a current source
• Now we have two biasing problems– Must precisely adjust both VI and ID to set quiescent point of Vo
– We'll solve these kinds of problems when looking at more practical biasing approaches, later in this course….
EE 214 Lecture 3B. Murmann 18
Intrinsic Gain
• With RL→∞, the basic common source stage achieves its maximum possible voltage gain or "intrinsic gain"
omDC rgA =
• As mentioned before, it is nearly impossible to develop a simpleequation that yields ro as a function of the device's bias point in modern technologies
• For the remainder of this course, we therefore regard the combined quantity "gm·ro" as a technology dependent parameter
– Must be characterized through measurements/simulations
• Interestingly, it will turn out that the gain of other, more complicated circuits is fundamentally linked to the intrinsic device gain "gm·ro"
39
EE 214 Lecture 3B. Murmann 19
Figure of Merit Summary
• Transit Frequency
gs
mT C
g=ω
• Transconductor Efficiency
D
mIg
• Intrinsic Gain
omrg
• Note that we can characterize any technology (MOS, BJT, …) with respect to these quantities
– Next lecture we'll look at our "EE214 technology" (0.35µm CMOS)
EE 214 Lecture 3B. Murmann 20
+ Width Independence
• These figures of merit are (to first order) independent of transistor width
– Simplifies technology characterization and enables design in a "normalized space"
WW
Cg
gs
m ∝WW
I
IdV
d
Ig
d
dgs
D
m ∝= W1Wrg om ⋅∝
40
EE 214 Lecture 4B. Murmann 1
Lecture 4EE214 Technology: gm/IDSubthreshold OperationShort Channel Effects
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 4B. Murmann 2
Overview
• Reading
– 1.8 (Weak Inversion)
– 1.7 (Short Channel Effects)
• Introduction
– Last lecture, we found that one important figure of merit for transistors is the transconductor efficiency, gm/ID. Today we'll look at gm/ID in our EE214 0.35µm technology. We will find that additional modeling is needed to explain the behavior of this parameter below and around Vt, and also the deviation from the long channel model in strong inversion. This leads us to an analysis of subthreshold current and an introduction to short channel effects.
41
EE 214 Lecture 4B. Murmann 3
gm/ID Simulation
$ gm/id vs. gate overdrive
$ Boris Murmann, September 2006
.param gs=1
vds d 0 dc 1.5Vvgs g 0 dc 'gs'mn1 d g 0 0 nch214 L=0.35um W=10um
.op
.dc gs 0.4V 1.2V 10mV
.probe ov = par('gs-vth(mn1)')
.probe gm_id = par('gmo(mn1)/i(mn1)')
.options post brief
.lib './ee214_hspice.txt' nominal
.end
EE 214 Lecture 4B. Murmann 4
Result
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.50
5
10
15
20
25
30
35
40
VOV
[V]
gm
/ID
[1/V
]
EE214 technology2/V
OVBJT (q/kT)
42
EE 214 Lecture 4B. Murmann 5
Observations
• Our long channel predication is fairly close for VOV > 150mV
• Unfortunately gm/ID does not approach infinity for VOV → 0
• It also seems that we cannot do better than a BJT, even though the long channel equation would predict that for 0 < VOV < 2kT/q ≅ 52mV at room temperature
• For further analysis, it helps to identify three distinct operating regions
– Strong inversion: VOV > 150mV• Deviations due to short channel effects
– Weak Inversion: VOV < 0 • Behavior similar to a BJT, gm/ID nearly constant
– Moderate Inversion: 0 < VOV < 150mV• Transition region, an interesting mix of the above
EE 214 Lecture 4B. Murmann 6
Weak Inversion
• Questions:
– What determines the current when VOV< 0, i.e. VGS< Vt?
– What is the definition of Vt?
• A closer look at the device current in our previous simulation:
-0.5 0 0.5 10
0.2
0.4
0.6
0.8
1
V [V]
I D [m
A]
-0.5 0 0.5 110
-5
10-4
10-3
10-2
10-1
100
V [V]
I D [m
A]
43
EE 214 Lecture 4B. Murmann 7
Definition of Vt
• Vt is defined as the VGS at which the number of electrons pulled to the surface equals the number of doping atoms
• Seems somewhat arbitrary, but makes sense in terms of surface charge control
EE 214 Lecture 4B. Murmann 8
Mobile Charge versus VOV
• Around Vt, the relationship between mobile charge in the channel and gate voltage becomes linear (Qn ~ Cox[VGS-Vt])
– Exactly what we assumed to derive the long channel model
0.E+00
1.E-07
2.E-07
3.E-07
4.E-07
5.E-07
6.E-07
-1.0 -0.5 0.0 0.5 1.0
VOV [V]
Ch
arg
e [
C]
Fixed Charge
Mobile Charge
Total Charge
44
EE 214 Lecture 4B. Murmann 9
Mobile Charge on a Log Scale
• On a log scale, we see that there are mobile charges before we reach the threshold voltage
– Fundamental result of solid-state physics, not short channels
1.E-16
1.E-15
1.E-14
1.E-13
1.E-12
1.E-11
1.E-10
1.E-09
1.E-08
1.E-07
1.E-06
-1.00 -0.50 0.00 0.50 1.00
VOV [V]
Mo
bil
e C
har
ge
[C]
EE 214 Lecture 4B. Murmann 10
BJT Similarity
• We have– An NPN sandwich, mobile minority carriers in the P region
• This is a BJT!– Except that the base potential is here controlled through a
capacitive divider, and not directly an electrode
45
EE 214 Lecture 4B. Murmann 11
Subthreshold Current
• We know that for a BJT
)//( qkTVSC
BEeII ⋅≅
• In our case we have
)//()(0
qnkTVVD
tGSeII −⋅≅
• n is given by the capacitive divider
ox
js
ox
oxjs
C
C
C
CCn +=
+= 1
where Cjs is the depletion layer capacitance
• In our technology n ≅ 1.5
EE 214 Lecture 4B. Murmann 12
Subthreshold Transconductance
• Similar to BJT, but unfortunately n (≅1.5) times lower
kT
qI
ndV
dIg D
GS
Dm
⋅== 1kT
q
ndV
dI
I
g
GS
D
D
m 1==
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.50
5
10
15
20
25
30
35
40
VOV
[V]
gm
/ID
[1/V
]
EE214 technology2/V
OVBJT (q/kT)1.5x
46
EE 214 Lecture 4B. Murmann 13
Moderate Inversion
• In the transition region between weak and strong inversion, we have two different current mechanisms
dxdn
qkT
dxdnD - (BJT)Diffusion
E - (MOS)Drift
µν
µν
==
=
• Both current components are always present
– Neither one clearly dominates around Vt
• Can show that ratio of drift/diffusion current ~(VGS-Vt)/(kT/q)
– MOS equation becomes dominant at several kT/q
• One way to close the gap between the two regimes is to work with a mathematical fit
– Sometimes useful for computer optimization, not so great for hand analysis…
EE 214 Lecture 4B. Murmann 14
A Curve Fitting Attempt
⎪⎪
⎩
⎪⎪
⎨
⎧
∞→
→
≅
⎟⎠⎞
⎜⎝⎛ ⋅++
=
OVOV
OV
OVD
m
VV
VkT
q
m
mkTqVkT
q
mI
g
;2
0;1
11
122
0 0.1 0.2 0.3 0.4 0.50
5
10
15
20
25
30
35
40
VOV
[V]
gm
/ID
[1/V
]
EE214 technology2/V
OVFitted Equation (m=1.7)
47
EE 214 Lecture 4B. Murmann 15
A Closer Look at Strong Inversion
• Long channel model overestimates gm/ID by roughly 10…20%
– Something worth looking into…
0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.50
5
10
15
VOV
[V]
gm
/ID
[1/V
]
EE214 technology2/V
OV
EE 214 Lecture 4B. Murmann 16
Short Channel Effects
• Velocity saturation due to high lateral field
• Mobility degradation due to high vertical field
• Vt dependence on channel length and width
• ro = f(VDS)
• …
• We will limit the discussion in EE214 to the first two aspects of the above list
– Focus on qualitative understanding, since we will not factor these effects into our hand calculations
48
EE 214 Lecture 4B. Murmann 17
Velocity Saturation (1)
• In lecture 1, we assumed that the carrier velocity is proportional to the lateral E-field, v=µE
• Unfortunately, the speed of carriers in silicon is limited
– At very high fields (high voltage drop across the conductive channel), the carrier velocity saturates
Approximation:
EE
1
µE ν
c+
=
EE 214 Lecture 4B. Murmann 18
Velocity Saturation (2)
• It is important to distinguish various regions in the above plot
– Low field, the long channel equations still hold
– Moderate field, the long channel equations become somewhat inaccurate
– Very high field across the conducting channel – the velocity saturates completely and becomes essentially constant (vscl)
• To get some feel for latter two cases, let's first estimate the E field using simple long channel physics
• In the forward active region, at pinch-off, the lateral field across the channel would be
m
V.
m.
mVe.g.
L
V E OV 610570
350200 ⋅==
µ
49
EE 214 Lecture 4B. Murmann 19
Field Estimates
• In our 0.35µm technology, we have for an NMOS device
m
V.
Vsm
.
sm
101.73
v E
5
sclc
62 1026
0280⋅=
⋅==
µ
• The above example shows that an 0.35µm NMOS device at VOV=200mV does not operate anywhere near Ec
• How about 0.13µm?
m
V.
m.
mV E 61051
130200 ⋅==
µ
• Still not too bad…
EE 214 Lecture 4B. Murmann 20
Short Channel Equation
• Bottom line is that most existing and future MOS analog circuitsare impaired, but not completely limited by velocity saturation
– The digital folks will tell you a different story• Why?
• A simple equation that captures the moderate deviation from the long channel forward active drain current is (see text)
( )OVc
OVcOVox
c
OVOVoxD
VLE
VLEV
L
WC
LEV
VL
WCI
+⋅⋅≅
⎟⎟⎠
⎞⎜⎜⎝
⎛ +⋅≅
µ
µ
21
1
121 2
"Parallel Combination"
50
EE 214 Lecture 4B. Murmann 21
Typical Values for EcL
• As long as VOV is "much less" than these voltages, the above simplified equation holds with reasonable accuracy
• We can use these numbers to check our earlier simulation data for gm/ID. With the correction factor, we have
NMOS PMOS
0.35µm 2 V 8V
0.13µm 0.6V 2.4V
902
2201
12
1
12.
V.V.g.e
LEVVI
g
OVOV
c
OVOVD
m ⋅≅⎟⎠⎞
⎜⎝⎛ +⋅
⎟⎟⎠
⎞⎜⎜⎝
⎛ +⋅≅
• Reasonable agreement with simulation data on slide 15
(Note that this expression is found by using 1.223 and 1.233 in the text, and not from the approximate ID expression on the previous slide)
EE 214 Lecture 4B. Murmann 22
Mobility Degradation due to Vertical Field
• In short channel MOSFETS, the oxide thickness has been continuously scaled down with feature sizes
– 6.5nm in 0.35µm, 2.2nm in 0.13µm technology
• As a result, there is a large vertical electric field that tries to pull the carriers closer to the "dirty" silicon surface
– Imperfections impede movement and thus mobility
• This effect can be included by replacing the mobility term with an "effective mobility"
( ) V.....
VOVeff
140101
=+
≅ θθµµ
• Yet another "fudge factor"
– Possible to lump with EcL parameter
51
EE 214 Lecture 4B. Murmann 23
Summary
• The long channel model does not predict gm/ID with reasonable accuracy in any operating regime
– Accuracy also tends to get worse in newer technology
• Once again, we'll find a way to deal with this in practice
• Simple trick: Change of design variables
– Instead of "thinking", in terms of VOV, we will use gm/ID as a design variable, and not as an unknown that is determined from our choice of VOV
• "gm/ID design methodology" - more later…
• Next lecture, we will first familiarize ourselves with the achievable ft and intrinsic gain in our EE214 technology
52
EE 214 Lecture 5B. Murmann 1
Lecture 5EE214 Technology: fT and Intrinsic Gain
gm/ID-Based Design
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 5B. Murmann 2
Overview
• References
– F. Silveira et. al. "A gm/ID based methodology for the design of CMOS analog circuits and its application to the synthesis of a silicon-on-insulator micropower OTA," IEEE Journal of Solid-State Circuits, Sept. 1996, pp. 1314-1319.
– D. Foty, M. Bucher, D. Binkley, "Re-interpreting the MOS transistor via the inversion coefficient and the continuum of gms/Id," Proc. Int. Conf. on Electronics, Circuits and Systems, pp. 1179-1182, Sept. 2002.
– Denis Flandre's Notes: "Méthodologie gm/ID: un chaînon entre l'analyse symbolique et la synthèse de circuits analogiques basse puissance," available athttp://www.comelec.enst.fr/taisa/Presentations/DenisFlandre.pps
53
EE 214 Lecture 5B. Murmann 3
Overview
• Introduction
– Today, we'll continue to characterize the EE214 technology. The two remaining figures of merit that are of interest to us as circuit designers are fT and intrinsic device gain. In conclusion, we find that VOV is not "directly" related to either performance metric we care about. Hence, we switch towards a strategy called "gm/ID-based design", in which gm/ID, rather than VOV, is used as a central design variable.
EE 214 Lecture 5B. Murmann 4
Performance Metrics of Interest
• Transit Frequency
gs
mT C
g=ω
• Transconductor Efficiency
D
mIg
• Intrinsic Gain
omrg
54
EE 214 Lecture 5B. Murmann 5
fT Simulation
$ ft vs. gate overdrive
$ Boris Murmann, September 2006
.param gs=1
vds d 0 dc 1.5Vvgs g 0 dc 'gs'mn1 d g 0 0 nch214 L=0.35um W=10um
.op
.dc gs 0.4V 1.2V 10mV
.probe ov = par('gs-vth(mn1)')
.probe ft = par('1/2/3.142*gmo(mn1)/(-cgsbo(mn1))')
.options post brief dccap
.lib './ee214_hspice.txt' nominal
.end
EE 214 Lecture 5B. Murmann 6
Result
223
21
LVf OV
Tµ
π=Long Channel:
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.50
5
10
15
20
25
30NMOS W/L=10/0.35
VOV
[V]
f T [G
Hz]
EE214 technologyLong Channel Fit
55
EE 214 Lecture 5B. Murmann 7
Observations - fT
• Again, a simple long channel model doesn't do a very good job
– Large fT discrepancy in weak inversion and in strong inversion, at large VOV
• The reasons for these discrepancies are exactly the same as the ones we came across when looking at gm/ID– Bipolar action in weak and moderate inversion
– Short channel effects at large VOV
• Less gm, hence lower gm/Cgs
• Same conclusion, we won't be able to make good predictions with a simple long channel relationship
EE 214 Lecture 5B. Murmann 8
Another Look at gm/ID·fT
2
321
Lf
Ig
TD
m µπ
=⋅Long Channel:
-0.1 0 0.1 0.2 0.3 0.4 0.50
20
40
60
80
100
120
140
160
NMOS W/L=10/0.35
VOV
[V]
gm
/ID
*fT [G
Hz/
V]
EE214 technologyLong Channel
Short channel effects
Sweet spot (?)
Long channel predicts too much gm/ID
56
EE 214 Lecture 5B. Murmann 9
Intrinsic Gain Simulation
$ gm*ro vs. vds$ Boris Murmann, September 2006
$ vt measured using .op run.param vt1=580m.param vt2=600m
mn1 d g1 0 0 nch214 L=0.35um W=10ummn2 d g2 0 0 nch214 L=0.40um W=10umvg1 g1 0 dc 'vt1+0.2' vg2 g2 0 dc 'vt2+0.2' vd d 0 dc 1
.op
.dc vd 0 3.3V 10mV
.probe av1 = par('gmo(mn1)/gdso(mn1)')
.probe av2 = par('gmo(mn2)/gdso(mn2)')
.options post brief
.lib './ee214_hspice.txt' nominal
.end
EE 214 Lecture 5B. Murmann 10
Intrinsic Gain Plot (VOV=200mV)
0 1 2 30
20
40
60
80
100
120NMOS W=10um
VDS
[V]
gm
*ro
0 0.2 0.4 0.60
20
40
60
80
100
120NMOS W=10um
VDS
[V]
gm
*ro
L=0.35umL=0.4um
L=0.35umL=0.4um
Long Channelgmro=const.
57
EE 214 Lecture 5B. Murmann 11
Observations – Intrinsic Gain
• The gm·ro product shows a strong dependence on VDS bias– Mostly due to varying ro, (see lecture 2, slide 20)
• Also, there is a very smooth transition from triode to somethingclose to forward active behavior– Long channel model would have predicted an abrupt change
to large intrinsic gain at VDS = VOV
– From the plots, we see that we need VDS > 1.5…3 VOV to get appreciable gain
• At high VDS, ro and thus gm·ro decrease due to SCBE (substrate current induced body-effect)– Highly technology dependent, and usually not present in
PMOS devices– If you are interested in more details, please refer to EE316
or a similar course
EE 214 Lecture 5B. Murmann 12
Why care about VOV?
• By now, it should be clear that VOV is not a very useful design variable
– There is no simple expression that accurately links either gm/ID, ft or gm·ro to VOV
– VOV does not even clearly define the onset of forward active operation ("Vdsat")
• The primary variables we care about from a design and performance perspective are gm/ID, ft and gm·ro
– So why not work exclusively with these?
– In case we need a rough estimate for "Vdsat", we can always use Vdsat ≅ 2/(gm/ID)
• E.g. gm/ID=10V-1 ⇒ Vdsat ≅ 200mV
• Let's go back to a design example using a simple CS stage
58
EE 214 Lecture 5B. Murmann 13
Design Example 1
• Given specifications
– DC gain=-2, ID ≤ 2mA, f-3dB=100MHz, CL=10pF
– Make transistor as small as possible
EE 214 Lecture 5B. Murmann 14
Hand Calculations (1)
mS.gRgA
pFMHzR
CRf
mLmDC
LLL
dB
612159
22
15910100
1211
21
3
=Ω
=⇒−=−=
Ω=⋅
=⇒=− ππ
• If we use all the available current, we have
V.
mAmS.
Ig
D
m 1362
612 ==
• We know that in our technology gm·ro ~> 40, even for minimum channel length. Hence, for a small signal gain of 2, the output impedance of the amplifier will be dominated by RL
– OK to use L=Lmin=0.35µm
• We can now find RL and gm simply as
59
EE 214 Lecture 5B. Murmann 15
Hand Calculations (2)
• How about using less current?
• Recall that for fixed transconductance W ~ 1/VOV ~ gm/ID– Using less current means larger gm/ID and thus larger device
• But specifications said to minimize device size
• As a last step, we need to determine the actual device width
– For design implementation and verification with Spice
• Using long channel equations will be very inaccurate
• Solution: Sizing chart
– Plot of ID/W as a function of the design variable gm/ID– Can generate this chart once for several channel lengths
and use it throughout the design process
EE 214 Lecture 5B. Murmann 16
Current Density Chart
0 5 10 15 20 250
10
20
30
40
50
60
70
80
90NMOS L=0.35um
gm
/ID
[1/V]
I D/W
[uA
/um
]
(VDS=1.5V)
60
EE 214 Lecture 5B. Murmann 17
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 2010
0
101
NMOS L=0.35um
gm
/ID
[1/V]
I D/W
[uA
/um
]
A Better Current Density Chart
23µA/µm
6.3 1/V
(VDS=1.5V)
EE 214 Lecture 5B. Murmann 18
Does VDS matter?
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 2010
0
101
NMOS L=0.35um
gm
/ID
[1/V]
I D/W
[uA
/um
]
VDS
=0.5V
VDS
=1.5V
VDS
=2.5VAt gm/ID = 6.3 1/V:
ID/W = 21.8 A/m (VDS=0.5V)
ID/W = 23.3 A/m (VDS=1.5V)
ID/W = 23.6 A/m (VDS=2.5V)
∴ Insignificant dependence on VDS; OK to use a single chart for design (e.g. VDS=1.5V)
61
EE 214 Lecture 5B. Murmann 19
Spice Verification
• So, the device width is
• How to determine VB?
• Adjust VB in Spice until gm/ID=6.3V-1
• Good initial guess– VB ≅ Vt+2/(gm/ID) ≅ 600mV+317mV
• Fortunately, we'll find a way to build circuits without VB
– No need to iterate in practice…
mm
WI
IWD
D µµ 8723
2000 ===
EE 214 Lecture 5B. Murmann 20
DC Operating Point
**** mosfetssubckt element 0:mn1 model 0:nch214 region Saturationid 2.0147mibs 0. ibd 0. vgs 845.0000mvds 1.6797 vbs 0. vth 569.7473mvdsat 201.1145m
beta 60.5664mgam eff 894.1238mgm 12.7994mgds 223.6427u
gmb 2.7928mcdtot 112.7289fcgtot 154.6957fcstot 294.9323f
cbtot 300.7032fcgs 107.6797fcgd 19.7903f
V.
mA.mS.
Ig
D
m 1376012812 ==
62
EE 214 Lecture 5B. Murmann 21
AC Response
100
101
102
0
1
2
3
4
5
6
7
f [MHz]
|vo
/vi|
[dB
]
5.87dB=1.97
EE 214 Lecture 5B. Murmann 22
Observations and Remarks (1)
• The design is essentially "dead on" target!– No need for any Spice tweaking
• We accomplished this by focusing on a performance related parameter (gm/ID) in the design process– Note that strictly speaking, device width is not a design
parameter, since it does not directly relate to any of the specs that we were given!
63
EE 214 Lecture 5B. Murmann 23
Observations and Remarks (2)
• The key advantage of gm/ID based design is that it allows you to transition from hand analysis to Spice without much of the "usual" modeling uncertainties
– Simply because we are incorporating relevant simulation data into the design process
– "Closed loop" design
– Enables you to optimize your circuit without even running a Spice simulation!
• To see why this is good, let's compare with some popular alternatives, as seen in many labs and cubicles around the country…
EE 214 Lecture 5B. Murmann 24
Design Methodology 1 (Worst)
• Have an existing design that somehow works for a different process, different specs, …
• Port this design over and tweak all 137 transistor geometries until I meet the specs…
• 500 Spice runs later, I am approaching the deadline with a design that somehow works, god knows how/why
64
EE 214 Lecture 5B. Murmann 25
Design Methodology 2 (Better)
• Remember the square law transistor model from class
• Go through the pain of estimating µCox, and do some hand calculations
• Plug my design into Spice and realize that everything is about 20…80% off
• Throw away my hand analysis and revert back to the "Spice Monkey" design flow from here
EE 214 Lecture 5B. Murmann 26
Design Methodology 3 (Much Better)
• Spend some time to characterize your technology using Spice– Current density as a function of gm/ID (Sizing Chart)– gm/ID vs. fT for different L– Intrinsic gain for different L– Junction capacitance estimate ("self loading")
• Do hand calculations using the generated technology data– Use MathCAD or Excel script– Can quickly iterate through tens of different designs, if
necessary
• Implement and verify in Spice– Only minor tweaking necessary (if any)– Done!
• More later…
65
EE 214 Lecture 5B. Murmann 27
gm/ID Considerations
• In case there is no "hard constraint" on gm/ID, e.g. via ωT or swing, a good starting point is to bias the device around the "technology hot spot" (highest gm/ID·fT)
0 5 10 15 20 25
20
40
60
80
100
120
140NMOS L=0.35µm
gm
/ID
[1/V]
gm
/ID
*fT [G
Hz/
V]
66
EE 214 Lecture 6B. Murmann 1
Lecture 6Design Example 2
Extrinsic Capacitance
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 6B. Murmann 2
Overview
• Reading
– 1.6.7 (Parasitic Elements)
• Introduction
– In today's lecture, we'll look at another CS amplifier design example – this time with a more realistic input source that has finite resistance. Through this example, we find that we need more modeling to accurately predict the resulting pole at the gate node. Our discussion leads to a discussion of parasitic extrinsic capacitors around the MOSFET - overlap and junction capacitance.
67
EE 214 Lecture 6B. Murmann 3
Design Example 2
• Given specifications
– DC gain=-4, ID ≤ 0.5mA
– RL=1k, Ri=10k
– Maximize and estimate bandwidth
gsiLom
i
oCsR1
1)R||r(g)s(v)s(v)s(H
+⋅−==
DC gain FrequencyDependence
EE 214 Lecture 6B. Murmann 4
Hand Calculation
• Just as in the previous design example, we know thatgm·ro >> ADC. Hence we simply find
mSk
g
RgA
m
LmDC
41
44
=Ω
=
−=−=
• In order to maximize bandwidth, we need Cgs as small as possible. Again, this is the case for using up all the availablecurrent, i.e. minimum gm/ID
VmA.mS
Ig
D
m 1850
4 ==
• In order to estimate the achieved bandwidth, we can now find Cgs using
?mS
fgC
T
mgs
421
21
ππ==
68
EE 214 Lecture 6B. Murmann 5
Transit Frequency Chart
0 5 10 15 20 250
5
10
15
20
25
30NMOS L=0.35um
gm
/ID
[1/V]
f T [G
Hz] 16GHz
EE 214 Lecture 6B. Murmann 6
Bandwidth
• Using the transit frequency chart, we find
fFGHzmS
fgC
T
mgs 40
164
21
21 ===
ππ
MHzfFkCR
fgsi
dB 39840101
211
21
3 =⋅
==− ππ
• As a last step, we'll now verify the design using Spice
– Use current density chart to find W
69
EE 214 Lecture 6B. Murmann 7
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 2010
0
101
NMOS L=0.35um
gm
/ID
[1/V]
I D/W
[uA
/um
]
Current Density Chart
15.5µA/µm
EE 214 Lecture 6B. Murmann 8
Spice Verification
• Device width
• Simulation circuit
m.
