Lecture L

ELEN E6316:
Analog-Digital Interfaces
Course Introduction
Dynamic Range (Noise and Linearity)

Columbia University
Spring 2020

Tod Dickson, Ph.D.

[email protected]
Research Staff Member, IBM T.J. Watson Research Center
Adjunct Professor, Columbia University

T. Dickson © 2020
 Course Information
 Instructor: Tod Dickson
 Email: [email protected]
 Office hours: Tues 5:30-6:30PM (1332A/1334A Mudd)
 TA: Aditya Gaonkar
 Email: [email protected]
 Office hours: TBD
 Class Tues. 7:00-9:30 PM (Mudd 834)
 Course material available online:
http://courseworks.columbia.edu
 Prerequisite: ELEN E4312 (Analog Electronic
Circuits)

T. Dickson © 2020
References
No required textbook. My lecture notes draw material
from a number of sources:

✓ M. Pelgrom, Analog-to-Digital Conversion (3rd ed),


Springer, 2017.
 F. Maloberti, Data Converters, Springer, 2008.

:
 R. Schreier & G. Temes, Understanding Delta-Sigma
Data Converters, McGraw-Hill, 2003.
 Chan Carusone, Johns & Martin, Analog Integrated
Circuit Design (2nd edition), Wiley, 2012.

I
 R. van de Plassche, Integrated Analog-to-Digital
and Digital-to-Analog Converters (2nd edition),
Springer, 2003.

Text in blue are available as an online Ebook ch 9

through Columbia Library.
.

T. Dickson © 2020
Assignments & Grading
 Grading:
 Midterm: 20% (March 12th)
 Problem Sets: 20%
 Project: 30%
 Final: 30% (TBD, tentatively May 14th)
 Computer Requirements:
 The Cadence design environment will be
used. Transistor model files will be
provided by the instructor.
 MATLAB may also be required for some
assignments.
T. Dickson © 2020
Dynamic Range
 Dynamic range is the range of signal
levels that can be (linearly) processed
by a system.
 Minimum signal level is set by the
noise floor of the system.
 Maximum signal level is set by the
nonlinearities within the system.

T. Dickson © 2020
Quantifying Power
Remember that a dB is a relative quantity.
Example: Signal 1 is X dB larger than signal 2 means
P
X = 10 log 1  (X has units of dB)
 2
P
How do we express the absolute power of a signal, when we
don’t have another signal to compare it to?

We choose an arbitrary reference: 1mW.

 P 
Y = 10 log 1  (Y has units of dBm)
 1mW 

Units of dBm mean “dB relative to 1mW”

T. Dickson © 2020
Noise

T. Dickson © 2020
Noise
 Two types of noise – interference and inherent noise
 Inference – corruption of one signal by another
 Examples – power supply noise, capacitive coupling
between wires (clock signals are very bad for this).
 May or may not be random.
 In ICs, board, and package design, can only improve
through careful layout/wiring/routing/planning.
 These notes do not deal with interference noise.
 Inherent Noise
 Noise from devices (transistors, resistors, etc).
 Can be reduced (but never eliminated) through
circuit design techniques, but not strongly dependent
on layout or wiring.
 We will only discuss inherent noise.

T. Dickson © 2020
 ( wt )
* For ult ) -_ aisin * For ult)= Aotaisinlwt )

terms =

Aq URms=ao2-A÷M
Noise in Time Domain
•Noise signals are random,
hence we analyze them
statistically.

•Noise has zero mean value

(i.e., no DC component).

'
•We cannot determine the
noise voltage at a given instant
-

in time.

t
1/ 2
•Root mean square (rms) value found by averaging the 1 T 2 
Vn ,rms =   vn (t )dt 
noise over a period of time T. Longer T gives more T 0 
accurate estimate of rms value.