WI
IWD
D µ32515
500 ===
70
EE 214 Lecture 6B. Murmann 9
DC Operating Point
**** mosfetssubckt element 0:mn1 model 0:nch214 region Saturationid 507.0226uibs 0. ibd 0. vgs 795.0000mvds 1.4930 vbs 0. vth 572.5882mvdsat 172.3503m
beta 22.1349mgam eff 894.1238mgm 4.0656mgds 70.2071u
gmb 897.2679ucdtot 42.9622fcgtot 56.5567fcstot 108.8000f
cbtot 112.7236fcgs 39.3251fcgd 7.2419f
V.
AmS.
Ig
D
m 10285070664 ==
µfF.Cgs 339= Good Agreement.
EE 214 Lecture 6B. Murmann 10
Frequency Response
• Simulation result: f-3dB=180MHz. Hand analysis: f-3dB=398MHz !
– Time to take a look at extrinsic capacitance…
100
101
102
0
2
4
6
8
10
12
14
f [MHz]
|vo
/vi|
[dB
]
3dB
11.6dB=3.8
71
EE 214 Lecture 6B. Murmann 11
Extrinsic Capacitance
• Overlap capacitance– Gate to source and gate to drain
• Junction capacitance– Source to bulk and drain to bulk
EE 214 Lecture 6B. Murmann 12
Overlap Capacitance
• Two components
– Direct overlap ~ CoxWLoverlap
– Additional component due to fringing fields• Non-negligible in modern technology (poly thickness large
compared to other feature sizes)
• EE214 technology parameters
– NMOS: Col= 0.23fF/µm
– PMOS: Col= 0.48fF/µm
72
EE 214 Lecture 6B. Murmann 13
Junction Capacitance
• Two components
– Area (AS, AD) and Perimeter (PS, PD)
mjswDB
jswmj
DB
jjdb
PBV1
CPS
PBV1
CASC
⎟⎠⎞
⎜⎝⎛ +
⋅+
⎟⎠⎞
⎜⎝⎛ +
⋅=
EE214 Technology
Cj Cjsw mj, mjsw PB
NMOS 0.85 fF/µm2 0.49 fF/µm 0.39 0.51V
PMOS 1.11 fF/µm2 0.48 fF/µm 0.48 0.93V
• Tedious to hand-calculate– In practice, we'll work with estimates and leave accurate
calculations for the final Spice tweak
• HSpice automatically calculates junction capacitance based on W and a geometry factor
– LTSpice does not (must specify AS, AD, PS, PD)
EE 214 Lecture 6B. Murmann 14
Layout Dependence
73
EE 214 Lecture 6B. Murmann 15
MOS Capacitor Summary
Subthreshold TriodeForward Active
Cgs Col ½ WLCox+ Col2/3 WLCox + Col
Cgd Col ½ WLCox+Col Col
Cgb ≅ small* ≅ small*
Csb Cjsb Cjsb+ CCB/2 Cjsb+ 2/3CCB
Cdb Cjdb Cjdb+ CCB/2 Cjdb
111
−
⎟⎟⎠
⎞⎜⎜⎝
⎛+
oxCB WLCC
* Channel is not a perfect "shield," hence there is some finite Cgb
EE 214 Lecture 6B. Murmann 16
Capacitance Simulation
• Gradual transition from weak inversion to forward active, and forward active to triode!
0 0.5 1 1.5 2 2.5 3 3.50
2
4
6
8
10
12
14
NMOS W/L=10/0.35, VDS
=0.5V
VGS
[V]
Ca
pa
cita
nce
[fF
]
Cgs
Cgd
Cgb
74
EE 214 Lecture 6B. Murmann 17
Does Junction Capacitance Matter?
• Let's plug in some numbers…– Using the data on slide 13, W=10µm and VDB=1.5V we find
• Cjdb,nmos = 11.3fF• Cjdb,pmos = 13.7fF
• How does this compare to Cgs?– In our technology, Cox=5.3aF/µm2, with W= 10µm and
L=0.35µm, we find• Cgs,nmos = 15fF• Cgs,pmos = 17fF
• Rule of thumb– Junction capacitance is comparable to Cgs!
• Cj ~ 0.8 Cgs (for minimum L at 1.5V bias)
• Unfortunately, this means that junction capacitance will be significant in many circuits…
EE 214 Lecture 6B. Murmann 18
PMOS Well Capacitance
• In the EE214 (N-well) technology, the PMOS transistor is a 5 terminal device!– G, D, S, B, Substrate
• N-well forms a PN junction with Substrate– Often "AC shorted" when N-well=VDD, Substrate=GND– Not shorted when we connect N-well to source!
• Resulting capacitance ~ 0.1fF/µm2
• Not modeled in Spice! Must add extra diode manually in this case
75
EE 214 Lecture 6B. Murmann 19
Complete Small Signal Model (Forward Active)
EE 214 Lecture 6B. Murmann 20
HSpice .OP Output Variables
cdtot 42.9622f
cgtot 56.5567f
cstot 108.8000f
cbtot 112.7236f
cgs 39.3251f
cgd 7.2419f
cdtot ≡ Cgd + Cdb
cgtot ≡ Cgs + Cgd + Cgb
cstot ≡ Cgs + Csb
cbtot ≡ Cgb + Csb+ Cdb
cgs ≡ Cgs
cgd ≡ Cgd
HSpice (.OP) Corresponding Small SignalModel Elements
76
EE 214 Lecture 6B. Murmann 21
+ A Note on Transcapacitance
• It turns out that this complicated model still doesn't describe the parasitics with 100% accuracy
– Since MOSFETS are 4-terminal devices, we are really dealing with a "4-terminal capacitor" and not with a network of simple two-terminal capacitors
• In practice, such "transcapacitance" effects are rarely relevantfor design & hand analysis
– Model based on two-terminal capacitors is usually good to within a few percent
– We'll simply ignore transcapacitance and use Spice as a final check to see if this was OK…
• In case you are curious about more details, please refer to section "Introducing Transcapacitance" in the HSpice manual
77
EE 214 Lecture 7B. Murmann 1
Lecture 7Design Example 2 (Continued)
Zero-Value Time Constant Analysis
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 7B. Murmann 2
Overview
• Reading– 7.1, 7.2.0, 7.2.1 (Miller Effect in CS Stage, only pp. 488-493)– 7.3.0, 7.3.1 7.3.2 (Zero-Value Time Constant Analysis)– 7.3.3 (Cascade Amplifier Frequency Response) – Supplementary document "Bandwidth estimation techniques," by
Tom Lee (optional, see website).
• Introduction– Last lecture, we found that using a simple circuit model based on
only intrinsic capacitance is not sufficient for accurate bandwidth prediction in our CS stage. Having learned about the involved extrinsic capacitances, we are now in a position to improve our hand analysis and match the Spice result with good precision. Tosimplify the analysis, we will utilize the so-called "Miller Approximation". Next, we will take at look at the the "Zero-Value Time Constant Analysis" as an alternative method, which is useful for a much broader class of circuits, beyond simple common source stages.
78
EE 214 Lecture 7B. Murmann 3
Design Example 2 Revisited
?)s(H =
EE 214 Lecture 7B. Murmann 4
Bandwidth Estimation (1)
• To simplify the problem, let's first neglect ro and Cdb
– We'll need to check later is the latter assumption was OK
• Next, we cut the circuit as shown below and calculate the equivalent admittance seen looking into the right side of the cut
)s(v)s(i)s(Y
gs
=
79
EE 214 Lecture 7B. Murmann 5
Bandwidth Estimation (2)
( ) ( ) 0=⋅−++⋅−= gdgsoL
ogsmgdogs sCvv
RvvgsCvvi
( )⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
+
−−=⋅−==∴
gdL
m
gd
Lmvgdvgs CsR
gC
sRg)s(AwithsC)s(A
)s(v)s(i)s(Y
1
11
EE 214 Lecture 7B. Murmann 6
Miller Approximation
• Assuming that the poles and zeros in Av(s) occur at much higher frequencies than the bandwidth we are trying to estimate, it is OK to replace Av(s) with its DC value
– This is known as the "Miller approximation"
– It is always a good idea to check (later) if the approximation was indeed valid!
• With the Miller approximation, we have
( ) gdLm sCRg)s(Y ⋅+≅ 1
• This is the same as a capacitor to GND with a value of (1+gmRL) times Cgd
– "Amplified capacitance"
80
EE 214 Lecture 7B. Murmann 7
Generalization
• Interesting cases
– Av=0 ⇒ Zin=Z (no surprise…)
– Av=1 ⇒ Zin=∞• "Bootstrapping"
– Av>1, e.g. Av=2 ⇒ Zin=-Z (negative!)
– Av<0, ⇒ Zin=Z/(1+|Av|)• Impedance reduction
( )vin
vtestvtest
test
test
testin
AYY
AZ
ZvAv
vivZ
−=
−=−==
1
1
EE 214 Lecture 7B. Murmann 8
Modified Input Network
• Very simple!
– At least much simpler than using exact expressions• See e.g. equation 7.19 in the text
• Next, we'll verify if the involved assumptions hold in our example circuit, and also see how accurately we can match Spice
– Need estimates of Cgd and Cgb
( )[ ]gdLmgbgsidB CRgCCR
f⋅+++
≅− 11
21
3 π
81
EE 214 Lecture 7B. Murmann 9
Extrinsic Capacitance Estimates
• Estimating Cgd is simple:fF.fF.
mfF.WCgd 4723032230 =⋅=⋅=µ
• How about Cgb?
• We can include Cgb by using a modified transit frequency chart
• Previously, we used
gs
mT C
gfπ21=
• Taking extrinsic capacitances into account, we can redefine
gg
m
gdgbgs
mT C
gCCC
gfππ 21
21 =
++=
• The chart on the following slide was generated with the Spice deck shown on slide 5, lecture 5, with the following modified line – .probe ft = par('1/2/3.142*gmo(mn1)/(-cggbo(mn1))')
EE 214 Lecture 7B. Murmann 10
Modified Transit Frequency Chart
80 5 10 15 20 250
5
10
15
20
25
30NMOS L=0.35um
gm
/ID
[1/V]
f T [G
Hz]
Improved estimate using Cgg
Previous estimate
11.25
82
EE 214 Lecture 7B. Murmann 11
Improved Bandwidth Estimate
• Using the transit frequency chart, we find
fFGHz.
mSfgC
T
mgg 57
25114
21
21 ===
ππ
( )[ ]
[ ]
[ ] MHzfF.fFk
CRgCR
CRgCCRf
gdLmggi
gdLmgbgsidB
1844745710
121
121
11
21
3
=⋅+Ω
=
+=
+++≅−
π
π
π
• Our simulation result from last lecture was
MHzf Spice,dB 1803 =−
• Pretty good! (2% error)
EE 214 Lecture 7B. Murmann 12
Assumption Check (1)
• It is interesting (and necessary in general) to check how good the Miller assumption was in this analysis
• We assumed that
LmgdL
m
gd
Lmv RgCsRgC
sRg)s(A −≅
⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
+
−−=
1
1
up to the frequency of interest (~180MHz)
• Let's check this by calculating the magnitudes of the pole and zero in Av(s)
GHzfF.
mSCg
GHz.fF.kCR
gd
m
gdL
8647
421
21
521471
1211
21
==
=⋅Ω
=
ππ
ππ
83
EE 214 Lecture 7B. Murmann 13
Assumption Check (2)
• We had also assumed that the impact of Cdb is negligible
– How can we make sure this is OK?
• One possible solution: Re-derive Av(s) with Cdb present
Cgs+Cgbgmvgs
+vgs-
+vo-
RL
Ri
vi
Cgd
Y(s)i(s)
Cdb
( ) 00 =⋅+⋅−++ dbgdgsoL
ogsm sCvsCvv
Rvvg
)s(v)s(v)s(A
gs
ov =
EE 214 Lecture 7B. Murmann 14
Assumption Check (3)
• Zero is unchanged, but pole frequency is lowered. From our Spice output, we know that Cdb=35.8fF, and hence
⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
++
−−==∴
)CC(sRgC
sRg
)s(v)s(v)s(A
dbgdL
m
gd
Lmgs
ov 1
1
GHz.)fF.fF.(k)CC(R dbgdL
73835471
1211
21 =
+⋅Ω=
+ ππ
• Still not too bad, since 3.7GHz >> 180MHz
– But what if we now add some load capacitance at the output of the amplifier?
• Will appear in parallel with Cdb
84
EE 214 Lecture 7B. Murmann 15
Effect of Output Loading (1)
⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
+++
−−==
)CCC(sRgC
sRg
)s(v)s(v)s(A
LdbgdL
m
gd
Lmgs
ov 1
1
• Suppose CL=10pF, then
! MHzMHzpFk)CCC(R LdbgdL
180161011
211
21 <<=
⋅Ω≅
++ ππ
EE 214 Lecture 7B. Murmann 16
Effect of Output Loading (2)
• Obviously, with the added CL=10pF, the Miller approximation does not hold any more
• So what is going on in this modified circuit, with large CL?– Cgd is "amplified" only at low frequencies, before CL
"destroys" the gain from vgs to vo
– The pole caused by Cgd no longer impairs the 3-dB bandwidth of the circuit
– Intuitively, the bandwidth must now be somehow set by CL
• After all, it destroys the gain from vgs to vo, which implies that the gain from vi to vo must also roll off
• With very large CL, it is intuitively clear what the approximate bandwidth of the circuit should be– But how about other cases and circuits that don't have a
straightforward Miller approximation?
85
EE 214 Lecture 7B. Murmann 17
Zero-Value Time Constant Analysis
• Fortunately, there is a more general method that allows us to estimate the bandwidth of arbitrary circuits (within limits)
– Without going though the pain of deriving the complete s-domain transfer function
• "ZVTC Analysis," or "Open Circuit Time Constant Analysis"
• Here's how it works
– Remove all but one capacitor. Short independent voltage sources, remove independent current sources
– Calculate resistance seen by capacitor and compute τj=RjoCj
– Repeat for all capacitors in the circuit
– Sum all time constants and calculate bandwidth estimate
∑≅−
jdB τ
ω 13
EE 214 Lecture 7B. Murmann 18
Example (1)
• Step 1:
11 CRi=τ
86
EE 214 Lecture 7B. Murmann 19
Example (2)
iLmLitest
testitestmLitest
test
testgsmLgs
test
testo RRgRR
i)iRig(RRi
i)ivg(Rv
ivR ++=++=
++==2
• Step 2:
• Step 3:
33 CRL=τ
22 C)RRgRR( iLmLi ++=τ
Lo RR =3
EE 214 Lecture 7B. Murmann 20
Bandwidth Estimate Based on ZVTC
• Reality check using numbers from design example 2
3213213
1211
21
CRC)RRgRR(CRf
LiLmLiidB ++++
=++
≅∴ − πτττπ
fF.CCfF.CC
fF.CCC
db
gd
gbgs
835
47
649
3
2
1
==
==
=+=
psfF.kpsfF.)kk(
psfF.k
368351377471105
49664910
3
2
1
=⋅Ω==⋅Ω+Ω⋅=
=⋅Ω=
τττ
MHzf dB 175121
3213 =
++≅∴ − τττπ
• Not bad!
– Spice simulation gave 180MHz
– Miller approximation result was 184MHz
87
EE 214 Lecture 7B. Murmann 21
Inclusion of CL
• What happens if again, we consider adding a large load capacitance (CL)?
pFCCCfF.CC
fF.CCC
Ldb
gd
gbgs
10
47
649
3
2
1
≅+=
==
=+=
ps,pFkpsfF.)kk(
psfF.k
00010101377471105
49664910
3
2
1
=⋅Ω==⋅Ω+Ω⋅=
=⋅Ω=
τττ
MHz.f dB 614121
3213 =
++≅∴ − τττπ
• Now the third time constant dominates and significantly reduces our bandwidth estimate
• Looks like this is a powerful method
– Miller effect is taken care of, output loading effect is included
• Most importantly though, the method provides us with insight about the limiting elements in our circuit!
EE 214 Lecture 7B. Murmann 22
How ZVTCs Work
• Intuition
– Each time constant relates to the bandwidth that we would get if no other capacitors were present
• "Local bandwidth bottleneck"
– For simplicity, the ZVTC method linearly combines the local bottlenecks to estimate the overall bandwidth
• Mathematically, the ZVTC method is based on the approximation
11 111
1 +≅
++++= −
− sbK
sb...sbsbK)s(H n
nn
n• Can show that
– b1 corresponds to the sum of all time constants in the circuit
– This approximation is OK unless there are "undamped" complex poles, or several limiting poles with comparable magnitude
88
EE 214 Lecture 7B. Murmann 23
A Simple Example
• Roughly -22% error
• ZVTC estimates tend to be conservative
– Actual bandwidth will almost always be at least as high as estimate
RC.
RC.
RC dBdB506440121
33 ==−⋅= −− ωω
Exact bandwidth: ZVTC Method:
EE 214 Lecture 7B. Murmann 24
ZVTC Accuracy and Other Caveats
• Accuracy tends to be OK when there is a single dominant pole– Not surprising, since the "approximation" shown on slide 22
makes no error for a single pole system– Fortunately, many practical circuits indeed have a somewhat
dominant pole
• Some elements, like AC coupling caps, must be eliminated before applying the ZVTC method– These caps are meant to be shorts at high frequencies, and
do not degrade the signal bandwidth– Can use method of "short circuit time constants" to
determine coupling cap sizes• See supplementary document by Tom Lee
• The ZVTC method tells us nothing about zeros in the transfer function!
89
EE 214 Lecture 8B. Murmann 1
Lecture 8Electronic Noise
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 8B. Murmann 2
Overview
• Reading
– 11.1 (Noise Introduction)
– 11.2.2 (Thermal Noise)
– 11.3.3 (MOS Transistor Noise)
– Supplementary Handout: "Introduction to Noise," by Daniel Cooley
• Introduction
– Electronic noise is a significant and fundamental issue in the design of high performance analog circuits. The noise level of a circuit affects the "fidelity" or accuracy of the signals that are beingprocessed. As we shall see, minimizing electronic noise is costly; there exists a steep tradeoff with power dissipation and bandwidth. Today's lecture will provide an introduction to electronic noise at the component level (resistors and MOSFETs). We will use these results in the remainder of the course to analyze the impact of noise in various circuits.
90
EE 214 Lecture 8B. Murmann 3
Types of Noise
• "Man made noise", interference noise
– Signal coupling
– Substrate coupling
– Finite power supply rejection
– Solutions• Fully differential circuits
• Layout techniques
• "Electronic noise" or "device noise" (focus of this lecture)
– Fundamental• E.g. "thermal noise" caused by random motion of carriers
– Technology related• "Flicker noise" caused by material defects and "roughness"
EE 214 Lecture 8B. Murmann 4
Significance of Electronic Noise (1)
"Signal" "Noise"
noise
2signal
noise
signal
P
V
P
PSNR ∝=
Signal-to-Noise Ratio
91
EE 214 Lecture 8B. Murmann 5
Significance of Electronic Noise (2)
• Example: Noisy image
http://www.soe.ucsc.edu/~htakeda/kernelreg/kernelreg.htm
EE 214 Lecture 8B. Murmann 6
Significance of Electronic Noise (3)
• The "fidelity" of electronic systems of often determined by their SNR
– Examples• Audio systems
• Imagers, cameras
• Wireless and wireline transceivers
• Electronic noise directly trades with power dissipation and speed
– In most circuits, low noise dictates large capacitors (and/or small R, large gm), which means high power dissipation
• Noise has become increasingly important in modern technologies with reduced supply voltages
– SNR ~ Vsignal2/Pnoise ~ VDD
2/Pnoise
• Designing a low power, high performance circuit requires good understanding of electronic noise!
92
EE 214 Lecture 8B. Murmann 7
Topics/Questions
• How to model noise of circuit components
• How to calculate/simulate noise performance of a complete circuit (e.g. SNR)
• Do we need to worry about electronic noise in all circuits?
– The answer must be no; digital folks never talk about this…
– Need to get a feel for situations where noise matters
EE 214 Lecture 8B. Murmann 8
Ideal Resistor
• Constant current, independent of time
• Non-physical
– In a physical resistor, carriers "randomly" collide with latticeatoms, giving rise to small current variations over time
i(t)
1V/1kΩ
93
EE 214 Lecture 8B. Murmann 9
Physical Resistor
• "Thermal Noise" or "Johnson Noise"– J.B. Johnson, "Thermal Agitation of Electricity in Conductors,"
Phys. Rev., pp. 97-109, July 1928.
• Can model random current component e.g. using a noise current source in(t)
i(t)
1V/1kΩ
EE 214 Lecture 8B. Murmann 10
Properties of Thermal Noise
• Present in any conductor
• Independent of DC current flow
• Instantaneous noise value is unpredictable since it is a result of a large number of random, superimposed collisions with relaxation time constants of τ ≅ 0.17ps
– Consequences:• Gaussian amplitude distribution
• Knowing in(t) does not help predict in(t+∆t), unless ∆t is on the order of 0.17ps (cannot sample signals this fast)
• The power generated by thermal noise is spread up to very high frequencies (1/τ ≅ 6,000Grad/s)
• The only predictable property of thermal noise is its average power!
94
EE 214 Lecture 8B. Murmann 11
Average Power
• For a deterministic current signal with period T, the average power is given by
( )∫−
⋅⋅=2/T
2/T
2av dtRti
T
1P
• This definition can be extended to capture non-deterministic random signals
– Assuming a real, stationary and ergodic random process
( )∫−
∞→⋅⋅=
2/T
2/T
2n
Tn dtRti
T
1limP
• For notational convenience, we typically drop R in the above expression and work with "mean square" current (or voltages)
( )∫−
∞→⋅=
2/T
2/T
2n
T
2n dtti
T
1limi
EE 214 Lecture 8B. Murmann 12
Thermal Noise Spectrum
• The so-called power spectral density (PSD) shows how much power a signal caries at a particular frequency
• In the case of thermal noise, the power is spread uniformly up to very high frequencies (about 10% drop at 2,000GHz)
PSD(f)
f
n0
• The total average noise power Pn in a particular frequency band can be found by integrating the PSD
( )∫ ⋅=2
1
f
f
n dffPSDP
95
EE 214 Lecture 8B. Murmann 13
Thermal Noise Power
• Nyquist showed that the noise PSD of a resistor is
( ) kT4nfPSD 0 ⋅==
• k is the Boltzmann constant and T is the absolute temperature
– 4kT = 1.66·10-20 Joules at room temperature
• The total average noise power of a resistor in a certain frequency band is therefore
( ) fkT4ffkT4dfkT4P 12
f
f
n
2
1
∆⋅=−⋅=⋅= ∫
EE 214 Lecture 8B. Murmann 14
Equivalent Noise Generators
• Can model the noise using either an equivalent voltage or current generator
fR
1kT4
R
Pi n2n ∆⋅⋅==fRkT4RPv n
2n ∆⋅⋅=⋅=
V4vMHz1f
Hz/nV4f
v
Hz
V1016
f
v
:1kΩR For
2n
2n
218
2n
µ∆
∆
∆
=⇒=
=
⋅=
=
−
nA4iMHz1f
Hz/pA4f
i
Hz
A1016
f
i
:1kΩR For
2n
2n
224
2n
=⇒=
=
⋅=
=
−
∆
∆
∆
96
EE 214 Lecture 8B. Murmann 15
Two Resistors in Series
( ) fRRkT4vvv 2122n
21n
2n ∆⋅+⋅⋅=+=
• Always remember to add independent noise sources using mean squared quantities
– Never add RMS values!
( ) 2n1n22n
21n
2
2n1n2n vv2vvvvv ⋅⋅−+=−=
• Since vn1(t) and vn2(t) are statistically independent, we have
EE 214 Lecture 8B. Murmann 16
MOSFET Thermal Noise (1)
• As one would expect, the noise of a MOSFET operating in the triode region is equal to that of a resistor
• In the forward active region, the thermal noise of a MOSFET can be modeled using a drain current source with spectral density
fgkT4i m2d ∆γ ⋅⋅⋅=
• For a long channel MOSFET γ=2/3
• For the past ten years, researchers have been debating over the value of γ in short channels
– Preliminary (wrong) results had suggested that in short channels γ can be as high as 2…5 due to hot carrier effects
97
EE 214 Lecture 8B. Murmann 17
MOSFET Thermal Noise (2)
• Fortunately, these discussions have come to an end with the conclusion that short channels have γ ≅ 1– A.J. Scholten et al., "Noise modeling for RF CMOS circuit
simulation," IEEE Trans. Electron Devices, pp. 618-632, March 2003.
EE 214 Lecture 8B. Murmann 18
Spice Simulation (1)
$ EE214 MOS device noise simulation
$ Boris Murmann, September 2006
vd dd 0 1.5vm dd d 0
vg g 0 dc 0.8 ac 1mn1 d g 0 0 nch214 L=0.35u W=10uh1 c 0 ccvs vm 1
.op
.ac dec 100 10k 1gig
.noise v(c) vg
.options post brief
.lib './ee214_hspice.txt' nominal
.end
dd1.5V
d
0V c
CCVS1V/A
98
EE 214 Lecture 8B. Murmann 19
10-2
10-1
100
101
102
103
10-24
10-23
10-22
f [MHz]
avg
(i d 2)/
df
[A2 /H
z]
HSpice ("outnoise")4kT*2/3*g
m
Spice Simulation (2)
???