&
1/ 2
1 T 2 
•Similarly, we can find the rms value of a noise current. I n,rms =   in (t )dt 
Vn2,rms T 0 
Pdiss = = Vn2,rms
=
•Power dissipated in a 1-Ω resistor is 1Ω
T. Dickson © 2020
-
 Adding noise sources
If we know the rms values of the two
noise sources, can we determine the
total rms noise?

vno (t ) = vn1 (t ) + vn 2 (t )

T T
1 2 1
=  vno (t )dt =  [vn1 (t ) + vn 2 (t )] dt = Vn21,rms + Vn22,rms + 2CVn1,rmsVn 2,rms
2
Vno2 ,rms
T 0 T 0

T
1
C is the correlation coefficient, and is
always between -1 and +1.

T 0
vn1 (t )vn 2 (t )dt
C≡
(C is the covariance divided by the product Vn1,rmsVn 2,rms
of the standard deviations)
T. Dickson © 2020
Correlation
Vno2 ,rms = Vn21,rms + Vn22,rms + 2CVn1,rmsVn 2,rms
Case 1) C = 0
Noise sources are uncorrelated (independent)

Vno2 , rms = Vn21,rms + Vn22,rms

Case 2) C = +1
Noise sources are fully correlated and in-phase

Vno2 ,rms = Vn21,rms + Vn22,rms + 2Vn1,rmsVn 2,rms = [Vn1,rms + Vn 2,rms ]

2

Case 3) C = -1
Noise sources are fully correlated and 180o out-of-phase
Vno2 ,rms = Vn21,rms + Vn22,rms − 2Vn1,rmsVn 2,rms = [Vn1,rms − Vn 2,rms ]
2

T. Dickson © 2020
 Noise in Frequency Domain
Time domain Frequency Domain
v(t) V2(f)

Periodic signals have

power at discrete
T0 frequencies.

time f0=1/T0 f

Random signals have

no power at discrete
frequencies! But
power is spread out
across the spectrum.

T. Dickson © 2020
 Noise in Frequency Domain

PSD RPS D

Noise in frequency domain characterized by the Power Spectral Density (PSD), or

by Root Power Spectral Density.

The noise power at an exact frequency is zero (otherwise the noise would have
some periodicity). Instead, PSD looks at the mean squared noise in a normalized
1-Hz bandwidth (look at the units of the PSD).

Can determine the rms noise voltage (or current) by integrating the PSD.
∞
Vn2,rms =  Vn2 ( f )df (Note: Only integrated over
0
one side of the spectrum)
T. Dickson © 2020
 Measuring Noise

If the noise power at an exact frequency is zero, then how do we measure?

The mean square value of noise at a given frequency is proportional to the

observation bandwidth ∆f centered about that frequency.

Spectrum analyzer adjusts the bandwidth of a bandpass filter – lower this

“resolution bandwidth” to lower the noise floor.

{
In plot above – Vn2(f) is about 1000 (µV)2/Hz at 0.1Hz. Hence for a 1-mHz

i. observation bandwidth, you would expect a measured mean squared value of

1 (µV)2 (i.e., 1000 times lower than for the normalized 1-Hz bandwidth).
T. Dickson © 2020
Adding noise sources in Frequency
Domain
Assuming noise sources are uncorrelated
(i.e. C = 0)

Vno2 ,rms = Vn21,rms + Vn22,rms

∞ ∞

Vn21,rms =  Vn21 ( f )df V 2

n 2 , rms =  Vn22 ( f )df
0 0
∞ ∞ ∞ ∞
[
Vno2 ,rms =  Vno2 ( f )df =  Vn21 ( f )df +  Vn22 ( f )df =  Vn21 ( f ) + Vn22 ( f ) df ]
0 0 0 0

∞ ∞

V
2
no ( f )df =  [Vn21 ( f ) + Vn22 ( f )]df
0 0

Differentiating both sides gives…

Total noise PSD is the
is
V ( f ) = V ( f )+V ( f )
2 2 2
no n1 n2 sum of the two PSDs
T. Dickson © 2020
‘White’ Noise
 Constant power spectral density with respect to
frequency (analogous to white light).
 Good model for many real noise sources.
 Noise power is infinite over all frequencies, but in
real circuits/systems with finite bandwidth the
noise power will be finite (more on this later)

T. Dickson © 2020
White Noise Example ( Noise Power)

What is the total noise power over the

-

range of frequencies from dc to 10kHz?