(gm=1.28mS)
EE 214 Lecture 8B. Murmann 20
1/f Noise
• Also called "flicker noise" or "pink noise"
• Caused by traps near Si/SiO2 interface that randomly capture and release carriers
• Occurs in virtually any device, but is most pronounced in MOSFETS
• One (empirical) way to model flicker noise:– Known as "NLEV=2" HSpice model
• For other models, see HSpice manual or– D. Xie et al, "SPICE Models for Flicker Noise in n-MOSFETs from
Subthreshold to Strong Inversion," IEEE Trans. CAD, pp. 1293-1303, Nov. 2000
• Kf is strongly dependent on technology; numbers for EE214 0.35µm CMOS technology:– Kf,NMOS = 0.5·10-25 V2F– Kf,PMOS = 0.25·10-25 V2F
f
f
LW
g
C
Ki
2m
ox
f2f/1
∆⋅
=
99
EE 214 Lecture 8B. Murmann 21
1/f Noise Corner
• By definition, the frequency at which the flicker noise density equals the thermal noise density
fgkT4f
f
LW
g
C
Km
co
2m
ox
f ∆γ∆ ⋅⋅=⋅
⎟⎠
⎞⎜⎝
⎛⎟⎟⎠
⎞⎜⎜⎝
⎛=
⋅=⇒
W
I
I
g
L
1
C
1
kT4
K
LW
g
C
1
kT4
Kf
D
D
m
ox
f
m
ox
fco
γ
γ
• Example: NMOS, L=0.35µm, gm/ID=10V-1 → ID/W=10A/m
⇒ fco = 244kHz
• In more recent technologies, 1/f corner frequencies can be on the order of 10MHz
EE 214 Lecture 8B. Murmann 22
1/f Noise Contribution (1)
• Just as with white noise, the total 1/f noise contribution is found by integrating its spectral density
⎟⎟⎠
⎞⎜⎜⎝
⎛⋅
=⎟⎟⎠
⎞⎜⎜⎝
⎛⋅
=
⋅= ∫
1
2m
ox
f
1
2m
ox
f
f
f
m
ox
f2tot,f/1
f
flog3.2
LW
g
C
1
kT4
K
f
fln
LW
g
C
1
kT4
K
f
f
LW
g
C
1
kT4
Ki
2
1
γγ
∆γ
• Total flicker noise depends on number of frequency decades
– Same flicker noise power in 1…10 Hz as in 1…10 GHz
– RMS noise proportional to sqrt(# of frequency decades)
• So, does flicker noise matter?
– Look at noise integral to see its relative contribution to totaldrain current noise
100
EE 214 Lecture 8B. Murmann 23
1/f Noise Contribution (2)
• For circuits with "high bandwidth", flicker noise is often insignificant
– Beware of exceptions…
102
104
106
108
10-23
10-22
10-21
10-20
f [Hz]
Sp
ect
rum
[A2 /H
z]
102
104
106
108
0
50
100
f [Hz]
Sq
rt(I
nte
gra
l) [n
A]
EE 214 Lecture 8B. Murmann 24
Lower Integration Limit
• Does the flicker noise PSD go to infinity for f→0?
– For a fun discussion, seeE. Milotti, "1/f noise: a pedagogical review," available at http://arxiv.org/abs/physics/0204033
• Even if the PSD goes to infinity, do we care?
– Let's say we are sensing a signal for a very long time (down to a very low frequency), e.g.
• 1 year ≅ 32 Msec, 1/year ≅ 0.03 µHz
– Number of frequency decades in 1/year to 100Hz ≅ 10• Means instead of 7 decades, we'd now have 17 in the previous
example
• sqrt(17/7) = 1.56 → Only 56% more flicker noise!
101
EE 214 Lecture 8B. Murmann 25
MOS Model with Noise Generator
Noiseless!
f
f
LW
g
C
KfgkT4i
2m
ox
fm
2d
∆∆γ⋅
+⋅⋅⋅=
EE 214 Lecture 8B. Murmann 26
Other MOSFET Noise Sources
• Gate noise
– "Shot noise" from gate leakage current
– Noise from to finite resistance of gate material
– Noise due to randomly changing potential/capacitance between channel and bulk
• Relevant only at very high frequencies
• Bulk noise
• Source barrier noise in very short channels (Navid & Dutton)
– Shot noise from carriers injected across source barrier
• More in EE314…
102
EE 214 Lecture 9B. Murmann 1
Lecture 9Electronic Noise (Continued)
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 9B. Murmann 2
Overview
• Reading
– 11.4 (Circuit Noise Calculations)
– 11.5 (Equivalent Input Noise Generators)
– 11.9 (Noise Bandwidth)
• Introduction
– Having established the basic noise mechanisms in MOSFETS, today's lecture looks at noise in circuits. We will learn how to calculate the signal to noise ratio in a basic RC circuit and a common source amplifier. These examples are useful prerequisites for analyzing more complicated circuits (e.g. OTAs, later in this class). Furthermore, we will use these simple to develop a basic feel for the relevance of noise and associated tradeoffs.
103
EE 214 Lecture 9B. Murmann 3
Noise in Circuits (1)
• Most interesting circuits have more than one relevant noise source
• In order to quantify the net effect of all noise sources, we must refer the noise sources to a single "interesting" node pair of the circuit
– Usually output or input
EE 214 Lecture 9B. Murmann 4
Noise in Circuits (2)
• Output referred noise– Refer noise to output via
individual noise transfer functions
– Physical concept, exactly what one would measure in the lab
• Input referred noise– Represent total noise via a
fictitious input source that captures all circuit-internal noise sources
– Useful primarily for "fair" comparisons
• E.g. independent of circuit gain
104
EE 214 Lecture 9B. Murmann 5
Datasheet Example
EE 214 Lecture 9B. Murmann 6
Circuit Example 1
• Let's calculate
– Output referred noise
– Input referred noise
– Signal-to-noise ratio at output
105
EE 214 Lecture 9B. Murmann 7
Output Referred Noise
( )2
222
out,nns1f2j1
1nV4sRC11RkT4
fv
⋅+=
+⋅⋅=
π∆
• Short input voltage, add noise source
• Calculate transfer function from noise source to output
• Multiply noise PSD with squared transfer function to get output PSD
EE 214 Lecture 9B. Murmann 8
Input Referred Noise
RVout
1k
C
1pF
v2
( )22
in,n nV4RkT4f
v=⋅=
∆
• Simply equal to noise PSD of resistor (in this example)
106
EE 214 Lecture 9B. Murmann 9
+ A Note on Equivalent Input Noise Generators
• In general, input referred noise must be modeled by an equivalent voltage and current source
– See section 11.5 in the text (optional)
• Modeling the noise with a voltage alone I sufficient when the circuit is indeed driven by an ideal voltage source (always an approximation)
• Or, when the input impedance of the circuit is large
– This is the case in MOS circuits at low to moderate frequencies
– Different story at RF frequencies• You'll learn more about this in EE314
EE 214 Lecture 9B. Murmann 10
Spice Simulation
$ EE214 RC circuit noise
$ Boris Murmann, September 2006
vin in 0 ac 1r1 in out 1k
c1 out 0 1pF
.ac dec 100 100 1gig
.noise v(out) vin
.options post brief
.lib './ee214_hspice.txt' nominal
.end
104
106
108
10-18
10-17
10-16
f [Hz]
No
ise
[V
2 /Hz]
outnoiseinnoise
107
EE 214 Lecture 9B. Murmann 11
Signal-to-Noise Ratio
• Over which bandwidth should we integrate the noise?
• Two interesting cases
– The output is measured or observed by a system with finite bandwidth (e.g. human ear)
• Use frequency range of that system as integration limits
• Applies on a case by case basis
– "Total integrated noise"
∫ ⋅
==2
1
f
f
2out,n
2peak,out
noise
signal
dff
v
V21
PP
SNR
∆
EE 214 Lecture 9B. Murmann 12
Total Integrated Noise (1)
• Interesting result
– Total noise at the output depends only on C (even though R is generating the noise)
CkTdf
RCf2j11RkT4v
0
22
tot,out,n =⋅+
⋅⋅= ∫∞
π
• Simply the integral over all frequencies (zero to infinity)
• Relevant metric when output is observed without (significant) band limiting– Or when the output is sampled (e.g. switched capacitor
circuits)
• Most suitable for getting a general feel about noise and to establish basic quantitative arguments
108
EE 214 Lecture 9B. Murmann 13
Total Integrated Noise (2)
• Increasing R increases the noise power spectral density, but also decreases the bandwidth
– R drops out in the end result
102
104
106
108
1010
10-22
10-20
10-18
10-16
10-14
f [Hz]
Sp
ect
rum
[V2 /H
z]
102
104
106
108
1010
0
20
40
60
f [Hz]
Sq
rt(I
nte
gra
l) [ µ
V]
R=1kR=100k
R=1kR=100k
EE 214 Lecture 9B. Murmann 14
SNR (1)
( ) dB8110122log10 6 =⋅
• Back to our example; plugging in numbers
( )6
2
222peak,out
noise
signal 10122V64V5.0
pF1kTV5.0
CkT
V21
PP
SNR ⋅=====µ
• SNR in dB
• Is this "good"?
• Typical system requirements
– Audio: SNR ≅ 100dB
– Video: SNR ≅ 60dB
– Gigabit Ethernet Transceiver: SNR ≅ 35dB
109
EE 214 Lecture 9B. Murmann 15
SNR (2)
• Assuming Vout,peak = 1V
SNR [dB] C [pF]20 0.0000008340 0.00008360 0.008380 0.83
100 83120 8300140 830000
Hard to make such small capacitors…
Designer will definitely be concerned about thermal noise, capacitor sizes set by SNR
"Hardcore" thermal noise battle…
• General rules of thumb
– Up to SNR ~ 40dB, integrated circuits are usually not limited by thermal noise
– Achieving SNR >100dB is extremely difficult• Must usually rely on external components or "tricks" (such as
oversampling, see EE315)
EE 214 Lecture 9B. Murmann 16
MDS and DR
• Minimum detectable signal (MDS)
– A somewhat arbitrary definition
– Quantifies the signal level in a circuit that yields SNR=1; i.e.noise power = signal power
• Dynamic range (DR)
MDSP
DR max,signal=
• If the noise level in the circuit is independent of the signal level (not always the case), it follows that the DR is equal to the "peak SNR", i.e. the SNR with the maximum signal applied
110
EE 214 Lecture 9B. Murmann 17
Circuit Example 2: Common Source Amplifier
( )
( )
( )
α
γ
γ
πγ
πγ
⋅=
+=
+=
⋅+⋅⎟⎠⎞
⎜⎝⎛ +=
⋅+⋅⎟⎠⎞
⎜⎝⎛ +=
∫∞
CkT
A1CkT
Rg1CkT
dfRCf2j1
RgR1kT4v
RCf2j1Rg
R1kT4
dffv
v
m
0
2
m2
tot,o
2
m
2o
[Ignoring 1/f noise for simplicity]
EE 214 Lecture 9B. Murmann 18
CS Stage Noise/Power Tradeoff
• Assuming that we're already using the maximum available signal swing, improving the SNR by 6dB means
– Increase C by 4x
– Decrease R by 4x to maintain bandwidth
– Increase gm by 4x to preserve gain
• Assuming that we can keep gm/ID constant, this means that we must increase ID by 4x
• Bottom line
– Improving the SNR in a noise limited circuit by 6dB ("1bit") QUADRUPLES power dissipation !
111
EE 214 Lecture 10B. Murmann 1
Lecture 10Body Effect
Common Gate Stage
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 10B. Murmann 2
Overview
• Reading– 1.6.6 (Body Transconductance)– 3.3.4, 7.2.4.2 (Common Gate Stage)– 3.4.2.2, 7.3.4 (up to p. 526) (Cascode Stage)
• Introduction– Having completed our discussion of simple common source
amplifiers, we now continue by exploring alternative ways of forming and using single transistor stages. First, we will look at the common gate stage, followed by a discussion of the common drain stage in future lectures. While CS stages can usually be configured with source and substrate nodes tied together, this is often not the case in the CG and CD configurations. Hence, we'll first need to take a look at the socalled "body effect," which becomes significant in this context.
Somewhat tedious to read…
112
EE 214 Lecture 10B. Murmann 3
The "Atoms" of Analog Circuit Design
• As we've seen from the discussion so far, a common source stage is sufficient for building a simple amplifier
– How about the other two possible configurations?
• We'll find that common gate and drain stages can be incorporated as valuable add-ons, for building "better" amplifiers
• Interestingly, many analog circuits can be decomposed into a combination of the above three fundamental building blocks
EE 214 Lecture 10B. Murmann 4
Body Connection
• In the EE214 (N-well) technology, only the PMOS device has an isolated body connection
• Newer technologies (e.g. 0.13µm CMOS) also tend to have NMOS devices with isolated body ("twin-well" process)
113
EE 214 Lecture 10B. Murmann 5
Body Connection Scenarios
EE 214 Lecture 10B. Murmann 6
Body Effect (1)
• With positive VSB, depletion region around source grows
• Increasing amount of negative fixed charge in depletion region tends to "repel" electrons coming from source
– Need larger VGS to compensate for this effect
VSB>0
114
EE 214 Lecture 10B. Murmann 7
Body Effect (2)
• This effect is usually factored in as an effective increase in Vt
• Detailed analysis shows
( )fSBftt VVV φφγ 220 −++=
• A change in Vt also means a change in drain current
– Define small signal body transconductance
SB
D
BS
Dmb V
IVIg
∂∂−=
∂∂=
fSBSB
t
D
GS
t
D
SB
t
m
mb
VVV
IV
VI
VV
gg
φγ
22 +=
∂∂=
∂∂
∂∂
∂∂−=
-1
EE 214 Lecture 10B. Murmann 8
gmb Simulation
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 210
12
14
16
18
20
22
24
NMOS W/L=10/0.35um, VGS
-Vt=0.2V
VBS
[V]
gm
b/gm
[%]
SpiceFit using 2φ
f=0.6V, γ=0.36V-1
VSB
115
EE 214 Lecture 10B. Murmann 9
Modified Small Signal Model
EE 214 Lecture 10B. Murmann 10
Common Gate Stage
RS
RL
Vo
-(gm+gmb)vi
+vi-
ii Cgs+Csb
Cgd+Cdb
ro
mbmmsbgsS gg'gCCC +=+=Define:
116
EE 214 Lecture 10B. Murmann 11
CG Current Transfer
11
1 >>+
≅ Sm
m
Si
o R'gfor
'gCsi
i
Sm
SSSm
Sm
SSm
m
i
o
R'gCRsR'g
R'gR
sC'g
'gii
++
⋅+
≅
++≅
11
11
1
EE 214 Lecture 10B. Murmann 12
CG Input Impedance (1)
testgs vv −=
( )
SoL
om
test
testin
o
o
o
testtestmtesttest
testoLmo
testmo
test
o
o
L
oo
sCrRr'g
viY
rv
rvv'gi:v@KCL
vr||R'gv
v'gr
vrv
Rv:v@KCL
++
≅=⇒
−+=
≅⇒
−−+=0
117
EE 214 Lecture 10B. Murmann 13
CG Input Impedance (2)
• At low frequencies
( )⎟⎟⎠
⎞⎜⎜⎝
⎛ +++
≅om
oLS
oL
omin r'g
rRsCrRr'gY 1
⎟⎟⎠
⎞⎜⎜⎝
⎛+≅=
o
L
minin r
R'gY
R 111
• Two interesting cases
– RL<<ro:
– RL>>ro:
min 'g
R 1≅
om
Lin r'g
RR ≅
(well known)
(not so well known…)
EE 214 Lecture 10B. Murmann 14
CG Output Impedance
Stestgs
gsmo
gs
o
testtest
Riv
v'grv
rvi
−=
++=
( )Smotest
testout R'gr
ivR +≅= 1
(Very high if g'mRS>>1 !)
118
EE 214 Lecture 10B. Murmann 15
CG Summary
• Current gain is unity up to very high frequencies
– Our "simple," ignorant device model predicts up to roughly fT
• Input impedance is very low
– At least when the output is also terminated with some reasonable impedance
• Can achieve very high output resistance
• In summary, a common gate stage is ideal for turning a decent current source into a much better one
– Seems like this is something we can use to improve our common source stage
• Which is indeed nothing but a decent (voltage controlled) current source
EE 214 Lecture 10B. Murmann 16
Cascode Stage
• Invented ~1920, in the context of vacuum tube circuits
– http://web.mit.edu/klund/www/cascode.html
( )12210
1 1 omoomi
mm r'grRgiigG +≅≅⋅=
( ) ( )222111221 1 omomomomomom rg~r'grgr'grgRG ⋅≅+=
119
EE 214 Lecture 10B. Murmann 17
High Frequency Benefits
• Very close to 1, for moderate values of RL
– Mitigates Miller effect
– Even if RL is large, there is often a load capacitance that provides a low impedance termination to help maintain this feature
• Additional benefit
– Cascode mitigates direct forward coupling from Vi to Vo at high frequencies
⎟⎟⎠
⎞⎜⎜⎝
⎛+≅=
22
11 1
o
L
m
mxm
i
x
rR
'ggZg
vv
EE 214 Lecture 10B. Murmann 18
High Frequency Issues
• Cascode causes pole around fT– Usually non-dominant
– Can be a headache for stability/phase margin in circuits with feedback
• More later…
m
sbgsi
o
'gCC
sii
++
≅1
1
120
EE 214 Lecture 10B. Murmann 19
Design Example 2 (Lecture 6) Revisited
• What we expect to see in Spice after adding the cascode device
– Gain should be closer to gm1·RL
– Bandwidth should increase (reduction of Miller effect)
– Non-dominant pole around some fraction of fT of cascode device
M2
M1
EE 214 Lecture 10B. Murmann 20
Frequency Response (1)
• Bandwidth increased from 180MHz to 250MHz
100
101
102
0
5
10
15
f [MHz]
|vo
/vi|
[dB
]
Without cascodeWith cascode
121
EE 214 Lecture 10B. Murmann 21
Frequency Response (2)
10-3
10-2
10-1
100
101
102
-100
-80
-60
-40
-20
0
20
f [GHz]
|vo
/vi|
[dB
]
Without cascodeWith cascode
EE 214 Lecture 10B. Murmann 22
Non-Dominant Pole Estimate
subckt element 0:m2 0:m1 model 0:nch214 0:nch214 region Saturati Saturatiid 514.6895u 514.6895uvgs 918.7732m 815.0000mvds 904.0836m 581.2268mvbs -581.2268m 0. vth 705.0212m 587.7659mvdsat 178.3723m 175.2946mbeta 22.3498m 22.1558mgam eff 926.5549m 894.1238mgm 4.0936m 3.9871mgds 94.1732u 108.6333ugmb 676.0386u 898.3368ucdtot 43.0143f 52.2437fcgtot 56.0163f 56.5703fcstot 91.3803f 108.9679fcbtot 93.8478f 121.9629fcgs 40.2737f 39.5319fcgd 7.1355f 7.2417f
( ) ( )fF2.7fF52fF40fF4.91fF40mS68.0mS4
21
CCC'g
21f
1db2sb2gs
m2p
−+−++≅
++≅
π
π
GHz2.5f 2p ≅
GHz4.11fF56
mS421f 2T
≅
≅π
For comparison:
122
EE 214 Lecture 10B. Murmann 23
Supply Headroom Issue
• Even if we adjust VB such that VDS1 is small, adding a cascode reduces the available signal swing
• This can be a big issue when designing circuits with VDD≅1V
– Typically need each VDS>~0.2V
– A severe dynamic range penalty
EE 214 Lecture 10B. Murmann 24
Cascode Noise
• It is typically argued that cascodes no not add a significant amount of noise
• A closer look reveals that cascodes can contribute significant noise at high frequencies
– Noise current A no longer compensates B at high frequencies
• We'll take a more quantitative look at this later in this course
VB
Vi
Vo
RL
VDD
i2n1
i2n2
VB
Vi
Vo
RL
VDD
i2n1
A
B
123
EE 214 Lecture 11B. Murmann 1
Lecture 11Common Drain Stage
(Source Follower)
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 11B. Murmann 2
Overview
• Reading
– 3.3.7 (Common Drain Stage)
– 7.2.2 (Frequency Response of Voltage Buffers)
– 5.3 (Source Follower as an Output Stage, optional)
• Introduction
– Last lecture, we have seen that a common gate stage has a fairly low input impedance, and high output impedance. The common drain stage that we'll analyze today exhibits the exact opposite features: High input impedance and low output impedance. After an analysis of relevant port characteristics, we will discuss some potential applications and also drawbacks of this circuit.
124
EE 214 Lecture 11B. Murmann 3
Common Drain Stage
EE 214 Lecture 11B. Murmann 4
CD Voltage Transfer (1)
LtotLtotgsm
gsm
i
o
RsCsCg
sCgvv
1+++
+=
( ) 01 =−−−⎟⎟⎠
⎞⎜⎜⎝
⎛++ oimgsi
LtotgsLtoto vvgsCv
RsCsCv
( )
Ltotm
Ltotgs
m
gs
Ltotm
m
i
o
Rg
CCsg
sC
Rg
gvv
11
1
1
+
++
+⋅
+=
omb
LLtotsbLLtot r||g
||RRCCC 1=+=
125
EE 214 Lecture 11B. Murmann 5
Low Frequency Gain
• Interesting cases
– RL→∞, ro→∞, gmb=0• PMOS, source tied to body, ideal current source
Ltotm
mv
Rg
ga 10
+= o
mbLLtot r||
g||RR 1=
10 =va
mbm
mv gg
ga+
=0– RL→∞, ro→∞, gmb≠0• NMOS, ideal current source)
(typically ≅ 0.8)
– ro→∞, gmb=0, RL finite• PMOS, source tied to body, load resistor
Lm
mv
Rg
ga 10
+=
EE 214 Lecture 11B. Murmann 6
High Frequency Gain
• Three scenarios
( )pszs
avvsa v
i
ov
−
−⋅==1
10
gs
m
Cgz −=
Ltotgs
Ltotm
CCR
gp
+
+−=
1
(infinite bandwidth !?)
|z|<|p| |z|>|p| |z|=|p|
126
EE 214 Lecture 11B. Murmann 7
CD Input Impedance
• By inspection
( ) ( ))s(asCCCsY vgsgbgdin −++= 1
• Gain term av(s) is real and close to unity up to fairly high frequencies
• Hence, up to moderate frequencies, we see a capacitor looking into the input
– A fairly small one, Cgd + Cgb, plus a fraction of Cgs
EE 214 Lecture 11B. Murmann 8
PMOS Stage with Body-Source Tie
• Gate-body capacitance is in parallel with Cgs
• gmb generator inactive
– Low frequency gain very close to unity
• Very small input capacitance
( )( )gdin
vgbgsgdin
sCY)s(aCCssCY
≅
−++= 1
127
EE 214 Lecture 11B. Murmann 9
Bootstrapped PMOS Stage
• "Extremely" small input capacitance
( ))s(a)s(asCY vNvPgdin −≅ 1
EE 214 Lecture 11B. Murmann 10
CD Output Impedance (1)
• Let's first look at an analytically simple case
– Input driven by ideal voltage source
• By inspection
( )sbgsmbmout CCsgg
Z++
= 11
• Low output impedance
– Resistive up to very high frequencies
128
EE 214 Lecture 11B. Murmann 11
CD Output Impedance (2)
• Now include finite source resistance
( )( )( )gsm
o
go
gsmgox
sCgvv
v
sCgvvi
+⎟⎟⎠
⎞⎜⎜⎝
⎛−=
+−=
1x
ox i
vZ =
igs
i
o
g
RsC
Rvv
+= 1
EE 214 Lecture 11B. Murmann 12
CD Output Impedance (3)
• Two interesting cases
( )
⎟⎟⎠
⎞⎜⎜⎝
⎛+
+≅
m
gs
gsi
mx
gsC
CsRg
Z1
11
Ri > 1/gmRi < 1/gm
Inductive behavior!
129
EE 214 Lecture 11B. Murmann 13
Equivalent Circuit for Ri > 1/gm
• This circuit is prone to ringing!
– L forms an LC tank with any capacitance at the output
1
1
22
21
−=
=
=
im
gsi
i
m
RgCR
L
RRg
R||R
EE 214 Lecture 11B. Murmann 14
Inclusion of Parasitic Input Capacitance
( )( )ii
m
gs
igsi
mx
CsRg
sCCCsR
gZ
+⎟⎟⎠
⎞⎜⎜⎝
⎛+
++=
11
11
• What happens to this result if we don’t neglect Ci=Cgd+Cgb?
gs
m
iiigsi Cg
CRCCR<<
+11
iigs
m
igsi CRCg
CCR11 <<
+
130
EE 214 Lecture 11B. Murmann 15
Application 1: Level Shifter
• Output quiescent point is roughly Vt+Vov lower than input quiescent point
EE 214 Lecture 11B. Murmann 16
Application 2: Buffer
• Low frequency voltage gain of the above circuit is ~gmRbig
– Would be ~gm(Rsmall||Rbig) without CD buffer stage
131
EE 214 Lecture 11B. Murmann 17
Issues
• Several sources of nonlinearity
– Vt is a function of Vo (NMOS, without S to B connection)
– ID and thus Vov changes with Vo
• Gets worse with small RL
• Reduced input and output voltage swing
– Consider e.g. VDD=1V, Vt=0.3V, VOV=0.2V• CD buffer stage consumes 50% of supply headroom!
– In low VDD applications that require large output swing, using a CD buffer is often not possible
– CD buffers are more frequently used when the required swing is small
• E.g. pre-amplifiers or LNAs that turn µV into mV at the output
EE 214 Lecture 11B. Murmann 18
Application 3: Load Device
• Advantages compared to resistor load
– "Ratiometric"• Gain depends on ratio of similar
parameters
• Reduced process and temperature variations
– First order cancellation of nonlinearities
• Disadvantage
– Reduced swing
22
10
mbm
mv gg
ga+
=
132
EE 214 Lecture 11B. Murmann 19
Summary – Elementary Transistor Stages
• Common source
– VCCS, makes a good voltage amplifier when terminated with a high impedance
• Common gate
– Typically low input impedance, high output impedance
– Can be used to improve the intrinsic voltage gain of a common source stage
• "Cascode" stage
• Common drain
– Typically high input impedance, low output impedance
– Great for shifting the DC operating point of signals
– Useful as a voltage buffer when swing and nonlinearity are not an issue
EE 214 Lecture 11B. Murmann 20
Biasing Issue
• Something we haven't talked about
– How to build VB?
• Especially when gain is large, output quiescent voltage is very sensitive to changes in VB
VDD
Vo
vi
VB
133
EE 214 Lecture 11B. Murmann 21
A Possible Solution
• VGS of M1 and M1' are equal, which implies that their drain currents are (approximately) equal
• gm1 is set by a current (IB), rather than a voltage– Independent of Vt
• Great!
• Problem: Cbig and Rbig
– High pass filter– Large area for low corner
frequency– Typically, this approach is
practical only in some RF circuits
EE 214 Lecture 11B. Murmann 22
Better Solution: Differential Amplifier
• Bias point of Vop and Vom set by IB– Independent of quiescent
value of Vip and Vim
• Differential output (Vop-Vom) depends only on differential input (Vip-Vim)
• More next lecture…
134
EE 214 Lecture 12B. Murmann 1
Lecture 12Differential Pair
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 12B. Murmann 2
Overview
• Reading
– 3.5.0, 3.5.3, 3.5.5 (Differential Pair)
• Introduction
– The differential pair is the most widely used two-transistor sub-circuit in analog ICs. As we have already seen last time, using a differential pair as a replacement for a simple common source stage eliminates the need for a precise gate bias voltage (VB). In addition, its differential input and output voltages are more immune to parasitic signal coupling. Today we will analyze some properties of differential pairs and introduce the notion of common- and differential-mode signal components.
135
EE 214 Lecture 12B. Murmann 3
Differential Pair
• When Vip=Vim, and both transistors are identical, we must have Id1=Id2=ITAIL/2
• How about Vip=Vim=1V versus Vip=Vim=2V?
– Makes no difference!
• From a signal perspective, we care only about the difference of the applied voltages
– Makes sense to introduce a new variable
• Vid=(Vip-Vim)
ITAIL
Id1
Vip
Id2
Vim
EE 214 Lecture 12B. Murmann 4
Differential and Common Mode (1)
• We now still need a second variable that describes the potential of nodes Vip and Vim
with respect to GND
– Could choose either Vip or Vim
• More elegant solution
– Cut Vid in half and define a new independent variable
– "Common mode" voltage Vic
ITAIL
Id1
+Vip-
Id2
Vid
+Vim-
136
EE 214 Lecture 12B. Murmann 5
Differential and Common Mode (2)
ITAIL
Id1
+Vip-
Id2
Vid/2+
Vim-
Vid/2
Vic
imipid VVV −=
2
2id
icim
idicip
VVV
VVV
−=
+=
2imip
ic
VVV
+=⇒
EE 214 Lecture 12B. Murmann 6
Coupling Noise Immunity
• Using differential pairs/signals not only solves the biasing issue, but also mitigates coupling noise issues
Single Ended Signaling Differential Signaling
137
EE 214 Lecture 12B. Murmann 7
Large Signal Transfer Function (1)
• Let's first do a simple analysis using long channel equations
ITAIL
Id1
M1
+Vip-
Id2
Vid/2+
Vim-
Vid/2
Vic
M2
LWC
IVV
LWC
IVV
IIIVVVV
ox
dtgs
ox
dtgs
TAILdd
gsimgsip
µµ2
21
1
21
21
22 +=+=
=+
−=−
2id
ox
TAILidox2d1dod V
LWC
I4VL
WC21III −=−=⇒
µµ
EE 214 Lecture 12B. Murmann 8
Large Signal Transfer Function (2)
2
21
21 ⎟⎟
⎠
⎞⎜⎜⎝
⎛−=−=⇒
OV
id
OV
id
TAIL
dd
TAIL
od
VV
VV
III
II
• We can turn this into a more elegant expression by using
2
21
2 OVoxTAIL V
LWCI µ=
where VOV is the quiescent point gate overdrive with Vid=0
• This equation predicts
– Iod/ITAIL = 0 when Vid=0, as expected
– Complete current steering (Iod/ITAIL=±1) takes place when Vid= ±VOV√2
138
EE 214 Lecture 12B. Murmann 9
Large Signal Plot
• Note that the equation on previous slide is only valid in the center of this transfer function (between saturation points)– Why?
Iod/ITAIL
⎝
2 1 0 1 2
1
0
1Slope = ITAIL/VOV
Vid/VOV-√2 √2
EE 214 Lecture 12B. Murmann 10
Observations
• Looks like something we have seen before
– A transfer function that is somewhat linear as long as Vid<<VOV
• For small signal analysis, we can find an equivalent transconductance by differentiation at the operating point
OV
TAIL
Vid
odm V
IdVdIG
id
===0
• Note that the transconductance of M1 and M2 is given by
OV
TAIL
OV
TAIL
OV
D,m V
IV
I
VIg === 2
2221
139
EE 214 Lecture 12B. Murmann 11
Does the Tail Node Move?
• Can show that
ITAIL
Id1
+Vip-
Id2
Vid/2+
Vim-
Vid/2
Vic
Vx
2
411 ⎟⎟
⎠
⎞⎜⎜⎝
⎛−−−=
OV
idOVticx V
VVVVV
• From this expression, we see that from a small signal perspective the tail node is pinned at Vic-Vt-VOV
– "AC ground"
EE 214 Lecture 12B. Murmann 12
Small Signal Equivalent
• Sufficient to work with half circuit!
– Can directly apply everything we've learned about single transistor stage
• Half circuit caveats
– Can not analyze nonlinearity using half circuits
– Assumes that M1 and M2 are identical
id1 id2
gmvid/2 Vx
-gmvid/2
iod/2
gmvid/2
iod
gmvid
idmddod vgiii =−= 21
140
EE 214 Lecture 12B. Murmann 13
Differential Voltage Amplifier Example
EE 214 Lecture 12B. Murmann 14
Short Channel Effects
• Can model short channel effects using long channels with source degeneration
ITAIL
Id1 Id2
Rsx Rsx
LWC
IVV
LWC
IVV
IIIRIVVRIVV
ox
dtgs
ox
dtgs
TAILdd
sxdgsimsxdgsip
µµ2
21
1
21
2211
22 +=+=
=+
−−=−−
mess... big⇒LV
gR
c
OV
msx ε
1=
141
EE 214 Lecture 12B. Murmann 15
Nonlinearity Comparison (1)
• Can look at this using Taylor expansions
• Without Rsx (long channel), we can show that
...VV
VV
II
OV
id
OV
id
TAIL
od −⎟⎟⎠
⎞⎜⎜⎝
⎛−⎟⎟⎠
⎞⎜⎜⎝
⎛≅
3
81
• With Rsx (short channel) this becomes
( ) ( ) ...VRg
VRgVRg
VII
OVsxm
id
sxmOVsxm
id
TAIL
od −⎟⎟⎠
⎞⎜⎜⎝
⎛⋅+⎟⎟
⎠
⎞⎜⎜⎝
⎛+
−⎟⎟⎠
⎞⎜⎜⎝
⎛⋅+
≅3
111
81
1
• Short channel differential pair is "more linear"
– But also has less gm at a given current
• For fairness, let's compare nonlinearity at same gm/ID
EE 214 Lecture 12B. Murmann 16
Nonlinearity Comparison (2)
• Recall that
OVsxmD
m
VRgIg 2
11
+≅
• Plugging this into the equations on the previous slide yields
...VIgV
Ig
II
idD
mid
D
m
TAIL
od −⎟⎟⎠
⎞⎜⎜⎝
⎛−⎟⎟⎠
⎞⎜⎜⎝
⎛≅
3
21
81
21
...VIg
RgV
Ig
II
idD
m
sxmid
D
m
TAIL
od −⎟⎟⎠
⎞⎜⎜⎝
⎛⎟⎟⎠
⎞⎜⎜⎝
⎛+
−⎟⎟⎠
⎞⎜⎜⎝
⎛≅
3
21
11
81
21
Long Channel
Short Channel
• Minor win for short channel in terms of linearity
– In practice, this may be overshadowed by additional effects that are not captured in the simple degeneration model
142
EE 214 Lecture 12B. Murmann 17
"Linear Region" of Transfer Function
• In the long channel case, we found that for good linearity we need Vid<<VOV
• As we have seen from the derivation above, this requirement can be generalized to capture both long and short channel cases simultaneously
• A few approximate rules of thumb
⎟⎟⎠
⎞⎜⎜⎝
⎛<<
D
mid
Ig
V 2
⎟⎟⎠
⎞⎜⎜⎝
⎛⋅<
⎟⎟⎠
⎞⎜⎜⎝
⎛⋅<
D
mid
D
mid
Ig
.V
Ig
.V 220250
For ~1.5% nonlinearity error: For ~0.1% nonlinearity error:
EE 214 Lecture 12B. Murmann 18
Voltage Amplifier Transfer Functions
• In a differential amplifier, we primarily want to have large gain that links only the two differential variables
id
oddm v
vA =
• Unfortunately, circuit nonidealities will also cause nonzero "parasitic" gain terms
ic
occm v
vA =id
occmdm v
vA =−ic
oddmcm v
vA =−
See text 3.5.6.9
143
EE 214 Lecture 12B. Murmann 19
Common Mode Gain
• Ideally zero (RTAIL=∞)
• With finite RTAIL:
ic
occm v
vA =
[ ]( )TAILmoTAILm
mcm RgrR
RggA 21
21⋅+⋅
⋅+−=
EE 214 Lecture 12B. Murmann 20
Common Mode Rejection Ratio
• Figure of merit that quantifies ratio desired/undesired gain
– Ideally infinite
cm
dm
AACMRR =
• For our simple resistively loaded differential pair, this becomes (assuming R<<ro and ignoring body effect)
TAILm
TAILm
mm R2g1
RR2g1
gRgCMRR ⋅+=
⋅⋅+
⋅≅
• Other important figures of merit
dmcm
dm
AA
− cmdm
dm
AA
−
144
EE 214 Lecture 12B. Murmann 21
Power Supply Rejection Ratio
• In practice, "noise" on the supplies will also propagate to the output
– In a differential system usually due to (half-) circuit imbalance
• Define
dd
od
vvA =+
ss
od
vvA =−
++ = A
APSRR dm
−− = A
APSRR dm
EE 214 Lecture 12B. Murmann 22
Input Referred Interpretation
• E.g. 1mV input signal, 100mV supply noise
– Need PSRR >> 100 (40dB)
• PSRR can be a very critical issue in highly integrated, complex integrated circuits
– Lots of potential supply noise sources• E.g. cross-talk between analog and digital sections
Amplifier
VDD
vi or vidvo or vod
+PSRRvdd
−PSRRvss
145
EE 214 Lecture 13B. Murmann 1
Lecture 13Offset Voltage
Current Mirrors
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 13B. Murmann 2
Overview
• Reading
– 3.5.6.6, 3.5.6.7 (Input Offset Voltage)
– 4.1, 4.2 (Current Mirrors)
• References– M. Pelgrom et al., "Matching properties of MOS transistors," IEEE J. Solid-
State Circuits, Oct. 1989.
– P.G. Drennan et al., "Understanding MOSFET mismatch for analog design," IEEE J. Solid-State Circuits, March 2003.
• Introduction
– So far, we have worked with perfectly balanced differential pairs. In practice, there is always some imbalance due to matching errors between circuit elements. In this lecture, we will consider the input referred offset of a differential pair as an example. Following this discussion, we will take a closer look at practical implementations of current mirrors with emphasis on high swing biasing techniques.
146
EE 214 Lecture 13B. Murmann 3
Offset Voltage
• Defined as the input voltage one needs to apply to make the differential output of an amplifier equal to zero
• There is no offset in a perfectly balanced differential amplifier
• Two fundamentally different sources of offset
– Systematic• Due to an "imbalance by design"
• See e.g. lecture 22, OTA with single ended output
– Random• Due to mismatch between nominally identical devices
EE 214 Lecture 13B. Murmann 4
Resistor Mismatch
• Want large gm/ID to mitigate this error (low VOV)
⎟⎠⎞
⎜⎝⎛ ∆+⋅=
RRRIRI dd 121
D
d
D
d
D
d
d
d
II
II
II
II
RR ∆+≅∆−
∆+==∆+ 1
211
211
12
1
RRV
Ig
II
osD
m
D
d ∆≅≅∆
RR
IgV
D
mos
∆⎟⎟⎠
⎞⎜⎜⎝
⎛≅
−1
147
EE 214 Lecture 13B. Murmann 5
Current Factor Mismatch
LWCoxµβ =
( ) 22
21 ovov VV βββ ∆+=
ββ∆+=12
2
21
ov
ov
VV
ββ∆+≅+≅
2111
2
1
OV
os
ov
ov
VV
VV
ββ∆
⎟⎟⎠
⎞⎜⎜⎝
⎛≅
−1
D
mos I
gV
EE 214 Lecture 13B. Murmann 6
Threshold Mismatch
tttos VVVV ∆=−= 21
148
EE 214 Lecture 13B. Murmann 7
Complete Expression
• Experimental data shows that it is reasonable to model the errorterms as uncorrelated random variables
– Need to add up variances
⎥⎦
⎤⎢⎣
⎡ ∆+∆⎟⎟⎠
⎞⎜⎜⎝
⎛+∆≅
−
RR
IgVV
D
mtos β
β1
⎥⎥⎦
⎤
⎢⎢⎣
⎡+⎟⎟
⎠
⎞⎜⎜⎝
⎛+≅ ∆∆
−
∆22
22
RR
D
mVV I
gtos
σσσσββ
• The error standard deviations depend on device geometries
– Usually larger area means smaller offset
EE 214 Lecture 13B. Murmann 8
Pelgrom Coefficients
• In 0.35µm technology: AVt ≅ 7mV-µm, Aβ ≅ 1%-µm
– AVt tends to scale down with technology, roughly proportional to gate oxide thickness
– Aβ has remained roughly constant with scaling
• Example: W=50µm, L=0.35µm, gm/ID=10V-1, ∆R=0
WLA
WLA
t
t
VV
β
ββσσ == ∆∆
( ) ( ) ( ) ( ) mV.mV.mV..
%V.
mVosV 681240671
350501
101
350507 22
2
2
2
=+=⋅
+⋅
=σ
• Vt mismatch usually dominates Vos, especially in moderate and weak inversion (large gm/ID)
149
EE 214 Lecture 13B. Murmann 9
Current Mirror Bias
• Objectives
– Want accurate mirror ratio ITAIL/IREF
– Want large RTAIL (and small CTAIL) for good CMRR
– Want small Vmin to maximize common mode input range
EE 214 Lecture 13B. Murmann 10
Basic Sizing Considerations
• Always use L1=L2
• Make W1/W2 or W2/W1 integer
– Use unit devices connected in parallel
– "m-factor" in Spice• E.g. M1 d g s b W=10u L=0.35u m=5
150
EE 214 Lecture 13B. Murmann 11
Inaccuracy due to Offset
• Want small gm/ID ("large VOV") to mitigate errors due to random and systematic offset in VGS
– Unfortunately this means large Vmin
( )2121 ososm VVgIII +≅−=∆
( )21 ososD
m VVIg
II +≅∆
EE 214 Lecture 13B. Murmann 12
Inaccuracy due to ∆VDS
• Two options
– Use device with large ro (Large L)
– Make V1 as close as possible to V2
oo rV
rVVIII ∆=−≅−=∆ 21
21
VDS
ID
∆V
∆Islope=1/ro
151
EE 214 Lecture 13B. Murmann 13
Output Resistance
0 0.5 1 1.5 2 2.5 3 3.50
500
1000
1500
2000
2500
NMOS W/L=10, VGS
=800mV, L=0.35µm...0.7µm
VDS
[V]
r o [k Ω
]
EE 214 Lecture 13B. Murmann 14
Output Resistance Zoom
0 0.1 0.2 0.3 0.4 0.5 0.6 0.70
100
200
300
400
500
600
700
NMOS W/L=10, VGS
=800mV, L=0.35µm...0.7µm
VDS
[V]
r o [k Ω
]
≅2/(gm/ID) ≅ 4/(gm/ID)
152
EE 214 Lecture 13B. Murmann 15
Higher Rout
• A cascode can help create a higher output resistance, e.g. to improve CMRR in a differential pair
• Even though the impedance is now high at the current mirror output, we still need V1=V2 to minimize systematic errors in the current ratio I2/I1
2omout rgR ≅
EE 214 Lecture 13B. Murmann 16
Solution 1
• Works, but VOUTmin is very large
– Using long channel algebra, we have
( ) OVttOVtminOUT VVVVVV 22 +=−+≅
• Note that using long channel equations tends to be OK for current mirrors
– Usually operated in strong inversion
– Channel length often not minimum
153
EE 214 Lecture 13B. Murmann 17
Solution 2
• Use some kind of "magic battery" that sets the cascode gate potential such that VOUTmin = 2VOV (minimum possible)
• "High swing" bias
EE 214 Lecture 13B. Murmann 18
Magic Battery 1
• Gate overdrive of ¼ device is twice as large as VOV of all other devices
– Vcas=Vt+2VOV
– VOUTmin = 2VOV
• Sensitive to body effect
154
EE 214 Lecture 13B. Murmann 19
Magic Battery 2
• Less sensitive to body effect
• Stacked device mimics transistor with length 4L
– May not match characteristics of devices with length L all that well
EE 214 Lecture 13B. Murmann 20
Magic Battery 3
• Insensitive to body effect
• (1/3) triode device will not exactly produce VOV
– Typically make device ~1/5 of W, to allow for safety margins and improved Rout
[ ] OVxxx
ttOVxoxOVox VVVVVVVVLWCV
LWCI =⇒⋅⎟
⎠⎞
⎜⎝⎛ −−++=≅
231
21 2
1 µµ
155
EE 214 Lecture 13B. Murmann 21
Magic Battery 4
• No extra current branch
• Needs lots of headroom on input side
• Sensitive to body effect
EE 214 Lecture 13B. Murmann 22
Magic Battery 5
• The previous circuits assumed that Lmirror=Lcascode
– Sometimes want Lmirror≠Lcascode
• The above circuit makes Vcas =Vt2 + VOV2 + VOV1
– VOUTmin becomes VOV2 + VOV1, as desired
I2I1
W1/L1 W1/L1
VcasW2/L2
W2/L2
I1
(2/3)*W1/L1
W1/L1
Vx
I1
W2/L2
156
EE 214 Lecture 13B. Murmann 23
Design Considerations
• In all of the above-discussed circuits, care must be taken to keep the devices forward active with sufficient safety margin, and not to "waste" too much ro
– Typically use ~1/6 in 1/4 size device approach (slide 18)
– Typically use ~1/5 in 1/3 triode device approach (slides 20-22)
• Work with integer ratios and unit devices as much as possible
– Using a 1/5 device really means that we work with a unit device of size 1, and use 5 of them elsewhere
• Keep mirror ratio (I2/I1) reasonably small
– Typically no larger than 10…20
EE 214 Lecture 13B. Murmann 24
Capacitive Coupling
• Can use decoupling capacitors to reduce the amplitude of noise coupling into bias nodes
• If noise is "deterministic" and occurs at the right point in time, you might be better off not decoupling, but making the bias node"fast" so it can recover quickly!
Vx
Low cap:fast recovery
High cap:slow recovery
big bounce
small bounce
t
t
Vx
157
EE 214 Lecture 13B. Murmann 25
Current Distribution
• Typically, we'll only have one single reference current generator on a chip
• Can generate/distribute currents across chip in two different ways
– Distribute gate voltage• Can cause big problems due to IR drop
• Usually limited to local distribution
– Distribute currents• Have one global bias cell close to reference that sends currents
into local biasing sub-circuits
• Disadvantage: Consumes additional current
158
EE 214 Lecture 14B. Murmann 1
Lecture 14Process Variations
Feedback
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 14B. Murmann 2
Overview
• Reading
– 8.0, 8.1, 8.2, 8.3 (Feedback)
– 9.2 (Relation Between Gain and Bandwidth)
– Supplementary Handout "Feedback Systems" by Tom Lee (see web, optional)
• Introduction
– In today's lecture we will first take a look at typical component variations in CMOS technology. In light of these numbers, we then consider negative feedback as a tool for building amplifiers with precisely defined gain. As we shall see, using a forward amplifier with "high", but arbitrary gain, combined with a ratiometric, passive feedback network allows us to build drift and process insensitive precision gain stages.
159
EE 214 Lecture 14B. Murmann 3
PVT
• So far, we've assumed that our Spice model accurately predicts the performance of every single chip we make
• We have also assumed that all our circuits run at room temperature, and VDD is precisely fixed
• In practice, it is the circuit designer's job to ensure that thecircuit works in presence of large variations
– PROCESS: Variations among production lots• "Slow, Nominal and Fast" corners• Sometimes there are even significant variations across wafers
and individual chips
– VOLTAGE: VDD is usually specified only within ±10%• E.g. VDD= 2.7…3.3V in our technology
– TEMPERATURE: Ambient temperature variations• 0…70°C (or -40…125°C)
EE 214 Lecture 14B. Murmann 4
Graphical Interpretation
[Razavi, p. 599]
160
EE 214 Lecture 14B. Murmann 5
Typical Variations
Parameter Lot-to-LotTemperature Coefficient
Device Matching (Area Dependent)
Vt ±100mV -2mV/°C ~1…10mV
µCox ±20% -0.33%/°C ±0.2%
Rpoly2 (50Ω/square) ±20% 0.2%/°C ±0.1%
Rnwell (1k Ω /square) ±40% 1%/°C ±1%
Cpoly-poly2 ±15% -30ppm/°C ±0.1%
EE 214 Lecture 14B. Murmann 6
Consequences
• Performance metrics that depend on absolute component values will show large variations
• Sometimes this is OK
– E.g. bandwidth of a circuit• Can overdesign, and make sure to have minimum required
bandwidth in presence of worst case variations
• Sometimes this is not OK
– E.g. need a precise gain of two in an A/D converter• gmRL or gmro will never be accurate enough
• Solution: Use negative feedback to desensitize the circuit to gmRL or gmro variations
161
EE 214 Lecture 14B. Murmann 7
Negative Feedback
• Harold S. Black, 1927
( )outinout fvvav −=
afa
vv
in
out
+=
1
• Interesting case:
fvvaf
in
out 11 ≅⇒>>
EE 214 Lecture 14B. Murmann 8
Interpretation
• As long as we have "large" gain in the forward path a, the overall gain will depend only on f
• Since vout/vin ~1/f, we often make f ≤ 1
– E.g. for a "closed loop" gain of two, we need f=0.5
• f ≤ 1 is easy to implement, and ratiometric!
– A "wire" (f=1) or resistive or capacitive voltage divider
1≅in
out
vv
1
21
RRR
vv
in
out +≅
162
EE 214 Lecture 14B. Murmann 9
Example
• Case 2: a=1000
5021 .fRR =⇒=
• Case 1: a=100
.....f
avv
in
out 96078150
1001
11
1 =+
=+
=
.....f
avv
in
out 996007150
10001
11
1 =+
=+
=
• 10x variation in forward gain, and only about 1.8% change in closed loop gain!
EE 214 Lecture 14B. Murmann 10
Gain Sensitivity
• Defineaf
aA+
=1
• Can show that
Taa
afaa
AA
+
∆
=+
∆
=∆11
• Fractional change in gain is reduced by product of a and f, which can be made arbitrarily large (conceptually)
– Loop gain T
• We will find that loop gain is a very meaningful parameter that will also appear in bandwidth and impedance calculations
– More later…
163
EE 214 Lecture 14B. Murmann 11
Effect of Negative Feedback on Nonlinearity
• Substitute (2) into (1), then compare coefficients to get
( ) ( ) ( )( ) 3
31
331
2
1
inin
!
out
outinoutinout
vbvbv
fvvafvvav
+=
−+−=
( )41
33
1
11 11 fa
abfa
ab+
=+
=
• Linear term as expected, reduced by (1+T)
• Cubic term reduced by (1+T)4!
EE 214 Lecture 14B. Murmann 12
Negative Feedback and Bandwidth
• Closed loop transfer function:
1
0
1ps
a)s(a.g.E−
=
fapsfa
af)s(a
)s(a)s(A
01
0
0
111
111
+−
⋅+
=+
=
• Bandwidth increases by (1+T)!
– But gain is reduced by (1+T)
• Product of gain and bandwidth remains constant
164
EE 214 Lecture 14B. Murmann 13
Bode Plot Illustration
EE 214 Lecture 14B. Murmann 14
Early Obstacles
• Today we are taking the concept of negative feedback for granted
• At the time of his "invention," Harold Black had a hard time convincing his colleagues that negative feedback was indeed something useful
– It was very hard to make a "high gain" forward amplifier using vacuum tubes
– Why have large gain and then throw it away by applying negative feedback?
• Long before negative feedback had been deemed useful, positive feedback was routinely applied to increase the gain of amplifiers
165
EE 214 Lecture 14B. Murmann 15
Positive Feedback
• Edwin H. Armstrong, 1915
( )outinout fvvav +=
afa
vv
in
out
−=
1
• Examplea
.a
vv.af
in
out 10901
90 =−
=⇒=
• Tenfold increase in gain!