T. Dickson © 2020
1/f (or ‘Flicker’) noise
 Some practical noise sources have a power spectral
density that increases at lower frequencies (i.e., when
you give more time for a random event to occur).
 Usually, we exclude very low frequencies from our total
integrated noise calculation (for example, 10-8 Hz takes
3+ years to observe)

Vril f)
=µ¥
"

V
→
If *

The units for Kv should

be volts
T. Dickson © 2020
Flicker Noise Calculation
Find the 1/f noise corner for a noise
source with PSD:

Answer:
Equate the white noise and the 1/f noise
(3.2 × 10 ) = (1 × 10 )
−6 2
−6 2 f =
(3.2 × 10 )
−6 2
= 10.24 Hz
f (1 × 10 )
−6 2

T. Dickson © 2020
 Filtered Noise

 Noise can be referred to the input or the output of a system.

 Output referred noise PSD equals the input-referred noise
multiplied by the square of the frequency response of the
system.

T. Dickson © 2020
1st-order lowpass filtered noise

T. Dickson © 2020
Equivalent Noise Bandwidth

 Note that result is the same as if A(s) were a

brickwall filter with bandwidth
fx, the “equivalent noise bandwidth”
T. Dickson © 2020
 Signal-to-Noise Ratio
•Definition of SNR at a given node in a system is the ratio of the
signal power to the noise power at that node.

-
 signal power  Vx2,rms  Vx ,rms 
SNR ≡ 10 log   = 10 log  2  = 20 log   (in dB)
 noise power  Vn ,rms  Vn ,rms 

•For an amplifier or filter, a useful metric is the noise factor (F)

or noise figure (NF).

vin vout SNRin (unitless)

F≡ (≥ 1)
SNRout
Power gain G, NF ≡ 10 log(F ) (dB)
vin2 ,rms Input-referred noise vna ,rms
2

SNRin =
vn2,rms (F is a measure of the noise added by the amplifier)

Gvin2 ,rms v 2
in , rms v 2
+v
n , rms
2
v
na , rms
2
na , rms
SNRout = = F= = 1+
Gvn2,rms + Gvna
2
, rms v 2
n , rms+v 2
na , rms v 2
n , rms v 2
n , rms

T. Dickson © 2020
 Noise in Cascaded Circuits
•Suppose we have two cascaded gain stages with gains AV1 and AV2,
and the input referred noise of each stage is Vn1 and Vn2 respectively.

* AV1 * AV2
Vn12(f) Vn22(f)
•We wish to know the total noise referred back to the input of
the first stage.

( f ) = Vn21 ( f ) + n 2 (2 f )
2
AV1 AV2 V
* 2
Vneq
AV 1
Vneq2(f)
•The noise of the second stage is divided by the gain of the first stage!

To minimize overall noise, it is critical to minimize the

noise contribution of the first stage.

T. Dickson © 2020
Noise from Circuit Components

T. Dickson © 2020
Resistor Noise
 Thermal noise due to random Brownian motion of
carriers.
 Can express noise spectrum as a voltage source in
series with the resistor, or a current source in
parallel with the resistor.

or

 T is absolute temperature (in Kelvin), k is Boltzmann’s

constant
 This is a “white noise” source – its PSD is constant with
respect to frequency.
J Easier to remember, for 1kΩ at room temperature:
k = 1.38 × 10 − 23
K VR ( f ) = 4nV / Hz VR2 ( f ) = 16(nV ) / Hz
2

T. Dickson © 2020
Resistor Noise
 IMPORTANT NOTE!! Only physical resistors generate
thermal noise.
 Resistors from “small-signal equivalent circuits”, i.e. rπ, ro
in transistor small-signal models are not physical
resistors and hence do not generate noise (there are
other noise sources to worry about in transistors).
Remember that these resistors arise from taking the
derivative of a particular I-V characteristic.
 Base resistance (i.e., rb or sometimes called rx) for a
bipolar transistor, and gate resistance of a MOSFET (i.e.,
rg) are both physical resistors and do generate thermal
noise.
 Bottom line – you have to know which resistors in a
small-signal equivalent circuit are physical and which are
not.