EE 214 Lecture 14B. Murmann 16
Feedback Using Ideal OpAmp
21
1
RRRf+
=
( )−+ −⋅= vvgvout
1
21
21
111 RRR
RRRg
gaf
avv
in
out +≅
++
=+
=
ga =
166
EE 214 Lecture 14B. Murmann 17
Inverting Configuration
• Circuit does not map directly into generic block diagram
– Cannot directly identify a and f
– a ≠ g
?f =
( )−+ −⋅= vvgvout
?a =
EE 214 Lecture 14B. Murmann 18
Superposition
a⇒
af−⇒
167
EE 214 Lecture 14B. Murmann 19
Inverting Configuration
gRR
Ra ⋅+
−=21
2
gRR
Raf ⋅+
−=−21
1
EE 214 Lecture 14B. Murmann 20
Result
2
1
21
2
21
1
RR
gRR
R
gRR
R
aaff −=
⋅+
−
⋅+==g
RRRa ⋅+
−=21
2
1
2
1 RR
afa
vv
in
out −≅+
=
• It can be quite tedious to try and morph arbitrary circuits into a generic "af" block diagram
– Especially when impedances come into play
• Elegant alternative: "Return Ratio Analysis"
– More later…
168
EE 214 Lecture 15B. Murmann 1
Lecture 15Fully Differential OpAmps and OTAs
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 15B. Murmann 2
Overview
• Reading
– 6.0, 6.1.1, 6.1.2, 6.1.2 (Basic Feedback Concepts)
– 12.1, 12.2, 12.3, 12.4 (Fully Differential OpAmps)
– 6.1.7 (Internal Amplifiers)
• Introduction
– Having discussed the basic properties of ideal feedback loops, we will now take steps toward a practical implementation of feedback amplifiers in CMOS technology. The first amplifier we consider uses a very crude gain stage, composed of a simple differential pair and an active current mirror load. For the time being, we will use this crude topology as a vehicle to gain insight into terminology, analysis and design tradeoffs, before moving to more complicated circuits in future lectures.
169
EE 214 Lecture 15B. Murmann 3
A Crude "High-Gain" Amplifier
• Output common mode not well defined
– Depends on mismatch in currents between top and bottom current mirror
EE 214 Lecture 15B. Murmann 4
Common Mode Feedback
• Feedback loop forces Voc=Voc,desired by adjusting tail current
170
EE 214 Lecture 15B. Murmann 5
Fully Differential vs. Single Ended (1)
• Symmetrical
– Helps mitigate parasitic coupling
– Good PSRR
– Easy to analyze
• Twice as much output swing compared to single ended
• Lower complexity
– No CMFB
– Fewer feedback components
• Can build non-inverting unity gain buffer without using any feedback components
EE 214 Lecture 15B. Murmann 6
Fully Differential vs. Single Ended (2)
• Most precision analog integrated circuits are based on fully differential stages
– Data converters, filters, etc.
• In contrast, printed circuit board circuits tend to be single ended
– Want minimum complexity and component count
• In EE214, we will study mostly fully differential circuits
– More to come in EE315
• In all of the upcoming assignments, we will use a very simple, ideal common mode feedback circuit to avoid distraction from the main design task
– Practical CMFB implementation examples will follow later in this course
171
EE 214 Lecture 15B. Murmann 7
Differential Mode Small Signal Half Circuit
EE 214 Lecture 15B. Murmann 8
Load Considerations
• Low load resistance will destroy the gain of our amplifier
– RL may may be an explicit load or due to parasitic loading from feedback network
• Need large gain in feedback loop for good precision
172
EE 214 Lecture 15B. Murmann 9
Solution 1: Buffer
• Can be very difficult to build
• Can cost lots of headroom
• Additional area, power
EE 214 Lecture 15B. Murmann 10
Solution 2: Capacitive Feedback
• Loads amplifier only at high frequencies, DC gain unaffected
• Issue: Charge initialization on caps
– Sometimes possible to use large resistors in parallel
173
EE 214 Lecture 15B. Murmann 11
Switched Capacitor Circuits (1)
During φ1 During φ2
EE 214 Lecture 15B. Murmann 12
Switched Capacitor Circuits (2)
• In switched capacitor (SC) circuits, the charge initialization problem is taken care of
– One phase for charge initialization, a second phase for amplification
• Predominant approach for precision signal processing in CMOS integrated circuits
• SC circuits tend to look very complicated, but always boil down to same design problem
– Build a good amplifier with capacitive feedback network
• More applications in EE315, in EE214 we'll just design the circuit during φ2, without worrying about the switches
174
EE 214 Lecture 15B. Murmann 13
OpAmps versus OTAs (1)
Operational Amplifier Operational Transconductance Amplifier
EE 214 Lecture 15B. Murmann 14
OpAmps versus OTAs (2)
OpAmp
• "General Purpose"
• Ideally a voltage controlled voltage source
• Low output impedance
• Can drive resistive and capacitive loads
• Essentially OTA + buffer
• Buffer increases complexity and power dissipation
OTA
• Most on-chip amplifiers are OTAs
• Ideally a voltage controlled current source
• High output impedance
• Cannot drive resistive loads
• Use capacitive (switched capacitor) feedback
175
EE 214 Lecture 15B. Murmann 15
OTA Gain Bandwidth Product
• GBW is independent of Ro
• Note that CLtot is the sum of all capacitors seen by the output– Including junction caps, etc.
Ltotoom
id
od
CsRRG
vv
+−=
11
Ltot
m
Loom C
GCR
RGGBWππ 211
21 =⋅=
EE 214 Lecture 15B. Murmann 16
OTA Unity Gain Frequency (1)
( )
1
0
11ps
aCsRRG
vvsa
Ltoto
om
id
od
−=
+−==
1pff
( )sa
0a
11
1
0!
p
u
ffj
a =−
uf
1101 =⇒>>
u
ppu f
faff
Ltot
mpu C
GGBWfafπ21
10 ==≅⇒
176
EE 214 Lecture 15B. Murmann 17
OTA Unity Gain Frequency (2)
• The above result holds to good precision when
– There is no second pole before or close to fu– |ao| is large, which implies fu>>fp1
• Important consequence
– No need to include large resistors (such as Ro) for analysis of circuit behavior around fu
1pff
( )sa
0a
uf
∞=oR With
177
EE 214 Lecture 16B. Murmann 1
Lecture 16
StabilityAnalysis of Feedback Circuits
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 16B. Murmann 2
Overview
• Reading
– 9.3 (Stability)
– 8.8 (Return Ratio Analysis)
– Supplementary Handout "Feedback Systems" by Tom Lee (see web, optional)
• Introduction
– This lecture covers basics on the analysis of feedback amplifiers. Using a simple OTA with capacitive feedback as an example, we will study the so called "Return Ratio" method as a tool to assess stability and calculate the bandwidth of feedback amplifiers. Interestingly, most relevant performance metrics follow directly from the "return ratio" or loop gain of the circuit, which highlights the significance of this parameter in feedback system design.
178
EE 214 Lecture 16B. Murmann 3
Stability
• Most general criterion: BIBO
– Bounded input – bounded output
– Applies to any system
• A continuous time linear system is BIBO stable if all its poles are in the left half of the s-plane
– Can calculate roots of 1+T(s) to check stability• Tedious and hard to do in general
)s(T)s(a
)s(f)s(a)s(a
vv)s(A
i
o
+=
+==
11
EE 214 Lecture 16B. Murmann 4
Methods for Checking Stability
• Nyquist Criterion
– Based on evaluating T(s) in a polar plot
– Works for arbitrary T(s)• Even if T(s) itself is unstable
– See books on control theory for details
• Bode Criterion
– A subset of the general Nyquist criterion that can be applied when T(s) itself is stable
• Safe to use in most electronic circuits
• Beware of exceptions
– System is unstable when |T(jω)| > 1 at the frequency where Phase(T(jω)) = -180°
– Can use simple bode plot to check for stability
179
EE 214 Lecture 16B. Murmann 5
Stability Measures
ωc
ω180
ωc
ω180
|T(jω)|
Phase[T(jω)]
( )[ ] cjTPhasePM ωωω =−°=180
( )180
1
ωωω
=
=jT
GM
Typically want GM ≥ 3…5
Typically want PM ≥ 60…70°
EE 214 Lecture 16B. Murmann 6
Closed Loop Peaking
ω/ωc
[Text, p.632]
Closed-loop gain, normalized to 1/f
180
EE 214 Lecture 16B. Murmann 7
OTA with Capacitive Feedback
• Important questions
– Is this circuit stable? What is its phase margin?
– What is the low frequency closed loop gain?
– What is the closed loop bandwidth?
EE 214 Lecture 16B. Murmann 8
Example: Simple Single Stage OTA
• Issues– Feedback network loads amplifier– Amplifier loads feedback network– At high frequencies, there exists a (capacitive) feedforward
path that overrides the amplifier• Cannot be modeled using the above generic feedback block
diagram
?
181
EE 214 Lecture 16B. Murmann 9
Solutions
• If all we needed was the closed loop transfer function, we couldsimply do a KCL/KVL based analysis
– Can be quite tedious, especially for more complex circuits
– Hard to assess stability and stability margin
• Two port feedback analysis
– "Shunt-series, shunt-shunt, series-shunt, series-series" feedback configurations
• See text, sections 8.4, 8.5, 8.6
– Attempts to identify amplifier (a) and feedback network (f) with loading effects included
– Some feedback circuits cannot be modeled using two-ports• E.g. bias circuits that use feedback tend to have only one port
EE 214 Lecture 16B. Murmann 10
Return Ratio Analysis (1)
• Does not attempt to identify forward gain and feedback network transfer functions separately
• Analysis entirely based on looking at gain around feedback loop
– Return ratio, loop gain, a·f, loop transmission• Different terminology for the same thing
• From the loop gain of a circuit, we can determine
– Stability, closed loop gain characteristics, node impedances
• Analysis can be applied to arbitrary feedback circuits, independent of topology, port structure, etc.
• We will first look at the complete framework of this technique
– Then identify a way to partition the analysis for our needs and reduce its algebraic overhead
182
EE 214 Lecture 16B. Murmann 11
Return Ratio Analysis (2)
1. Set all independent sources to zero
2. Break the feedback loop by disconnecting a controlled source in the feedback loop
• E.g. VCCS, VCVS, …
3. Inject a test signal st at the breakpoint
• Current or voltage, depending on type of removed source
4. Find the return signal sr generated by the dependent source that was disconnected from the circuit in step 2.
5. The return ratio (loop gain) of the circuit is given by T=–sr/st
• The text uses the symbol R , we will use T for simplicity and to acknowledge that this quantity is the same as T=af in the ideal feedback block diagram
EE 214 Lecture 16B. Murmann 12
Example
ox vv ⋅= β
xsf
f
CCCC
++=β
⎟⎟⎠
⎞⎜⎜⎝
⎛⋅−=
tott sC
Riv 100
( ) fLtot CCC β−+= 1
( ) ( )[ ]fLom
totm
t
r
CCsRRG
sCRG
iisT
βββ
−++⋅⋅=⎟⎟
⎠
⎞⎜⎜⎝
⎛⋅⋅=−=
1111
00
183
EE 214 Lecture 16B. Murmann 13
Return Factor
• In literature, β is often called "feedback factor"
– We will call it "return factor" to avoid confusion with f of thegeneric feedback block diagram
( ) ( )[ ]fLom CCsR
RGsTβ
β−++
⋅⋅=11
10
Feedback loadingCapacitive feedback divider, includes amplifier input
capacitance Cx
xsf
f
CCCC
++=β
EE 214 Lecture 16B. Murmann 14
Phase Margin
pωω
( )ωjT
0T
cω
omRGT ⋅= β0
( )[ ]fLop CCR β
ω−+
=11
( ) fL
mc CC
Gβ
βω−+
≅1
1>>≅ omp
c RGβωω( )[ ]ωjTPhase
°0
°− 45
°−90
ωpω
( )[ ] °−≅⎟⎟⎠
⎞⎜⎜⎝
⎛−= −
=901
p
cc
tanjTPhaseωωω ωω
°≅°−°≅ 9090180PM
184
EE 214 Lecture 16B. Murmann 15
Closed Loop Bandwidth
• From lecture 14, we know that the bandwidth of the closed loop amplifier is given by
( ) 003 1 TT ppdB ωωω ≅+=−
( )[ ]fLo
omdB CCR
RGβ
βω−+
≅− 13
( )[ ] cfL
mdB CC
G ωβ
βω ≅−+
≅− 13
• Closed loop bandwidth is equal to the unity gain (crossover) frequency of |T(jω)|
EE 214 Lecture 16B. Murmann 16
Remaining Questions
• What is the low frequency closed loop gain of the amplifier?
• How does capacitive feedforward at high frequencies affect the closed loop transfer function?
• Both of these questions could be answered by deriving the closed loop transfer function from first principles (KVL/KCL)
• Alternatively, it is possible to calculate the complete closed loop transfer function based on T(s)
( ))s(T
d)s(T
)s(TAvvsA
i
o
++
+== ∞ 11
• A∞ is the closed loop gain with ideal feedback (Gm→∞)
• d is the direct signal feedthrough with the controlled source removed (Gm→0)
⎟⎟⎠
⎞⎜⎜⎝
⎛+
=af
aff
A :y similaritNote1
1
185
EE 214 Lecture 16B. Murmann 17
Comparison
• Two-Port Analysis
• a, f
• 1/f
• Return Ratio Analysis
• T
• A∞
T1d
T1TAA
++
+= ∞af1
aff1A
+=
(feedforward effects are not modeled)
EE 214 Lecture 16B. Murmann 18
Finding A∞
• With infinite Gm, vx must be zero, and there is no current though Cx
fosi sCvsCv +=0
f
s
CCA −=∞
186
EE 214 Lecture 16B. Murmann 19
Finding d
• Writing a general expression for d will be quite messy
• Better to distinguish the two cases of interest
– Low frequency (Ro<<|1/jωC|)
– High Frequency (Ro>>|1/jωC|)
EE 214 Lecture 16B. Murmann 20
High Frequency Result
• Pole as expected
• Zero at Gm/Cf, typically a very high frequency
– Far beyond ω-3dB if CL is large
• In most cases, calculating feedforward zeros is easier using simple KCL analysis, with large resistors removed
)s(Td
)s(T)s(TA)s(A
++
+= ∞ 11
( ) ( )[ ]m
fL
m
f
f
s
GCCs
GsC
CCsA
ββ−+
+
−−= 1
1
1
187
EE 214 Lecture 16B. Murmann 21
Low Frequency Result
• No feedforward, capacitors are open circuits
• Define static gain error
0
0
0
00 11 T
dT
TAA+
++
= ∞
0
000 1
0T
TCCAd
f
s
+−=⇒=
∞
∞ −=A
AA 0ε ⎟⎟⎠
⎞⎜⎜⎝
⎛−−≅
+−=
+−=−=
∞ 0
0
0
00 11111
111
11T
TT
TAAε
0
1T
≅ε
EE 214 Lecture 16B. Murmann 22
Summary – OTA with Capacitive Feedback
• To find the static gain error– Write an expression for the low frequency loop gain, neglect
all impedances due to capacitors– Note that capacitive voltage dividers still work at low
frequency!
• To find the closed loop bandwidth– Write an expression for the high frequency loop gain, neglect
finite output resistance of all devices– The 3-dB bandwidth of the closed loop circuit is
(approximately) equal to the unity gain frequency of T(jω)
• To assess stability and phase margin– Determine phase of (high-frequency) T(jω) at its unity gain
frequency• Boring for single pole systems, more interesting if two or more
poles are involved…
188
EE 214 Lecture 17B. Murmann 1
Lecture 17Loop Gain Simulation
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 17B. Murmann 2
Overview
• Introduction
– Last lecture, we've seen that return ratio analysis is a useful tool for characterizing feedback amplifiers in terms of stability, bandwidth and static precision. Since we are always interested in verifying our hand analysis results using simulations, it is desirable to have an equivalent method available in Spice. In this lecture, we will discuss the so called Middlebrook method for loop gain simulation. This particular approach helps overcome the issue that ideal loop breakpoints, e.g. at controlled sources, are not directly accessible in Spice.
189
EE 214 Lecture 17B. Murmann 3
References
• H.W. Bode, Network Analysis and Feedback Amplifier Design, Van Nostrand, New York, 1945.
• R.D. Middlebrook, "Measurement of Loop Gain in Feedback Systems," Int. J. Electronics, Vol. 38, No.4, .pp. 485-512, 1975.
• S. Rosenstark, "Loop Gain Measurement in Feedback Amplifiers," Int. J. Electronics, Vol. 57, No.3., pp. 415-421, 1984.
• P.J. Hurst, "Exact Simulation of Feedback Circuit Parameters," Trans. on Circuits and Systems, pp.1382-1389, Nov. 1991.
• P.J. Hurst, S.H. Lewis, "Simulation of Return Ratio in Fully Differential Feedback Circuits," Proc. CICC 1994, pp.29-32.
• Ken Kundert, "A Test Bench for Differential Circuits," Online: http://www.designers-guide.com/Analysis/diff.pdf
EE 214 Lecture 17B. Murmann 4
Circuit Example 1
• What is the loop gain in this circuit?
190
EE 214 Lecture 17B. Murmann 5
Return Ratio Analysis
• As we've seen last lecture, such a hand calculation is pretty straightforward
• How can we simulate T(jω) in Spice?– Using "real" transistor models
( )t
r
iisT −=
EE 214 Lecture 17B. Murmann 6
Spice MOSFET AC Simulation Model
• Nodes of ideal controlled source are not accesible!
– Cannot break loop at gm generator
191
EE 214 Lecture 17B. Murmann 7
Popular Simulation Approach
• Inaccurate
• Hard to estimate mock load
• May get different results for different breakpoints
• Ideally, we'd like to avoid all of the above issues
• Solution: Middlebrook method
( )t
r
vvjT −≅ω
EE 214 Lecture 17B. Murmann 8
Problem Generalization
• Middlebrook argued that any single single loop feedback circuit can be partitioned as shown below
• Hence, there is always some "nonideal" breakpoint between impedances– How can we use this breakpoint to find the loop gain?
available breakpoint
21
21
ZZZZg)s(T m +⋅=
192
EE 214 Lecture 17B. Murmann 9
Double Injection Trick
• No “DC“ break in the loop, all loading effects included!
• Measure Tv and Ti separately, then calculate actual T
1
22 Z
ZZgTvv
mvx
y +⋅=≡−
2
11 Z
ZZgTii
mix
y +⋅=≡
21
21
ZZZZgT m +
⋅=True Loop Gain:
Solving yields:
21++−=iv
iv
TTTTT
iv T11
T11
T11
++
+=
+
EE 214 Lecture 17B. Murmann 10
Simulation Example
• Two options
– Run two copies of the same circuit simultaneously, or
– Run two simulations with different stimuli
x
yv v
vT −=
x
yi i
iT =
193
EE 214 Lecture 17B. Murmann 11
Spice Code
.param ai=1 av=0
m1 vo x 0 0 nch214 L=0.35u W=10u
i1 0 vo 100u
cf vo y 200f
rf vo y 100gig
cs y 0 400f
xt x y looptest ai='ai' av='av'
.op
.ac dec 10 1e3 10e9
.options post brief
.lib './ee214_hspice.txt' nominal
.include utils.sp
.alter
.param ai=0 av=1
.end
* utils.sp
.subckt looptest x y ai=0 av=0
vx middle x dc 0
vy middle y ac 'av'
it 0 middle ac 'ai'
.ends looptest
EE 214 Lecture 17B. Murmann 12
Matlab Postprocessing
% load signals from first run and calculate Ti
m = loadsig('middlebrook1.ac0');
ti = evalsig(m, 'I_xt_vy')./evalsig(m, 'I_xt_vx')
% load signals from second run and calculate Tv
m = loadsig('middlebrook1.ac1');
tv = -evalsig(m, 'y')./evalsig(m, 'x')
f = 1e-6*evalsig(m, 'HERTZ');
% calculate loop gain
t = (tv.*ti-1)./(2+tv+ti)
% calculate magnitude, phase and plot
% ...
194
EE 214 Lecture 17B. Murmann 13
Simulation Result
10-2
10-1
100
101
102
103
104
-40
-20
0
20
40
f [MHz]
Ma
gn
itud
e [d
B]
10-2
10-1
100
101
102
103
104
-100
-50
0
f [MHz]
Ph
ase
[de
gre
es]
TvTiT
TvTiT
EE 214 Lecture 17B. Murmann 14
How to do this in Awaves?
• Generally much more tedious, but some shortcuts make it possible to do this
– Remember that 1/(1+T) = 1/(1+Tv) + 1/(1+Ti)
– Plugging in voltages & currents for Tv and Ti, this becomes 1/(1+T) = vx/(vx-vy) + ix/(ix+iy)
– With unity test voltage and current, this simplifies to1/(1+T) = vx/1V + ix/1A
– Probe real and imaginary parts of vx and ix in HSpice using• .probe ac vxreal = par('vr(x)')
• .probe ac vximag = par('vi(x)')
• .probe ac ixreal = par('ir(xt.vx)')
• .probe ac iximag = par('ii(xt.vx)')
• Can now create expressions for phase and magnitude of T and plot using real and imaginary components of vx and ix
195
EE 214 Lecture 17B. Murmann 15
Circuit Example 2
Cs
+vid-
+vod-
Cs
Cf
Cf
CL
CL
EE 214 Lecture 17B. Murmann 16
Fully Differential Testbench
xmxp
ymypv vv
vvT
−−
−=xmxp
ymypi ii
iiT
−−
=
196
EE 214 Lecture 17B. Murmann 17
Differential Looptest Circuit
.subckt difflooptest xp xm yp ym ai=0 av=0
vxp middlep xp dc 0
vyp middlep yp ac 'av'
vxm middlem xm ac 'av'
vym middlem ym dc 0
it middlem middlep ac 'ai'
.ends difflooptest
EE 214 Lecture 17B. Murmann 18
Simulation Result
10-2
10-1
100
101
102
103
104
-40
-20
0
20
40
f [MHz]
Ma
gn
itud
e [d
B]
10-2
10-1
100
101
102
103
104
-100
-50
0
f [MHz]
Ph
ase
[de
gre
es]
TvTiT
TvTiT
197
EE 214 Lecture 17B. Murmann 19
Alternative Approach
• Use baluns to convert differential signals into CM/DM components and use simple single ended loop test circuit
• The above setup is particularly for use with Spectre simulator– Spectre has built-in Middlebrook analysis, called "stb"– Simply replace looptest block with "iprobe" element– Spectre automatically finds and plots T, no post-processing
needed
EE 214 Lecture 17B. Murmann 20
xfmr
xfmr
Ideal Balun
• Useful for separating CM and DM signal components
• Bi-directional, preserves port impedance
• Uses ideal, inductorless transformers that work down to DC
• Not available in all simulators (e.g. not available in LTSpice)
= .subckt balun vdm vcm vp vm
e1 vp vcm transformer vdm 0 2
e2 vcm vm transformer vdm 0 2
.ends balun
198
EE 214 Lecture 17B. Murmann 21
+ New Middlebrook Method
• Recently, Middlebrook came up with an alternative way of looking at feedback circuits
– A more "design oriented analysis"
• He titled this approach "The General Feedback Theorem – A final solution for feedback systems"
• If you are curious about this new method, please refer to
– http://rdmiddlebrook.com
– http://ardem.com/free_downloads.asp
EE 214 Lecture 17B. Murmann 22
Multi Loop Considerations
• Any practical feedback circuit has multiple feedback loops
– Fully differential circuits have CM/DM loops
– Local device feedback through Cgd, Rsource
– ...
• Solutions
– Decompose fully differential circuit into CM/DM loops
– If a local feedback loop can be modeled as a combination of a stable controlled source and passive impedances, the multi-loop circuit reduces to a single loop [Hurst 94]
– If there is a common breakpoint that breaks all feedback loops simultaneosly, stability can be checked by finding the return ratio at the single breakpoint [Hurst 94]
199
EE 214 Lecture 17B. Murmann 23
Last Resort: General Nyquist Criterion
[Bode 45]:
“If a circuit is stable when all its tubes have their nominal gains, the total number of clockwise and counterclockwise encirclements of the critical point must be equal to each other in the series of Nyquist diagrams for the individual tubes obtained by beginning with all tubes dead and restoring the tubes successively in any order to their nominal gains“
[You may want to take a controls class if you are interested in this...]
EE 214 Lecture 17B. Murmann 24
[Bode 45]:
“... thus the circuit may sing when the tubes begin to lose their gain because of age, and it may also sing, instead of behaving as it should, when the gain increases from zero as power is supplied to the circuit...“
Another Useful Quote
Always run one or more transient analyses for a "true" stability check!
200
EE 214 Lecture 18B. Murmann 1
Lecture 18Differential Mode Voltage Range
Two-Stage OTA
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 18B. Murmann 2
Overview
• Reading
– 6.3 (Basic Two-Stage MOS Amplifiers)
– 12.6.1 (Fully Differential Two-Stage-Amplifier)
– 12.6.5 (Neutralization)
• Introduction
– As we will see in today's lecture, the available output swing and open-loop gain of our simple single stage OTA are fairly limited. Hence, we will begin to consider a two-stage OTA architecture as an alternative. Unfortunately, with the addition of a second stage, we also introduce a second pole, which makes it difficult to achieve reasonable phase margin. The compensation techniques discussed in the following lecture will provide solutions to this problem.
201
EE 214 Lecture 18B. Murmann 3
Output Swing of Simple OTA
• Available output swing depends on input and output common mode levels
• May be limited by headroom for differential pair device (Vminn) or active load (Vminp)
EE 214 Lecture 18B. Murmann 4
Maximum Available Swing
• Input and output common mode adjusted such that all devices operate at "edge" of forward active region
– Well defined using long channel model, very gradual transition in short channels
• Unfortunately, the choice of Vic and Voc are often dictated by the circuits that interface with the amplifier
– E.g. Vic=Voc=1.5V( )tminnminpminDDmax,odpp VVVVV −−−= 2
202
EE 214 Lecture 18B. Murmann 5
Example Vic=Voc=VDD/2
• Assuming that we are limited by Vminn, and Vminn~VOV, the available differential peak-to-peak swing is ~4Vt
• Since the transition to triode is smooth, which criterion shouldwe use find the "exact" output range of an amplifier?
EE 214 Lecture 18B. Murmann 6
Gain vs. Output Swing DC Simulation
• In EE214, we arbitrarily define output range as the peak-to-peak swing that causes no more than 30% drop in Vod/Vid
-1.5 -1 -0.5 0 0.5 1 1.540
50
60
70
80
90
Vod
[V]
Vod
/Vid
[V/V
]
-30%
Vodpp,max
203
EE 214 Lecture 18B. Murmann 7
How Much Gain Can We Get?
• Small signal gain (around Vid=Vod=0):
vop
von
mn
mpvon
op
onvon
opon
oponmnvo
aa
gga
rra
rrrr
ga+
=+
=+⋅
⋅=1
1
1
1
( )( ) vop
von
nDm
pDmvonvo
aa
I/gI/gaa
+=
1
1
vonvonvo a||aa = ( ) ( )nDmpDm I/gI/g for =
• E.g. avon=avop=50, (gm/ID)n= (gm/ID)p ⇒ avo=25
• Static gain error ~1/To ~1/avo ~1/25=4%
– Not precise enough for many applications
EE 214 Lecture 18B. Murmann 8
Two-Stage Amplifier
• More output swing, more gain ~(gmro/2)2
• Output range no longer depends on input common mode
204
EE 214 Lecture 18B. Murmann 9
Common Mode Feedback
• Same as for single stage OTA!
• Common mode of first stage output is set via VGS of common source device in second stage
EE 214 Lecture 18B. Murmann 10
Simplified AC Half Circuit with Feedback
( ) ( ) ( ) ( )sasasa
ps
ps
RgRgsT mm ⋅=⋅=
⎟⎟⎠
⎞⎜⎜⎝
⎛−⋅⎟⎟
⎠
⎞⎜⎜⎝
⎛−
⋅= βββ 21
21
2211
11 111
1CR
p −=22
21CR
p −=
205
EE 214 Lecture 18B. Murmann 11
How About Miller Effect?
• Two ways to deal with Miller amplification of Cgd in first stage– Use cascodes
• Often needed for high gain, anyways
– Use "neutralization caps"• See text, section 12.6.5
EE 214 Lecture 18B. Murmann 12
Bode Plot of Loop Gain
• If ωp1 and ωp2 are close to each other, the loop will essentially have no phase margin!
ω
1pω
°0
°−90
°−180 ω
2pω
( )ωjMag
( )ωjPhase
( )sa1
( )sa⋅β
( )sa2
( )sa⋅β
206
EE 214 Lecture 18B. Murmann 13
Introducing a Dominant Pole
• The problem is solved if we somehow manage to make ωp1<< ωp2
– Or ωp2<< ωp1
• Loop behaves close first order system around crossover frequency
ω
1pω
°0
°−90
°−180 ω
2pω
( )ωjMag
( )[ ]ωjTPhase
( )sa1
( )sa⋅β
( )sa2
cω
EE 214 Lecture 18B. Murmann 14
Phase Margin
• At the crossover frequency, the dominant pole has shifted the phase by about -90°
• The non-dominant pole's phase at ωc is given by -tan-1(ωc/ωp2)
⎟⎟⎠
⎞⎜⎜⎝
⎛−°−°≅ −
2
190180p
ctanPMωω
⎟⎟⎠
⎞⎜⎜⎝
⎛≅ −
c
ptanPMωω 21
ωp2/ωc PM
1 45°2 63°3 72°4 76°5 79°
207
EE 214 Lecture 18B. Murmann 15
Creating a Dominant Pole
• Numerical example:
mSGG mm 121 == Ω== kRR 10021 pFC 12 =
MHz.CR
f p 612
1
222 ==
π
°= 72PM
kHzf
f pc 530
32 == Hz
RGRGff
mm
cp 106
22111 =
⋅⋅=β
50.=β
nFRf
Cp
152
1
111 =
⋅⋅=
π
• Two issues
– Very low fc, which means low closed loop bandwidth
– Huge capacitor• Get roughly 1fF/µm2 in CMOS technology
• C1 would occupy about 4mm x 4mm !
EE 214 Lecture 18B. Murmann 16
Utilizing the Miller Effect
• Purposely connect an additional capacitor between gate and drain of M2 (Cc = "Compensation capacitor")
• Two interesting things happen
– Low frequency input capacitance of second stage becomes large – exactly what we need for low ωp1
– At high frequencies, Cc turns M2 into a diode connected device – low impedance, i.e. large ωp2 !
208
EE 214 Lecture 18B. Murmann 17
Pole Splitting
• More analysis next lecture…
c
209
EE 214 Lecture 19B. Murmann 1
Lecture 19Compensation
Noise in Feedback OTAs
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 19B. Murmann 2
Overview
• Reading
– 9.4.1, 9.4.2, 9.4.3 (Compensation)
• Introduction
– In this lecture, we will continue to look at pole splitting as amethod for achieving sufficient phase margin in a two-stage amplifier. While using a Miller capacitance for compensation helps move the non-dominant pole to higher frequency, it also introduces an undesired zero in the transfer function. Today, we'll discuss several options on how to cope with this artifact.
– In addition, we will derive useful expressions for the total integrated noise in OTAs.
210
EE 214 Lecture 19B. Murmann 3
Two-Stage OTA with Cc
• Detailed analysis shows:
( ) ( ) ( )[ ] ( )1c2c21212
c122m1c12c2
2m
c22m11m
i
o
CCCCCCRRsCRRgRCCRCCs1gCs1RgRg
vvsa
++++++++
⎟⎟⎠
⎞⎜⎜⎝
⎛−⋅⋅
==
• Very messy…
EE 214 Lecture 19B. Murmann 4
Dominant Pole Approximation
• We can write the denominator as
( )21
2
2121
11111pp
spp
sps
pssD +⎟⎟
⎠
⎞⎜⎜⎝
⎛+−=⎟⎟
⎠
⎞⎜⎜⎝
⎛−⋅⎟⎟
⎠
⎞⎜⎜⎝
⎛−=
• Since in a practical design outcome we'll have |p1|<<|p2|, we can approximate
( )21
2
1
11pp
sp
ssD +⎟⎟⎠
⎞⎜⎜⎝
⎛−≅
• With this simplification, we can now easily identify p1 and p2 by comparing the coefficients with the expression from the previousslide
211
EE 214 Lecture 19B. Murmann 5
Final Result
• Questions
– How can we design an amplifier with these complex expressions?
– What will the zero in the transfer function do to us?
( )⎟⎟⎠
⎞⎜⎜⎝
⎛−⋅⎟⎟
⎠
⎞⎜⎜⎝
⎛−
⎟⎠⎞
⎜⎝⎛ −
⋅≅
21
0
11
1
ps
ps
zs
asa v
c
m
Cgz 2+=
( ) ( ) cmcmcc CRRgCRRgCCRCCRp
12212222111
11 −≅++++
−≅
2121
22
CCCCC
gp
c
m
++−≅
EE 214 Lecture 19B. Murmann 6
RHP vs. LHP Zero
Phase
°0
°− 45
°−90
ωzω
zω zω−
Phase
°0
°+ 45
°+ 90ωzω
212
EE 214 Lecture 19B. Murmann 7
OTA Transfer Function with RHP Zero
• RHP zero will reduce phase margin
– Unless gm2 >> βgm1
c
mc C
g 1βω ≅
c
mz C
g 2=ω
1
21
m
m
c
z
gg
βωω =
ωz
(assuming zero is beyond crossover)
EE 214 Lecture 19B. Murmann 8
Mitigating the Impact of RHP Zero
• Create unilateral feedback through Cc
– Source follower
– Cascode compensation• Ahuja, IEEE J. Solid-State Ckts., 12/1983
• Ribner, IEEE J. Solid-State Ckts., 12/1984
• "Nulling resistor"
– Push zero to infinity
– Push zero into LHP and cancel nondominant pole (!)
213
EE 214 Lecture 19B. Murmann 9
Source Follower
• Mitigates feedforward path issue
• Problems: Reduced output swing, additional power dissipation
EE 214 Lecture 19B. Murmann 10
Cascode Compensation (1)
• Cc sees low impedance, output of first stage sees high impedance looking into cascode device
– Reduced feedforward
214
EE 214 Lecture 19B. Murmann 11
Cascode Compensation (2)
• Benefits
– Tends to push ωp2 to higher frequencies, when load capacitor is large (see text)
• Can use smaller Cc, less power in first stage
• Issues
– Additional power dissipation
– Bias current mismatches cause input referred offset
• The above two issues can be addressed by feeding back to cascode device embedded in first stage (Ribner, JSSC 12/1984)
– New issue: Complex design problem (3rd order system)
EE 214 Lecture 19B. Murmann 12
Nulling Resistor (1)
• New transfer function becomes
( )⎟⎟⎠
⎞⎜⎜⎝
⎛−⋅⎟⎟
⎠
⎞⎜⎜⎝
⎛−⋅⎟⎟
⎠
⎞⎜⎜⎝
⎛−
⎟⎟⎠
⎞⎜⎜⎝
⎛−−
⋅≅
321
20
111
11
ps
ps
ps
Rg
sCasa
zm
c
v
• p1 and p2 unchanged, new pole p3, and a knob to tune the zero
215
EE 214 Lecture 19B. Murmann 13
Nulling Resistor (2)
• Rz=1/gm2 pushes the zero to +∞– Can use a transistor in triode region to implement resistor
• Helps track process variations
EE 214 Lecture 19B. Murmann 14
Nulling Resistor (3)
• Rz=(1+C2/Cc)/gm places zero on top of ωp2!– Cancels p2
– Good in theory, may be troublesome in practice– If the pole and zero don't fall exactly on top of each other, we
get a so-called pole-zero doublet• Can cause very slow settling tails in transient response• See supplementary handout "Effect of Doublet on Amplifier
Settling Time" (optional)
• Third pole
1
2
13
1Cg
CRm
zp ≅≅ω
11
11
23 >>= typically CC
gg c
m
m
c
p
βωω
216
EE 214 Lecture 19B. Murmann 15
Other Compensation Methods
• Nested Miller compensation
– >2 gain stages
– Higher order response presents design challenge
– Not (yet?) used much
– Ref: R. G. H. Eschauzier and J. H. Huijsing. Frequency Compensation Techniques for Low-Power Operational Amplifiers. Kluwer, 1995.
• Lag/lead compensation
– See handout "Feedback Systems" on website
– An attempt to improve bandwidth/phase margin by adding additional zeros to T(s)
– May introduce doublets and worsen noise performance
EE 214 Lecture 19B. Murmann 16
Noise in Feedback OTAs
• Single stage amplifier
• Single stage amplifier with cascode
• Two stage amplifier
217
EE 214 Lecture 19B. Murmann 17
Single Stage Amplifier
CL+vi-
+vo-
M1
M2VB
Cf
Cs
⎟⎟⎠
⎞⎜⎜⎝
⎛+⎟
⎟⎠
⎞⎜⎜⎝
⎛++=
⎟⎟⎠
⎞⎜⎜⎝
⎛+=
∞1m
2m
f
1gg
1m
2m2tot,o
gg1
CC
A1CkT
gg1
CkTv
γ
βγ
• Make gm2 as small as possible, i.e. use small gm/ID for current source device
– Issue: Smaller gm/ID means less available swing
• Small Cgg1, i.e. high fT helps reduce noise
( )β−++= 1CCCC fjunctionL
EE 214 Lecture 19B. Murmann 18
Useful Integrals
218
EE 214 Lecture 19B. Murmann 19
Single-Stage Amplifier with Cascode
• Analysis shows
⎟⎟⎠
⎞⎜⎜⎝
⎛+=
⎟⎟⎠
⎞⎜⎜⎝
⎛+=
1m
2m
2p
c
1m
2m2tot,o
gg
k111
CkT
gg11
CkTv
β
ωω
β
• Make gm2 as small as possible, i.e. use small gm/ID for cascode device
– Reduces gm2/gm1 and Cx
– Issue: Smaller gm/ID means less available swing
Ltot
1mc C
gβω =
x
2m2p C
g=ω
( )β−++= 1CCCC fjunctionLc
2pkωω
=
EE 214 Lecture 19B. Murmann 20
Two-Stage Amplifier
• Analysis shows
– See supplementary handout "Noise in a Two-Stage Amplifier"
• Need large Cc for low noise
• Noise from second stage can be significant if CL is small and β is large
Noise from second stage
≅ 1
219
EE 214 Lecture 19B. Murmann 21
SNR in Differential Circuits
• In fully differential circuits, the effective output swing is doubled
– But there are two half circuits that contribute noise
• Signal power increases by 4x, noise power increases by 2x
– 3dB win in terms of SNR
– At the cost of twice the power consumption (no free lunch…)
CkTVSNR
2o∝
220
EE 214 Lecture 20B. Murmann 1
Lecture 20OTA Design Considerations
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 20B. Murmann 2
Overview
• Reading
– Supplementary handout "OTA Design Example"
• Introduction
– Today, we will review a possible design strategy for two-stage OTAs. Since there exist many degrees of freedom in this topology, a thorough hand analysis is essential, and can help minimize, if not completely eliminate, the time needed for a large number of Spice iterations.
221
EE 214 Lecture 20B. Murmann 3
Two-stage OTA Design Strategy
• The design equations for a two-stage OTA are fairly complex, and involve several capacitances that may not be known a priori
• Solution: Iterative hand analysis/design– Most efficiently done using e.g. Excel, Matlab or MathCAD
• Possible design flow1. Pick Cc based on noise spec, and Cgg1, Cgg2 based on
heuristics (see following slides) 2. Use initial guesses for junction capacitances
– Conservative starting point: Cjunction≈Cgs≈Cgg
3. Find gm, fT, gm/ID4. Iterate, using different choices in step 15. Find W for Spice verification6. Resolve potential discrepancies
– E.g. refine Cjunction estimates if necessary
EE 214 Lecture 20B. Murmann 4
A Note on Discrepancies
• Discrepancies between design script and Spice are usually on the order of 10%
– Mostly due to• Bias point dependence (VDS); charts generated for VDS=VGS
• Inaccuracy in Cjunction estimate
• First order nature of design equations
• Good news
– It is always possible to track discrepancies down if needed• Look at gm/ID , CGG, Cjunction in the bias point output
• Big difference to square law design using µCox, VOV, …
– These quantities simply don’t exist in your circuit…
222
EE 214 Lecture 20B. Murmann 5
Component Identification
1junction2gg1 CCC +=
( ) 2junctionfL2 CC1CC +−+= β 1ggsf
f
CCCC++
=β
• For algebraic convenience, we'll neglect the junction caps in the following discussion…
– Straightforward to add back in in your spreadsheet
EE 214 Lecture 20B. Murmann 6
Closed-Loop Bandwidth
• Small Cgg1 helps, but there is diminishing return if becomes small compared to Cs, Cf
• Typically a good starting point: Cgg1 ≈ Cs
– Sometimes optimum
1ggsf
f
CCCC++
=β
c
mmm
cmcdB C
gRgRgCRRg
12211
1223
1 ββωω =⋅⋅≅≅−
223
EE 214 Lecture 20B. Murmann 7
Nondominant Pole
• Heuristic: Can try to make Cgg2 ≅ CLtot
– Choosing Cgg2 much smaller may mean that we need large gm2/Cgg2=ωT2 and therefore get small gm/ID
– Choosing Cgg2 much larger may cost too much power with diminishing returns
Ltotggc
ggLtot
mp
CCCCC
g
++≅
22
22
ω ( ) fLLtot CCC β−+≅ 1
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛+
+⎟⎟⎠
⎞⎜⎜⎝
⎛+≅
c
Ltot2gg
2ggLtot
2m
Ltot
2m
2gg
2p CCC
CC
1gC
gC1
ω
EE 214 Lecture 20B. Murmann 8
Outcome
• Interesting to look at Cgg2=CLtot (not necessarily optimum) to get further qualitative insight
– With this choice, we have
⎟⎟⎠
⎞⎜⎜⎝
⎛+≅
c
Ltot
2T2p C2C121
ωω
c
mcdB C
g 13 βωω ≅≅−
• Cc large means that we can use lower ωT2, i.e. save power in the second stage
– But larger Cc also means more gm and thus more power in the first stage
• Seems like there may be an optimum value for Cc
– But Cc may ultimately be constrained by noise requirements…
224
EE 214 Lecture 20B. Murmann 9
225
EE 214 Lecture 21B. Murmann 1
Lecture 21Step Response
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 21B. Murmann 2
Overview
• Reading
– 7.5 (Relation Between Frequency and Time Response)
• Reference
– H.C. Yang and D.J. Allstot, "Considerations for fast settling operational amplifiers," IEEE Trans. Ckts. and Syst., March 1990, pp. 326 – 334.
• Introduction
– OTAs with capacitive feedback are primarily used in switched capacitor circuits, where fast settling to voltage steps at the input is critical. In today's lecture we will look at first and second order behavior in the step response of feedback OTAs.
226
EE 214 Lecture 21B. Murmann 3
Motivation
• In switched capacitor circuits, the amplifiers have to respond to voltage steps
• How fast of a switched capacitor circuit can we build?
EE 214 Lecture 21B. Murmann 4
Analysis
• Assuming a single pole system, we have
( )
c
0
0
f
s
in
out
s1
1T1
TCC
)s(V)s(VsA
ω+
⋅+
−=≅
Ltot
mc C
G⋅≅ βω
insf
f
CCCC
++=β
omRGT ⋅= β0
( ) fLLtot CCC ⋅−+= β1
227
EE 214 Lecture 21B. Murmann 5
Step Response
)s(V)s(A)s(V inout ⋅=
)s(V)s(AL)t(V in1
out ⋅= −
( )τ/t
0
0step
f
sstep1out e1
T1TV
CC
sV
)s(AL)t(V −− −⋅+
⋅⋅−=⎭⎬⎫
⎩⎨⎧
⋅=c
1ω
τ =
Ideal Response
Static Error
Dynamic Error
0ε⇒ dε⇒
EE 214 Lecture 21B. Murmann 6
Graphical Illustration
0 2 4 6 8 100
0.2
0.4
0.6
0.8
1
t/τ
Vou
t/Vou
t,ide
al
Dynamic Error
Static Error
228
EE 214 Lecture 21B. Murmann 7
Design Considerations (1)
• Need large DC loop gain for small static error
– |ε0| ~ 1/T0
– E.g. need T0>1000 for better than 0.1% precision
• Need small τ (large bandwidth) for fast settling
• Can define settling time ts based on tolerable dynamic error
τε /ttol,d
se−−=−
( )tol,ds lnt ετ ⋅−=
( )tol,dc
s ln1t εω
⋅−=
EE 214 Lecture 21B. Murmann 8
Design Considerations (2)
• Going from 1% dynamic precision to 10-6 necessitates only ~3x increase in settling time
εd,tol ts/τ
1% 4.6
0.1% 6.9
0.01% 9.2
10-6 13.8
229
EE 214 Lecture 21B. Murmann 9
Design Considerations (3)
ε fc/fCLK
1% 1.5
0.1% 2.2
0.01% 2.9
10-6 4.4
• A switched capacitor circuit operates in two clock phases
• Fitting the required number of time constants within ½ period lets us relate fCLK to a minimum bandwidth requirement
( )CLK
max,dc
s f1
21ln
f21t <⋅⋅
−= επ
( )max,dCLK
c ln1ff ε
π−>
EE 214 Lecture 21B. Murmann 10
How Fast Can We Go?
• fc cannot be larger than about 1/3 of the nondominant amplifier pole frequency (stability)
– In a cascoded amplifier, the nondominant pole occurs at a frequency around fT/3
• Assuming that two junction caps equal to Cgg are present
– At a reasonable bias, the NMOS transit frequency in our technology is roughly 8GHz (nominal process and temperature)
• Putting all of this together, and assuming that we'll need to settle to within 0.01% (~9 time constants), we have
MHz26630f
9f
9.21f TT
max,CLK =≅⋅=
230
EE 214 Lecture 21B. Murmann 11
Practical Aspects
• In practice, it is hard to achieve fCLK> fT/50
– Amplifier topology restriction• E.g. sometimes forced to use PMOS in signal path
– Restrictions on power dissipation• Near the technology limits, power tends to grow out of bounds
– Timing overhead to produce non-overlapping clocks• Have somewhat less than half clock cycle to settle
– Other transient effects• We'll look at some of those next…
EE 214 Lecture 21B. Murmann 12
Impact of Non-Dominant Pole
[Yang, IEEE TCAS 3/1990]
231
EE 214 Lecture 21B. Murmann 13
Simulation Example
• Using simple single stage (single pole) OTA
• Parameters– Cs=Cf=500fF, CL=10pF, β=0.48, Gm=1mS, GmRo=85, Vidstep=-10mV
EE 214 Lecture 21B. Murmann 14
Result
( )ns21
GC1C1
m
fL =−+
=β
βτ mV76.9
RG1RGVV
0m
0midstepfinal,od =
⋅+⋅−=β
β
0 50 100 150 200 250 3000
2
4
6
8
10
Time [ns]
Vo
ltag
e [m
V]
Vid
Vod
(simulation)V
od (theory)
232
EE 214 Lecture 21B. Murmann 15
Another Run
• Changed CL from 10pF to 300fF
• What's this ?
5 10 15 20 25-5
0
5
10
Time [ns]
Vo
ltag
e [m
V] V
idV
od (simulation)
Vod
(theory?)
EE 214 Lecture 21B. Murmann 16
Capacitive Feedforward
• In the first instant after the input step has been applied, the output is completely determined by capacitive voltage division
• Half circuit during initial transient:
Lf
f
Lf
Lfins
s
idstep
odstep
CCC
CCCC
CC
CVV
+⋅
+++
=
233
EE 214 Lecture 21B. Murmann 17
Analysis
• Can analyze this effect in two (equivalent) ways
– Using capacitive divider to find new starting point of exponential
– Using inverse Laplace transform of A(s) with feedforward zero included
• Recall from lecture 14, that A(s) is more precisely given by
( )0
0
f
sT1
T
ps1zs1
CCsA
+⋅
−
−−= f
m
CGz =
( ) fL
m
C1CGpβ
β−+
−=
EE 214 Lecture 21B. Murmann 18
New Result
• For our example:
⎟⎟⎠
⎞⎜⎜⎝
⎛⎥⎦⎤
⎢⎣⎡ −−⋅
+⋅⋅−=
⎭⎬⎫
⎩⎨⎧
⋅= −− τ/t
0
0idstep
f
sstep1od e
zp11
T1TV
CC
sV
)s(AL)t(V
New
( )( ) ( )
Lf
ffL
fL
fL
ffL
CCC
1
1C1C
CCC1C
CC1Czp1
+−
=−++
=−+
+−+=−
βββββ
( ) mV44.11mV10)0t(V4.1
fF300fF500fF50048.01
1od −=−≅=⇒=
+−
• Good agreement with simulation
234
EE 214 Lecture 21B. Murmann 19
New Settling Time
• Settling time for given precision increases due to feedforward, since the settling range is artificially enlarged
• E.g. in our simulation example, the time to settle within 0.1% dynamic error increases from 6.9τ to 7.3τ– Not extremely significant, especially when β is low and CL is
at least comparable to Cf
⎟⎟⎠
⎞⎜⎜⎝
⎛
⎥⎥⎦
⎤
⎢⎢⎣
⎡
+⋅−⋅−=
Lf
ftol,d
cs CC
C1ln1t βε
ω
<1
235
EE 214 Lecture 22B. Murmann 1
Lecture 22Slewing
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 22B. Murmann 2
Overview
• Reading
– 9.6.1, 9.6.2, 9.6.5 (Slew Rate)
• Introduction
– Today we'll complete our discussion of transient behavior in OTA circuits with capacitive feedback. Aside from the feedforward artifact we've discovered last time, there exists another effect called "slewing". Whenever the differential input pair is driven outside its "linear range", the differential output current is limited by the available tail bias. In typicalswitched capacitor circuits, slewing can cause a significant speed reduction, and is therefore an important effect to consider.
236
EE 214 Lecture 22B. Murmann 3
Simulation Example from Last Lecture
• Using simple single stage OTA (See lecture 15, slide 3)
• Parameters– Cs=Cf=500fF, CL=10pF, β=0.48, Gm=1mS, GmRo=85
EE 214 Lecture 22B. Murmann 4
Another Simulation
• Set Vidstep=-1V (CL=10pF ⇒ insignificant feedforward to output)
• What causes this discrepancy ?
0 50 100 150 200
0
200
400
600
800
1000
Time [ns]
Vo
ltag
e [m
V]
-Vid
Vod
(simulation)V
od (theory?)
237
EE 214 Lecture 22B. Murmann 5
Capacitive Divider at OTA Input
• Half circuit during initial transient:
mV480fF500fF40fF500
fF500V1
CCCC
CC
CVV
Lf
Lfins
sidstepxdstep −=
++−≅
+++
=
• Initially -480mV across differential pair input!
EE 214 Lecture 22B. Murmann 6
Differential Pair Characteristics
• Differential output current begins to saturate ~|Vid| > 1.4·2/(gm/ID)
• Beyond this point, current will be much less than that predictedby linear model (slope at origin)
Iod/ITAIL
⎝
2 1 0 1 2
1
0
1Slope = Gm
Vid/(2/[gm/ID])-√2 √2
238
EE 214 Lecture 22B. Murmann 7
0 50 100 150
-500
-400
-300
-200
-100
0
Time [ns]
Vxd
[mV
]
0 50 100 150-300
-200
-100
0
Time [ns]
Diff
. pa
ir I od
[ µA
]
Differential Pair Input Voltage vs. Output Current
-1.4·2/gm/ID)
"Slewing" "Linear Settling"
EE 214 Lecture 22B. Murmann 8
Slewing
• During "slewing", the amplifier drives its output with a constant current (equal to tail bias)
• The slewing behavior ends when |Vid| has become smaller than about 1.4·(2/gm/ID)
– This is the point when the differential pair re-enters its "linear region"
– Hence, the remaining portion of the settling is often called "linear settling"
• Even though the output voltage really settles with a (1-et/τ) relationship
• The total settling time of the amplifier in presence of slewing can be calculated as shown in the following derivation
239
EE 214 Lecture 22B. Murmann 9
Slew Rate
• In order to find the time it takes to complete slewing, we can first calculate the "ramp speed" at which the output changes
– This quantity is called "Slew Rate" (SR)
( ) fL
TAIL
Ltot
TAILod
C1CI
CI
dtdVSR
β−+===
EE 214 Lecture 22B. Murmann 10
Slewing Time
• The input of the differential pair changes at a rate equal to β·SR, where β is given by the usual capacitive feedback divider
• Hence, the time it takes to complete slewing is given by
( )SR
I/g/8.2Vt Dmxstepslew ⋅
−=
β
• In our example, we have
sV20
pF10A200
CISR
Ltot
TAIL
µµ =≅=
ns21
sV2048.0
mV280mV480tslew =⋅
−=
µ
240
EE 214 Lecture 22B. Murmann 11
Subsequent Linear Settling
• Once slewing is completed, the differential output voltage is
mV420SRtV slewslew,od =⋅=
• The final settling value in our example is roughly 1V
– Almost half way there after slewing
• This means that the dynamic error budget for the remaining settling portion has increased
– E.g. if we wanted to settle within 0.1% of the final value (~1V), we only need to complete the remaining transient to within 0.1%·1V/0.58V=0.17%
– Not a very big win, usually a negligible change in the number of required time constants
• 0.1%→6.9τ, 0.17% → 6.4τ
EE 214 Lecture 22B. Murmann 12
Complete Expression for Settling Time
• Note that circuits with large closed loop gain tend to slew less
– Since Vidstep cannot be larger than Voutputswing/Gain
– E.g. Voutputswing=2V, Gain =8 ⇒ Vxdstep < Vidstep < 250mV• Won’t slew at all if gm/ID < 2.8/250mV= 11.2V-1
( ) ( )tol,dDmxstep
linslews lnSR
I/g/8.2Vttt ετ
β−
⋅−
≅+=
( ) fL
TAIL
Ltot
TAIL
C1CI
CISR
β−+==
fins
sidstep
Lf
Lfins
sidstepxdstep CCC
CV
CCCC
CC
CVV++
≅
+++
=
insf
f
CCCC
++=β
<1
241
EE 214 Lecture 22B. Murmann 13
Slewing in a Two-Stage OTA (1)
• Nodes V1p and V1m stay roughly constant
– These nodes move at a rate equal to the slew rate Vop, Vom
divided by ~ intrinsic gain of second stage
– Second stage acts as "integrator"
EE 214 Lecture 22B. Murmann 14
Slewing in a Two-Stage OTA (2)
• Must design circuit such that slew rate is limited by ITAIL
– IB2 limitation would cause asymmetric slewing• One branch slew limited by IB2, other slew limited by ITAIL
– Asymmetric slewing causes CM shift• Causes slow transients due to slow CMFB
242
EE 214 Lecture 22B. Murmann 15
Slewing in a Two-Stage OTA (3)
• The maximum slew rate at which output Vom can move down
( )⎭⎬⎫
⎩⎨⎧
+=
−=−
Ltotc
2B
c
TAILp1omMAX, CC
2/I,C
2/Imindt
VVdSR
• The maximum slew rate at which output Vop can move up
c
TAILm1opMAX, C
2/Idt
)VV(dSR =
−=+
• To make slew rates equal, we need
Ltotc
2B
c
TAILCC
IC
I+
≤
EE 214 Lecture 22B. Murmann 16
Slewing in "Continuous Time" Circuits
• Slewing is not only an issue in switched capacitor circuits, it can also limit the large signal performance of continuous time circuits
• Example:
CL
iout
VoutVin
( ) ( )tVtv oo ωsinˆ=
( )dtdvCti o
Lout =
oLout VCi ⋅⋅= ω
• At large frequency and/or amplitude, the peak output current needed can exceed the maximum current available from the amplifier
( )tVC oL ωω cosˆ⋅⋅=
243
EE 214 Lecture 22B. Murmann 17
Resulting Waveform
EE 214 Lecture 22B. Murmann 18
Design Considerations
• When slewing is an issue, it can be mitigated by biasing the relevant transistors at lower gm/ID– Increase ID, keep gm constant
– Slewing performance improves, because of larger ID and also because the differential pair input range increases (2.8/[gm/ID])
– Small signal performance remains virtually unchanged or improves if fT is a limiting factor (since fT increases)
– Issue: Lower gm/ID means higher power consumption
244
EE 214 Lecture 22B. Murmann 19
How to Incorporate Slewing in Design Flow
• Set up a spreadsheet for small signal design as usual– Settling time requirement translates into minimum "linear"
bandwidth spec (based on linear analysis, without slewing)
• Introduce a bandwidth spec scale factor K≥1 in your design script– If the circuit slews, we will need more bandwidth than
predicted from linear analysis
• Perform design optimization as usual, begin with K=1
• Calculate slewing time, add to linear settling time– Done if tslew=0
• Increase K until design meets settling time spec– In the process, you may consider to optimize gm/ID values to
minimize power in presence of slewing
245
EE 214 Lecture 23B. Murmann 1
Lecture 23Feedback and Port Impedances
OTA Variants, CMFB Implementation
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 23B. Murmann 2
Overview
• Reading– 8.8.2 (Closed-Loop Impedance Formula using Return Ratio)– 6.7 (MOS Active-Cascode Amplifiers)– 12.5 (CMFB Circuits)– 12.6 (Fully Differential Amplifiers)– 9.4.4 (Compensation of Single Stage CMOS OTAs)
• Reference– K. Bult, G.J.G.M. Geelen, "A fast-settling CMOS op-amp for SC circuits with
90-dB DC gain," IEEE J. Solid-State Circuits, Dec. 1990, pp. 1379 – 1384.
• Introduction– In order to complete our framework for feedback circuit analysis, we will
study the effect of feedback on port impedances in today's lecture. A useful analytical tool for quick calculations is Blackman's Impedance Formula, which allows us to find port impedances based on a superposition of simple sub-analyses.
– Next, we survey common OTA architectures and associated implementation aspects. Most OTAs used in practice are derivatives of the basic single- or two-stage topology, with add-ons such as simple cascodes or gain boosted cascodes. The survey concludes with a brief analysis of the most commonly used common mode feedback implementations.
246
EE 214 Lecture 23B. Murmann 3
Using Feedback to Modify Port Impedances
• Our initial motivation for using feedback was to build precise gain elements that are insensitive to gmR variations
• In addition, feedback can be used to increase/decrease port impedances
– In fact, we have already seen one example of such behavior• Closed loop bandwidth of OTA feedback amplifier was (1+T)
higher than open loop bandwidth
• This means the impedance seen by the load capacitor must have dropped in presence of feedback
• We can calculate the port impedances of arbitrary feedback circuits using "Blackman's Impedance Formula"
– Based on loop gain calculations
– Extremely useful and easy to use
EE 214 Lecture 23B. Murmann 4
Blackman's Impedance Formula
1. First, find port impedance with feedback loop broken
– E.g. set gm=0
2. Calculate loop gain in circuit with port under consideration shorted
3. Calculate loop gain in circuit with port under consideration open
• In many cases, either the "shorted" or "open" loop gain is zero
( ) ( )( )open portT
shortedportTkZZ portport ++⋅==1
10
247
EE 214 Lecture 23B. Murmann 5
Example 1
( ) ( )m
voutout gaRkR 100 ====
( ) 0= shortedportT
( ) vaopen portT =
vmout ag
R+
=1
11
EE 214 Lecture 23B. Murmann 6
Example 2: Shunt-Shunt Stage
( ) oFmin rRgR +== 0
( ) 0= shortedinputT
( ) omrgopen inputT =
( ) ⎟⎟⎠
⎞⎜⎜⎝
⎛+≅
++=
o
F
momoFin r
Rgrg
rRR 111
1
( ) omout rgR == 0
( ) 0= shortedoutputT
( ) omrgopen outputT =
momoout grg
rR 11
1 ≅+
=
248
EE 214 Lecture 23B. Murmann 7
Example 3: Active Cascode
• Also referred to as "Regulated Cascode" or "Gain Boosting" technique
( ) 22010 omvout rgraR ⋅≅=
( ) vmbm
mv a
ggga shortedportT ≅+
≅22
2
( ) 0=open portT
( )vom
vomout
argrargrR
⋅⋅≅+⋅⋅≅
2201
2201 1
EE 214 Lecture 23B. Murmann 8
Basic Implementation
• Can use simple CS stage as auxiliary amplifier
– Issue: Costs headroom
• See literature for more advanced implementations
– E.g. using fully differential folded cascode amplifier as an auxiliary amplifier
332201 omomout rgrgrR ⋅⋅≅
249
EE 214 Lecture 23B. Murmann 9
Properties
• Must compensate local feedback loop such that its crossover frequency ωc occurs before non-dominant pole at source of M2.
• Consideration of the total impedance at the output node (including CL) shows that gain boosting introduces a pole zero doublet around ωc
– Can result in slow step response, if not designed carefully
– See Bult, JSSC 12/1990 for design considerations
• In typical designs, "gain boosting" adds about 20-30% to the total power dissipation of an OTA
EE 214 Lecture 23B. Murmann 10
OTA Variants: Telescopic OTA
• Approximately same gain as two stage amplifier, with only two current legs
– Maximum power efficiency
• Issue: Low swing
– Especially if input and output common mode cannot be chosen freely
250
EE 214 Lecture 23B. Murmann 11
Biasing
• Typically use at least 20% of tail current in auxiliary biasing branch
– Must avoid slow recovery of this node during transients
• Can use several device in series to implement MB1
– "1/3rd or 1/5th" device
– Helps avoid extremely small width for MB1
EE 214 Lecture 23B. Murmann 12
Folded Cascode OTA
• Large input common mode rage
• Slightly improved output range
• Folding adds power dissipation
251
EE 214 Lecture 23B. Murmann 13
Compensation
• Nondominant pole ωp2=gm1a/Cp
– Cp=Cgs1a+Cj1a+Cj1
• Easy to compensate
– Make k·β·gm1/CL < ωp2
– k=3 for 72° phase margin
AC Half Circuit for Telescopic OTA AC Half Circuit for Folded Cascode OTA
EE 214 Lecture 23B. Murmann 14
Current Mirror OTA
• Gm=k·gm1,2, ωp2 ~ ωT/(1+k)
• Large swing
• Good for low speed, low power applications– See e.g. Yao, JSSC 11/2004
k 1 1 k
252
EE 214 Lecture 23B. Murmann 15
All NMOS Signal Path (1)
• Capacitive level shift allows NMOS in second stage
[Feldman et al., JSSC 10/1998]
EE 214 Lecture 23B. Murmann 16
All NMOS Signal Path (2)
• Differential pair and separate common mode feedback in second stage
[Yang et al., JSSC 12/2001]
253
EE 214 Lecture 23B. Murmann 17
Gain Boosted Gain Boosters
• Gain ~ gmro6, design achieved av0=130dB in 0.18µm technology
[Chiu et al., ISSC 2004]
EE 214 Lecture 23B. Murmann 18
Common Mode Feedback
• Implementation aspects
– How to sense
– How to compare to desired value
– How to provide a "knob" for adjusting Voc
254
EE 214 Lecture 23B. Murmann 19
Knob
• Typically generate ~80% of tail current with fixed bias, leave remaining 20% as tuning range for CMFB loop
EE 214 Lecture 23B. Murmann 20
Comparison Circuit
• Low frequency loop gain T0 ≅ 0.25·gmx·rop1 · gmp2/gmx
– Loop will control Voc more accurately if Mp1 is cascoded
255
EE 214 Lecture 23B. Murmann 21
Sensing
• Using a resistive divider may destroy differential gain
• Solutions
– Use source followers to drive divider (headroom issue)
– Purely capacitive sensing
EE 214 Lecture 23B. Murmann 22
CMFB Implementation Example
[Feldman et al., JSSC 10/1998]
• Circuit uses switched capacitors (CM) to set the voltage across sensing capacitors (CCM)
256
EE 214 Lecture 23B. Murmann 23
"Passive" CMFB (1)
• During φ1: Initialize voltage across Ccmfb to Voc,desired - VB
• During φ2: Activate feedback loop
– If Voc>Voc,desired, Vcntrl becomes >VB and lowers Voc
EE 214 Lecture 23B. Murmann 24
"Passive" CMFB (2)
• OTA cannot be used during φ1, because the common mode feedback mechanism is inactive
– Often not a problem in switched capacitor circuits, where the OTA is active only during one half-cycle
• Can use switched capacitor scheme shown on slide 22 to enable uninterrupted common mode feedback
• Unfortunately, this simple circuit cannot be used if an additional inversion is needed in the common mode feedback loop
– E.g. won't work for a two-stage OTA that uses a single common mode feedback loop (see e.g. slide 9, lecture 18)
– Will work for the two-stage OTA with separate CMFB loops as shown on slide 16
257
EE 214 Lecture 23B. Murmann 25
Common Mode Half Circuit
• Low frequency loop gain:
2CC
Cr
2gT
xcmfb
cmfbop
mx0
+⋅≅
• Loop crossover frequency
xcmfb
xcmfbL
mxc
C5.0CC5.0C
C
g21
+⋅
+≅ω
• Nondominant pole
y
mn2p C
g≅ω
EE 214 Lecture 23B. Murmann 26
Design Considerations
• The required bandwidth of the common mode loop strongly depends on the amount of expected imbalance, common mode transients or ac components
– In an ideal world, the common mode is not affected by the signal and hence stays constant
• In this case, the bandwidth of the CMFB loop is unimportant
• For robustness in practical implementations, the bandwidth of the common mode loop is often chosen to be about 30% of the differential signal path bandwidth
– In a typical switched capacitor circuit with 10 time constants differential settling, this means that the common mode has about 3 time constants to settle
• Enough time to remove 95% of common mode disturbance
258
EE 214 Lecture 24B. Murmann 1
Lecture 24OTAs with Single Ended Outputs
Output Stage Examples
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 24B. Murmann 2
Overview
• Reading
– 4.3.5 (Differential Pair with Current Mirror Load)
– 6.3.3 (Systematic Offset in Two-Stage Amplifier)
– Chapter 5 (Output Stages)
• Introduction
– This lecture concludes our discussion of amplifier implementations by looking at a number of missing bits and pieces. First, we will discuss subtleties of OTA implementations that use a single ended output. While rarely used in the signal path of high performanceintegrated circuits, OTAs with single ended outputs can be useful as auxiliary amplifiers in biasing circuits. Finally, we will take brief look at output stages that are suitable for driving resistive loads. While most on-chip loads tend to be capacitive in nature, low impedance drivers are often needed when interfacing to off-chip components and wire lines.
259
EE 214 Lecture 24B. Murmann 3
Single Ended OTA
• Current mirror performs "differential to single-ended conversion"
EE 214 Lecture 24B. Murmann 4
Mirror Doublet
• Half of the output current comes directly from the differential pair, the other half goes through a current mirror with finite bandwidth
– Result: Pole-zero doublet
ps1
p2s1
vg
ps1
121
21vgi
idm
idmo
−
−=
⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
−+=
• In the circuit on the previous slide p ≅ -ωTp/2 (neglecting junctions)
260
EE 214 Lecture 24B. Murmann 5
Unity Gain Buffer
• Closed loop gain ≅ 1
• Output impedance ≅ 1/gm
• Output range equal to common mode input range
– Advantageous to use a folded cascode OTA architecture
EE 214 Lecture 24B. Murmann 6
Inverting Amplifier
• Cx = ?, T(s) = ?
– Big mess…
261
EE 214 Lecture 24B. Murmann 7
Two-Stage OTA with Single Ended Output
EE 214 Lecture 24B. Murmann 8
Systematic Offset
• No offset if VGS6=VGS3
balances the current in the output branch
• Input referred systematic offset is (VGS6,balance-VGS3)/av1, where av1 is the gain of the first stage
262
EE 214 Lecture 24B. Murmann 9
Output Stages
• Needed to drive resistive loads (low R)
– Integrated continuous time RC-filters
– Off-chip resistive loads
– Line drivers• E.g. twisted pair (Ethernet, ISDN, ADSL)
• Solutions
– Use OTA + source follower output stage• Swing issue
– Use OTA + "low gain" common source stage• E.g. make gm·RL~1
• E.g. 20mS·50Ω =1, IBIAS=20mS·10V-1 = 2mA
– "Sophisticated" output stages• Examples on following slides
EE 214 Lecture 24B. Murmann 10
Output Stage Nomenclature (1)
• Class-A
– Output devices conduct for entire cycle of output sine wave
– E.g. source follower with constant current source bias
• Class-B
– Output devices conduct for ≅50% of sine wave cycle
– E.g. back-to back PMOS/NMOS source followers• NMOS and PMOS each conduct for about one half cycle
• No quiescent current during zero crossing of sine wave
• Class-AB
– Output devices conduct for >50%, but <100% of cycle
– E.g. a simple inverter• Problem: How to set quiescent current around zero crossing
• Needs local or global feedback to mitigate nonlinearity
263
EE 214 Lecture 24B. Murmann 11
Output Stage Nomenclature (2)
Class-A Class-B
Class-AB Class-C
EE 214 Lecture 24B. Murmann 12
Selected References on Output Stages
• D. M. Monticelli, "A quad CMOS single-supply op amp with rail-to-rail output swing," IEEE J. Solid-State Ckts., pp. 1026-1034, Dec. 1986.
• R. Hogervorst, J. P. Tero, R. G. H. Eschauzier, and J. H. Huijsing, "A compact power efficient 3 V CMOS rail-to-rail input/output operational amplifier for VLSI cell libraries," IEEE J. Solid-State Ckts., pp. 1505 - 1513, Dec. 1994.
• G. Palmisano, G. Palumbo, and R. Salerno, "CMOS Output Stages for Low-Voltage Power Supplies," IEEE Trans. Ckts. and Syst. II, pp. 96-104, Feb. 2000.
264
EE 214 Lecture 24B. Murmann 13
Class-AB Output Stage
[Hogervorst]
EE 214 Lecture 24B. Murmann 14
Output Current (Vo=0)
• Output transistors never turn off
• Quiescent current set by transistor ratios
• Large drive capability
ID(M25)
ID(M26)Iout
Iin1=Iin2 [A]
Cur
rent
[A
]
265
EE 214 Lecture 24B. Murmann 15
Complete OpAmp
[Hogervorst]
EE 214 Lecture 24B. Murmann 16
Low Voltage Variant
[Palmisano]
<Vtn+|Vtp|
266
EE 214 Lecture 24B. Murmann 17
Complete OpAmp
[Palmisano]
<Vtn+|Vtp|
EE 214 Lecture 24B. Murmann 18
A Note on General Purpose OpAmps
• Issue: Compensation
– The designer of a general purpose OpAmp does not know anything about the feedback network of the particular application
– General purpose OpAmps are typically compensated for the "worst case", i.e. unity feedback configuration
• Tends to be wasteful, since much less compensation is needed for smaller return factors
– Some general purpose OpAmps provide an external pin to let the user decide on the required compensation capacitor
267
EE 214 Lecture 25B. Murmann 1
Lecture 25Supply Insensitive Biasing
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 25B. Murmann 2
Overview
• Reading
– 4.4.2 (Supply Insensitive Biasing)
• Introduction
– Throughout this course, we have ignored the question on how to generate the reference bias currents used in most of our circuits. Today, we will examine a variety of current reference implementations, with primary focus on supply independence.
268
EE 214 Lecture 25B. Murmann 3
Poor Man's Bias
• Issue: Current is essentially proportional to VDD
– E.g. if VDD, varies by X%, bias current will roughly vary by the same amount
RVVVII OVtDD
INOUT−−=≅
EE 214 Lecture 25B. Murmann 4
"Vt" Referenced Bias
• By using a sufficiently large device, we can make VOV << Vt, and achieve
2
ox
INt
2
OVt
2
1GSOUT R
LWC
I2V
RVV
RVI
µ+
≅+≅=
2
tOUT R
VI ≅
• Question: By how much will IOUT
change given some variation in VDD?
269
EE 214 Lecture 25B. Murmann 5
Sensitivity
• The sensitivity of a parameter y to a change in parameter x can be approximately found using
yx
!S
xy
yx
x/xy/y
x/xy/y =
∂∂=
∂∂≅
∆∆
• In our case, we are looking for
%1.7V1.0V6.0
V1.021.g.e
VVV
21
II
II1
SSS
1OVt
1OV
IN
OUT
OUT
IN
II
IV
IV
OUT
IN
IN
DD
OUT
DD
=++
≅
∂∂⋅⋅≅
⋅=
• Not bad, but also not all that great…
EE 214 Lecture 25B. Murmann 6
BJT Version
⎟⎟⎠
⎞⎜⎜⎝
⎛==
S
IN
22
1BEOUT I
Ilnq
kTR1
RVI
%7.3mV700mV26.g.e
Vq
kT
SBE
IV
OUT
DD==
270
EE 214 Lecture 25B. Murmann 7
Stability
( )
2p1p
22m
22m11m s1
1s1
1Rg1
RgRgsT
ωω+
⋅+
⋅+
⋅≅
• Loop gain greater 1 at low frequencies, two poles
– Means that we must make one of the poles dominant to guarantee sufficient phase margin
• E.g. use large capacitance to ground at drain of T1
EE 214 Lecture 25B. Murmann 8
Self Biasing
• In the above discussed bias generator circuits, the supply sensitivity is still fairly high, because IIN is essentially directly proportional to VDD
• Idea: Mirror output current back to input instead of using supply dependent input current!
271
EE 214 Lecture 25B. Murmann 9
Start-Up Circuit
• Unfortunately, self-biasing comes with a built in "chicken and egg problem"
– There exists a stable operating point with all currents =0
– Can use a simple start-up circuit to solve this problem
(WEAK)
EE 214 Lecture 25B. Murmann 10
VBE Reference
• Utilizes "parasitic" substrate PNP transistor available in any CMOS technology
RV
I 1BEOUT =
272
EE 214 Lecture 25B. Murmann 11
Temperature Dependence (1)
• Similar to the sensitivity to supply variations, we can establish an expression for temperature variations
– Fractional temperature coefficient
OUT
OUT
F IT
I
TC ∂∂
=
• For the circuit on the previous slide we have
RVI 1BE
OUT =⎟⎟⎠
⎞⎜⎜⎝
⎛∂∂−
∂∂=
∂∂−
∂∂=
∂∂
TR
R1
TV
V1I
TR
RV
TV
R1
TI
1BE
1BEOUT
21BE1BEOUT
EE 214 Lecture 25B. Murmann 12
Temperature Dependence (2)
• Final result
• Temperature dependence is usually quite high
TR
R1
TV
V1TC 1BE
1BEF ∂
∂−∂∂=
K/%33.0mV600
K/mV2T
VV
1 1BE
1BE
−=−≅∂∂
resistor)(poly K/%2.0TR
R1 +≅∂∂
• E.g. ∆T=100K ⇒ ∆I=-53% (!)
K/%53.0TCF −=
273
EE 214 Lecture 25B. Murmann 13
∆VBE Reference
( )nlnq
kTII
IIln
qkT
IIln
qkT
IIln
qkT
VVRI
IN
2S
1S
IN
2S
OUT
1S
IN
2BE1BEOUT
=
⎟⎟⎠
⎞⎜⎜⎝
⎛=
⎟⎟⎠
⎞⎜⎜⎝
⎛−⎟⎟⎠
⎞⎜⎜⎝
⎛=
−=⋅
( )nlnq
kTR1IOUT =
EE 214 Lecture 25B. Murmann 14
TC of ∆VBE Reference
( )
( ) ⎟⎟⎠
⎞⎜⎜⎝
⎛∂∂−
∂∂=
∂∂−
∂∂
=∂∂
TR
R1
TV
V1
RVnln
RTRV
TVR
nlnT
I
T
T2
T
2
TT
OUT
TR
R1
T1
TR
R1
TV
V1TC T
TF ∂
∂−=∂∂−
∂∂=
qkTVT =
• ExampleK/%13.0K/%2.0
K3001TCF =−=
• TC of resistor and kT/q partially cancel!
274
EE 214 Lecture 25B. Murmann 15
∆VGS Reference (1)
• Strange result, why is this useful?
2
1OV
REF
1OV
2OV1OV
2GS1GS2REF
Rm11V
I
m11V
VVVVRI
⎟⎠⎞
⎜⎝⎛ −
≅
⎟⎠
⎞⎜⎝
⎛ −≅
−=−=⋅
[Lee, 2nd ed. p.326]
m1
EE 214 Lecture 25B. Murmann 16
∆VGS Reference (2)
• The transconductance of M1 is approximately
21OV
1REF
1OV
1D1m R
m112
VI2
VI2g
⎟⎠
⎞⎜⎝
⎛ −⋅===
• Transconductance of M1 and other devices biased using this circuit depend only on m and R2!
– This is why this bias circuit is more appropriately called "constant gm reference"
• Design aspects
– Can use off-chip resistor to set gm precisely
– Large VOV helps reduce mismatch errors
– Small I⋅R2 makes circuit less sensitive to body effect
275
EE 214 Lecture 25B. Murmann 17
Constant Settling Time Bias (1)
• Reference
– I. E. Opris, L. D. Lewicki, "Bias optimization for switched capacitor amplifiers,"IEEE TCAS II, pp. 985-989, Dec. 1997.
Settling time with constant bias current
EE 214 Lecture 25B. Murmann 18
Constant Settling Time Bias (2)
• Settling time of amplifier is given by
m
slew,olinslews g
CNI
CVttt
⋅+
⋅≅+=
β
• Constant gm bias helps, but is not optimum for minimizing process variations
• Really want a bias circuit that keeps ts constant
276
EE 214 Lecture 25B. Murmann 19
Constant Settling Time Bias (3)
• Can make setting time "constant" by choosing
RVII 0
∆+=
⎟⎟⎠
⎞⎜⎜⎝
⎛⋅
+≅⇒
⋅+
⋅≅
mslew,o
s
m
slew,os
gINV
tCI
gCN
ICV
t
β
β
CNk
11t2R
tCV
Is
s
slew,o0 ⋅
⎟⎠
⎞⎜⎝
⎛ −⋅⋅⋅≅
⋅≅
β
EE 214 Lecture 25B. Murmann 20
Constant Settling Time Bias (4)
277
EE 214 Lecture 26B. Murmann 1
Lecture 26Bandgap Reference
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 26B. Murmann 2
Overview
• Reading
– 4.4.3 (Temperature Insensitive Biasing)
• Introduction
– In this lecture we will introduce the basic idea behind the frequently used "bandgap" voltage reference. Conceptually, a bandgap reference simply combines two quantities with opposite temperature behavior to generate a voltage with (approximately) zero TC.
278
EE 214 Lecture 26B. Murmann 3
Key Idea
• kT/q has a positive temperature coefficient
– "PTAT" proportional to absolute temperature
• VBE of a BJT decreases with temperature
– "CTAT" complementary to absolute temperature
• Can combine PTAT + CTAT to yield an approximately zero TC voltage reference
– Useful in circuits that require a stable reference voltage• E.g. A/D converters
EE 214 Lecture 26B. Murmann 4
Conceptual Block Diagram
279
EE 214 Lecture 26B. Murmann 5
A Closer Look at VBE
• Even though kT/q increases with temperature, VBE decreases because IS itself strongly depends on temperature
⎟⎟⎠
⎞⎜⎜⎝
⎛=
S
CBE I
Ilnq
kTV
⎟⎟⎠
⎞⎜⎜⎝
⎛−≅
⎟⎟⎠
⎞⎜⎜⎝
⎛≅
C
00G
)q/kT/(V
0
CBE
IIln
qkTV
eIIln
qkTV 0G
• I0 is a device parameter, which is (unfortunately) not completely independent of temperature– We'll ignore this for now
• VG0 is the bandgap voltage of silicon "extrapolated to 0°K"
EE 214 Lecture 26B. Murmann 6
Extrapolated Bandgap
1.205eV
V205.1q
eV205.1V 0G ==
[Pierret, Advanced Semiconductor
Fundamentals, p.85]
280
EE 214 Lecture 26B. Murmann 7
Temperature Coefficient of VBE
• Assuming that both I0 and IC are constant over temperature
TVV
IIln
qk
dtdV 0GBE
C
0BE −=⎟⎟⎠
⎞⎜⎜⎝
⎛−≅
• Example
⎟⎟⎠
⎞⎜⎜⎝
⎛−≅
C
00GBE I
Ilnq
kTVV
KmV02.2
K300V205.1V6.0
dtdVBE −=−≅
EE 214 Lecture 26B. Murmann 8
CTAT + PTAT
• Returning our initial idea, we can now find the condition that gives us a temperature independent voltage
⎟⎟⎠
⎞⎜⎜⎝
⎛⎟⎟⎠
⎞⎜⎜⎝
⎛−+≅
+⎟⎟⎠
⎞⎜⎜⎝
⎛−≅+
C
00G
C
00GBE
IIlnM
qkTV
qkTM
IIln
qkTV
qkTMV
• Combining VBE and an appropriately scaled version of kT/q produces a temperature independent voltage, equal to VG0
281
EE 214 Lecture 26B. Murmann 9
Simple CMOS Realization
( )
( )nlnq
kTRRVV
nlnq
kTR1
RV
II
1
21BEout
11
BE21
+=
===∆
EE 214 Lecture 26B. Murmann 10
Choice of n
• Usually make n=integer2-1, e.g. n=8
• Layout:
282
EE 214 Lecture 26B. Murmann 11
Design Example
• From measurement data, we know that |VBE| of a unit device is 700mV at room temperature and I=50µA
• Decided to use n=8
( ) ( )
( ) ( ) ΩΩ
Ωµ
k34.98lnmV26
V7.0V205.1k08.1nln
qkT
VVRR
k08.18lnmV26A50
1nlnq
kTI1R
1BEGO12
21
=−=−
=
===
EE 214 Lecture 26B. Murmann 12
Nonidealities
• "Curvature"
– Temperature dependence of I0
• Offset voltages and TC of offset voltages
• Resistor mismatch and TC
• Finite β and β mismatch
• More next lecture…
283
EE 214 Lecture 27B. Murmann 1
Lecture 27Bandgap Reference
(Continued)
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 27B. Murmann 2
Overview
• Reading
– 4.4.3 (Temperature Insensitive Biasing)
• Introduction
– Today's lecture will cover several important details that we've left out in our previous analysis of bandgap references. We will discuss nonidealities such as "curvature" and the impact of offset voltages. Finally, we will take a brieflook at state-of-the art implementations and performance.
284
EE 214 Lecture 27B. Murmann 3
Selected References (1)
• R. J. Widlar, "New developments in IC voltage regulators," IEEE J. Solid-State Circuits, pp. 2-7, Feb. 1971.– First report, LM309 5V regulator
• A. P. Brokaw, "A simple three-terminal IC bandgap reference,"IEEE J. Solid-State Circuits, pp. 388-393, Dec. 1974.– A classic implementation
• C. Palmer and R. Dobkin, "A curvature corrected micropower voltage reference," IEEE Int. Solid-State Conference, pp. 58-59, Feb. 1981.
• G. Nicollini et al., "A CMOS bandgap reference for differential signal processing," IEEE J. Solid-State Circuits, pp. 41-50, Jan. 1991.– Offset compensated amplifier
EE 214 Lecture 27B. Murmann 4
Selected References (2)
• T.L. Brooks et al., "A low-power differential CMOS bandgap reference," IEEE Int. Solid-State Conf., pp. 248-249, Feb. 1994.– Differential output, stacked diodes
• H. Banda et al. "A CMOS bandgap reference circuit with sub-1-V operation," IEEE J. Solid-State Circuits, pp. 670 - 674, May 1999 .
• P. Malcovati et al., "Curvature-compensated BiCMOS bandgap with 1-V supply voltage," IEEE J. Solid-State Circuits, pp. 1076-1081, July 2001.
285
EE 214 Lecture 27B. Murmann 5
VBE Revisited
• Last lecture, we assumed that
⎟⎟⎠
⎞⎜⎜⎝
⎛−≅
C
00GBE I
Ilnq
kTVV
• A more accurate, but empirical model is given by
⎟⎟⎠
⎞⎜⎜⎝
⎛ ⋅−≅C
r1
0GBE ITKln
qkTVV
• The temperature dependence inside the logarithm slightly curves the VBE vs. temperature characteristic– TC of VBE is not quite independent of temperature
• Parameter r depends on technology, typically 3…6
EE 214 Lecture 27B. Murmann 6
Curvature
[Lee, 2nd ed., p.322]
286
EE 214 Lecture 27B. Murmann 7
Collector Current
• Another superficial assumption was to assume that the current IC is independent of temperature
– Actually PTAT in the example circuit from last lecture
– Also affected by TC of resistors
• To capture the temperature behavior of IC, we can introduce yet another empirical fudge factor and write
⎟⎟⎠
⎞⎜⎜⎝
⎛−≅ n
r
0GBE TTKln
qkTVV
• The factor n is 1 for ideal PTAT current behavior
EE 214 Lecture 27B. Murmann 8
Modified TC of VBE
• With this refinement, we have
( )
( )
( )T
qkTnrVV
nrTTKln
qk
TTK
TnrKq
kTTTKln
qk
TTKln
qkT
dTd
dTdV
BE0G
n
r
n
r
1nr
n
r
n
rBE
−+−−=
⎥⎦
⎤⎢⎣
⎡−−⎟⎟
⎠
⎞⎜⎜⎝
⎛−≅
−−⎟⎟⎠
⎞⎜⎜⎝
⎛−≅
⎥⎦
⎤⎢⎣
⎡⎟⎟⎠
⎞⎜⎜⎝
⎛−≅
−−
287
EE 214 Lecture 27B. Murmann 9
Modified Condition for Zero-TC
• Or, more interestingly
( )
( )
qkT
qkTnrVV
M
Tq
kTnrVV
qkTM
dTdV
qkTM0
BE0G
BE0GBE
⎥⎦
⎤⎢⎣
⎡ −+−=⇒
−+−−=+=
( )q
kTnrVV 0Gout −+=
• Must design for Vout that is a few kT/q higher than VGO
EE 214 Lecture 27B. Murmann 10
Modified Output Voltage
• Getting good temperature stability typically requires some tweaking or calibration
– Some semiconductor foundries provide tried and true bandgap cells that are optimized for a particular technology
[Lee, 2nd ed., p.323]
288
EE 214 Lecture 27B. Murmann 11
Curvature Compensation
• Control current such that the impact of "(r-n) term" is minimized
• In practice, curvature compensation may not be all that effective if other nonidealities dominate…
[Palmer]
EE 214 Lecture 27B. Murmann 12
Offset Voltage
( )
( )
( ) ⎟⎟⎠
⎞⎜⎜⎝
⎛−−⎟⎟
⎠
⎞⎜⎜⎝
⎛+=
⎟⎟⎠
⎞⎜⎜⎝
⎛−++=
⎟⎟⎠
⎞⎜⎜⎝
⎛−===
1RRVnln
qkT
RRV
Vnlnq
kTRRVVV
Vnlnq
kTR1
RV
II
1
2OS
1
21BE
OS1
2OS1BEout
OS11
BE21
∆
289
EE 214 Lecture 27B. Murmann 13
Issues
• VOS appears amplified at the bandgap output and can cause a large absolute error in Vout
– Since R2/R1-1 ≅ 8, this means that for VOS=5mV, the error in Vout will be about 40mV or roughly 3.3%
• If the bandgap is trimmed after manufacturing, e.g. to yield an output of VGO+2kT/q, then this is no longer the point of zero TC in presence of VOS
• In CMOS, VOS drift is typically 1…10µV/K– This means Vout will drift at least 8µV/K, which corresponds
to about 6.6 ppm/K• Good CMOS bandgaps achieve about 10…50ppm/K
• Possible solutions– Mitigate impact of offset by stacking two VBE
– Cancel offset or use low offset BJT differential pair
EE 214 Lecture 27B. Murmann 14
SC Bandgap with Offset Cancellation
[Nicollini]
290
EE 214 Lecture 27B. Murmann 15
Bandgap with stacked VBE
[Brooks]
EE 214 Lecture 27B. Murmann 16
Sub-1-V Bandgap
• Idea: Add currents proportional to VBE and kT/q, instead of stacking voltages
≅1.2V
( )
⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
+=3R
Nlnq
kT
2RVRV BE
4ref[Banba]
291
EE 214 Lecture 27B. Murmann 17
Variant with Curvature Compensation
[Malcovati]
Low offset amplifier
EE 214 Lecture 27B. Murmann 18
Performance
292
EE 214 Lecture 28B. Murmann 1
Lecture 28Technology Scaling
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 28B. Murmann 2
Overview
• Introduction
– The trend of continuously shrinking feature sizes in integrated circuits has resulted in enormous performance gains in both digital and analog circuits. However, since device scaling necessitates the use of smaller supply voltages, it is often argued that noise limited circuits can no longer benefit from scaling. In this lecture, we will take a closer look at this argument and also review basic analog device performance trends in light of feature size reduction.
293
EE 214 Lecture 28B. Murmann 3
The Age of "Moore's Law"
• In 1965, Gordon Moore predicted that there will be an exponential growth in the number of transistors per integrated circuit
EE 214 Lecture 28B. Murmann 4
… And He was Right
[Moore, ISSCC 2003]
294
EE 214 Lecture 28B. Murmann 5
Smaller
Min
imum
Fea
ture
Siz
e
EE 214 Lecture 28B. Murmann 6
Faster
295
EE 214 Lecture 28B. Murmann 7
Cheaper
Ave
rage
Tra
nsi
stor
Pric
e
EE 214 Lecture 28B. Murmann 8
Misnomer
• The term "Moore's Law" was coined by the press
– Of course, the exponential progress rate is not set by fundamental law
– Merely a rate of progress that makes sense for the industry and keeps things predictable for all involved players
• Device technologists
• Makers of manufacturing equipment
• Circuit designers
• Sales and marketing
• The dirty truth is that Moore's law is mostly just a gigantic economic feedback loop
– With lot's of great innovation fueled by $$$
296
EE 214 Lecture 28B. Murmann 9
Moore's Law in Action
$$$
EE 214 Lecture 28B. Murmann 10
Worldwide Semiconductor Sales
[European Nanotechnology
Roadmap]
297
EE 214 Lecture 28B. Murmann 11
State-of-the-Art Semiconductor Fab
• Cost ~ $3,000,000,000
EE 214 Lecture 28B. Murmann 12
State-of-the-Art Silicon Technology
• 90nm feature sizes, electrical channel length of 50nm
• Gate oxide thickness 1.2nm, roughly 5 atomic layers
• 8 Layers of metal routing
298
EE 214 Lecture 28B. Murmann 13
State-of-the-Art Chips
Pentium 4 Processor, 125 Million Transistors[Schutz, ISSCC 2004]
Single-Chip 802.11 Transceiver
[Zargari, ISSCC 2004]
EE 214 Lecture 28B. Murmann 14
Impact of Technology Scaling
• Scaling is great from a digital perspective!
• Can show that scaling down features and voltages achieves three things simultaneously
– Higher speed
– More transistors/area
– Lower energy per operation
• How about analog circuits?
299
EE 214 Lecture 28B. Murmann 15
Quotes
• [Vertregt, ESSCIRC 2004]
– "Significant power efficiency improvements are predicted as a result of scaling to deep sub-micron technology nodes."
• [Annema, IEEE J. Solid-State Circuits, 12/2005 ]
– "In summary: unlike digital designs, analog circuits can benefit from technology scaling if the supply voltages are not scaled down."
• [Nauta, ESSCIRC 2005]
– "The evolution of CMOS technology will continue for many years to come, which is beneficial for digital circuits but which is not so for analog."
EE 214 Lecture 28B. Murmann 16
List of Concerns
• Reduced supply voltage
• Low intrinsic gain
• Variability
• Distortion
• Gate leakage
• Isolation
• …
• Cost (mask & wafer)
• Model accuracy
• …
300
EE 214 Lecture 28B. Murmann 17
gm/ID and fT trends
• gm/ID essentially unaffected by scaling
• Very high fT in recent technologies
– Enables RF CMOS
-0.4 -0.2 0 0.2 0.4 0.60
200
400
600
800
1000
1200(b)
VGS
-Vt [V]
(gm
/I D)*
f T [G
Hz*
S/A
]
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.60
10
20
30
40(a)
gm
/I D [S
/A]
VGS
-Vt [V]
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.60
40
80
120
160
f T [G
Hz]
180nm130nm90nm
180nm130nm90nm
EE 214 Lecture 28B. Murmann 18
Available Signal Swing
0
1
2
3
4
5
0.50um 0.35um 0.25um 0.18um 0.12um 90nm 65nm 45nm 32nm
Technology Node
VD
D [
V]
> 4kT/q
> 4kT/q
VDD
qkT8V SwingAvailable DD −<
301
EE 214 Lecture 28B. Murmann 19
Noise Limited Circuit Performance
C/kTSwingDR
CgBWIVP
2m
DDD ∝∝⋅∝
D
m2
DDDD I
gVSwingV
PDRBW ⋅⎟⎟
⎠
⎞⎜⎜⎝
⎛⋅∝⋅
• Low VDD is generally bad news, but– Analog designers have worked hard to maintain or even
improve Swing/VDD
• Typical ADC in 0.5µm: Swing/VDD=2/5• Typical ADC in 90nm: Swing/VDD=0.5/1
– How about gm/ID?
EE 214 Lecture 28B. Murmann 20
Leveraging fT
• Example– fT = 50GHz, 130nm: gm/ID = 8S/A, 90nm: gm/ID = 16S/A
• For "fixed-speed" applications, high fT can be leveraged to mitigate low VDD penalty
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.60
10
20
30
40
gm
/I D [S
/A]
VGS
-Vt [V]
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.60
40
80
120
160
f T [G
Hz]
180nm130nm90nm
302
EE 214 Lecture 28B. Murmann 21
Further Considerations
• Analog building blocks are never completely limited by thermal noise
– Not uncommon to have ~50% dynamic power
• Decreases with scaling
• Designers are continuing to develop/refine low-voltage design techniques
– Recent publications show very good analog building block performance at 1V
• Bottom line
– Analog design is challenging at 1V, but it's neither impossible nor detrimental
EE 214 Lecture 28B. Murmann 22
Intrinsic Gain
• A real issue
– How to design a high-gain op-amp with devices that have intrinsic gain of ~10?
• How much worse does this get at 45nm/65nm?
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.60
50
100
150
200
250(a)
VDS
[V]
I D [ µ
A]
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.60
10
20
30
40
50(b)
VDS
[V]
gm
/gd
s
180nm130nm90nm
180nm130nm90nm
(VGS-Vt=100mV)
303
EE 214 Lecture 28B. Murmann 23
Intrinsic Gain in the Near Future
• Pretty bad…
• Solutions– Use non-minimum length device (NML-device)– Use asymmetric device without drain-side pocket implant (A-
device)– Or, don't try to build op-amps in these technologies…
• E.g. "Digitally Assisted ADC" research in my group
0 0.2 0.4 0.6 0.8 10
5
10
15
VDS
[V]
gm
/gd
s
45nm (TCAD)
65nm (TCAD)
90nm (BSIM4)(VGS-Vt=100mV)
EE 214 Lecture 28B. Murmann 24
A-Device vs. Standard Device
HoleImpactIonizationSigned Log
0
2.9
5.8
8.7
11.6
14.5
17.4
20.3
23.2
26.52
-0.16 -0.12 -0.08 -0.04 0 0.04 0.08 0.12 0.16
-0.1
6-0
.12
-0.0
8-0
.04
00.
040.
080.
120.
160.
20.
240.
280.
320.
360.
40.
44
HoleImpactIonizationSigned Log
0
2.9
5.8
8.7
11.6
14.5
17.4
20.3
23.2
26.58
-0.16 -0.12 -0.08 -0.04 0 0.04 0.08 0.12 0.16
-0.1
6-0
.12
-0.0
8-0
.04
00.
040.
080.
120.
160.
20.
240.
280.
320.
360.
40.
44
• Removing halo widens depletion region– Reduces impact ionization and improves output resistance
304
EE 214 Lecture 28B. Murmann 25
Intrinsic Gain of Alternate Devices (45nm)
• For both NML and A-device Lphysical=80nm (Lphysical=24nm for minimum length device)
• Great, lots of gain!
– But how about fT?
(VGS-Vt=100mV)
0 0.2 0.4 0.6 0.8 10
50
100
150( )
VDS
[V]
gm
/gd
s
Min. length
NML-device
A-device
(VGS-Vt=100mV)
EE 214 Lecture 28B. Murmann 26
gm/ID and fT for Alternate Devices (45nm)
• fT much lower than for minimum length 45-nm device
– But still better than minimum length device in 90nm…
• Who needs fT > 200GHz in an op-amp…?
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.50
1000
2000
3000
4000
5000(b)
VGS
-Vt [V]
(gm
/I D)*
f T [G
Hz*
S/A
]
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.50
10
20
30
40(a)
VGS
-Vt [V]
gm
/I D [S
/A]
-0.2 -0.1 0 0.1 0.2 0.3 0.4 0.50
200
400
600
800
f T [G
Hz]
Minimum length
NML-device
A-device
Minimum length
NML-device
A-device
305
EE 214 Lecture 28B. Murmann 27
Variability (1)
[Courtesy A. Bowling, Texas Instruments]
EE 214 Lecture 28B. Murmann 28
Variability (2)
• Must keep in mind that even if we draw nice colorful rectangles in our CAD tools, things won't come out like that…
• Device shape strongly depends on surroundings
• Good old layout recipes become more important and must be applied even when only "moderate" matching in required
– Use unit devices
– Match surroundings of unit devices
306
EE 214 Lecture 28B. Murmann 29
Variability (3)
• Device mismatch larger than process corner variations!– For small "digital" transistors…
[Marcel Pelgrom, Philips]
EE 214 Lecture 28B. Murmann 30
Variability (4)
Analog
• A well known problem
– Designers are used to "caring" about mismatch
• Lots of options and potential solutions
– Layout techniques, analog or digital calibration, dynamic element matching, larger device area, …
• Usually care about matching for a few up to a few hundred transistors
Digital
• A "new" problem
– Significant impact on achievable performance, yield, design methodology, EDA, …
• Big difference compared to analog
– Care about millions if not billions of devices!
307
EE 214 Lecture 28B. Murmann 31
An Interesting Hike Lies Ahead…
Bag of Tricks
A
D
EE 214 Lecture 28B. Murmann 32
Cost – The "Real" End of The Roadmap?
• Reference point: 30 mm2 die in 0.12µm CMOS
[Marcel Pelgrom, Philips][Marcel Pelgrom, Philips]
308
EE 214 Lecture 29B. Murmann 1
Lecture 29Class Summary
Project Discussion
Boris MurmannStanford University
Copyright © 2006 by Boris Murmann
EE 214 Lecture 29B. Murmann 2
Transistor Models
• Any model is an approximation of the real world
– Must leave many details out
– Must retain the important details (to be useful)
– Appropriate level depends on questions you want to answer• BSIM model/Spice
• gm/ID approach
• Long channel equations
• When designing and analyzing circuits, we are usually forced to use much simpler models than the ones available in Spice
– gm/ID methodology partially closes this gap
• A "good" IC designer is always on the lookout for modeling limitations!
309
EE 214 Lecture 29B. Murmann 3
Transistor Figures of Merit
• Transit Frequency
gs
mT C
g=ω
• Transconductor Efficiency
D
mIg
• Intrinsic Gain
omrg
EE 214 Lecture 29B. Murmann 4
Circuit Analysis
• Once we have appropriate models, we can in principle analyze any circuit using KCL/KVL
– Usually too difficult and also won’t allow us to reason about design choices and tradeoffs
• Crutches for circuit analysis
– Small signal approximation
– Zero value time constant method
– Miller approximation
– Return ratio analysis
– Blackman's impedance formula
– …
• Again, a good designer will always be on the lookout for potential limitations of the respective analysis method
310
EE 214 Lecture 29B. Murmann 5
The "Atoms" of Analog Circuit Design
• Common source
– Basic voltage amplifier
• Common gate
– Good for "shielding"
– Can help boost output impedance, mitigate Miller effect
• Common drain
– Buffer, level shifter
• Differential pair
EE 214 Lecture 29B. Murmann 6
Feedback (1)
• Desensitizes circuit to (forward-) gain variations
• Modifies port impedances
• Central quantity of interest: Loop gain T
– Quantifies static error
– Used to assess stability, phase and/or gain margin
– Helpful in calculating closed loop port impedances
• Finding T is simple
– E.g. break loop at transconductor, inject test current, calculate ratio of return and test current
• Typically need a dominant pole
– Hard to use more than 2-3 amplifier stages and maintain sufficient phase margin
311
EE 214 Lecture 29B. Murmann 7
Feedback (2)
• Impact of phase margin on step response – Excessive "ringing" for phase margin <60°– Fast settling for phase margin ~70°
• Impact of phase margin on closed loop AC response– Gain peaking for small phase margin; leads to larger 3-dB
bandwidth as a side effect
• Compensation techniques– Can use simple load compensation for single stage OTAs– Miller compensation is most popular for two-stage designs
• Can push parasitic zero to infinity using nulling resistor• Beware of tricks such as pushing the zero into the LHP to avoid
pole-zero doublets
– Cascode compensation• For designers with experience…
EE 214 Lecture 29B. Murmann 8
OTAs with Capacitive Feedback
• Primary application: Switched capacitor circuits
• Important performance metric: Settling time
– Small signal model fails when input exceeds "linear range"
– Slew rate ~ I/C
– ts = tlin + tslew
• High performance OTAs are usually implemented as fully differential circuits
– Need CMFB
– Better XXX-rejection than single ended circuits
– Easy to analyze, since circuit is "perfectly" balanced• Exceptions occur during transients, in presence of mismatch,
etc.
312
EE 214 Lecture 29B. Murmann 9
Electronic Noise
• Tends to set achievable power dissipation in circuits with high DR requirements
• Fundamental noise
– Thermal noise
• Technology related noise
– 1/f noise
• Increasing (noise limited) precision by "one bit" quadruples power dissipation!
EE 214 Lecture 29B. Murmann 10
References
• Self-biasing concept
– Beware of stability issues!
• Current references
– VDD, VGS, ∆VGS, VBE, ∆VBE - based approaches
– To first order, ∆VGS bias yields constant gm over temperature and process
• Voltage reference
– Bandgap
– Useful as a reference whenever there's a need to convert "bits to Volts" or "Volts to bits"
– Conceptually simple, but lots of second order issues
313