T. Dickson © 2020
High-Frequency Bipolar Hybrid-π
Model

Base resistance –
physical resistor that
generates noise!

rπ – not a ro – not a
physical resistor and physical resistor
does not generate noise ∂VCE
∂VBE ro =
rπ = ∂I C
∂I B
(There are other noise generators besides resistor thermal noise in a bipolar
transistor, which we shall cover shortly)
T. Dickson © 2020
Capacitor Noise
 Ideal capacitors and inductors are noiseless.
 However, they do store noise generated by other
elements (such as resistors).

∞ ∞
2

4kTR
DE
Vno2 ( f ) = H ( jω ) VR2 ( f ) =

4kTR
1
1 + (2πfRC )
2
4kTR
∞
kT
Vno2 ,rms =  Vno2 ( f )df =  df = tan −1
(2πfRC ) =
0 0 1 + (2πfRC )2
2πRC 0 C
 Very important (and not an intuitive) result.
 As R increases, the increase in noise is balanced by
the decrease in bandwidth.
 For noise-sensitive circuits, don’t overdesign the
bandwidth.
T. Dickson © 2020
 Etf .IE?gIvoE::n:4scdemam
E÷¥÷::÷÷÷÷ :*
Replace W by 21T f
:*
f Hlf)
&
integrate over →
"

itzitfrc
 Capacitor Noise derivation
Csl ( Capacitor
- -

✓
-

no
→

!E%¥±I¥
noise ,

17¥ -

↳ Noise of a resistor
can be model as a lossless
resistor in series with a

Noise
Voltage source .

Von tf)
To find
division
we can use
voltage function
finding by the transfer .

Hls )=s # =

p
Hcjw ) -21
Rt 1- Rsc t s ( Rjwt ,
Sc

Itkiwsl !
1-
ftlciwy ,¥,
-

12-CRWCQVnofst-VBCSIIHCJWYZ-HKTRH.ES#
-
Diode Noise
 Forward biased diode with DC bias current ID
 While DC current ID gives the average rate at which
carriers pass through the p-n junction, in reality
since carriers are discrete quantities this rate may
deviate from the mean value at a given instant in
time. This is called shot noise.

Like thermal noise,

shot noise is “white” –
its PSD is constant or
with frequency.

T. Dickson © 2020
Bipolar Transistor Noise
 Four noise sources:
 Collector shot noise (due to DC current through a p-n junction)
 Base shot noise (again due to DC current through a p-n junction)
 Base flicker (1/f) noise (due to generation-recombination in the base)
 Thermal noise from the physical base resistance

T. Dickson © 2020
 Bipolar Transistor Noise Biff
Collector shot noise
Base resistance thermal noise (4kTrb) can be referred back
to the input…

as a noise voltage
(2qIC/gm2)

OR as a noise current
(2qIC/|β(f)|2)

BUT DO NOT
COUNT THE
Base shot noise current (2qIB) COLLECTOR SHOT
NOISE TWICE!!
Base flicker noise current (2qKIB/f),
where K is a technology constant
T. Dickson © 2020
MOSFET Noise
Lecture 2 01/28/20
 Two noise sources:
 Flicker (1/f) noise (due to random trapping of carriers in silicon-
silicon dioxide interface)
 Channel thermal noise (from channel resistance)

?
me

→ ( flicker noise )

(in saturation region)

(in triode region)

T. Dickson © 2020
Subthreshohdthegion
I. I,
"hTh¥I Fe
⇐ e n = , +

" H #I
C, e-
I Do I.
fussy → IT IS) 2g I +
-
=

=
2 ID
q
 MOSFET Noise: Low-to-Moderate
Frequency Model
In saturation region…
Can represent as a single noise
voltage source

Thermal d ↳ flicker
noise noise

 For low-to-moderate frequencies (< fT/10), we can refer the channel

thermal noise back to the input as a voltage simply by dividing by gm2.
 For high frequencies (like RF applications), noise modeling is more
complex (beyond the scope of this course).

T. Dickson © 2020
Op-Amp Noise
 Represented by three noise sources.
 Input noise voltage
 Two noise currents, one for each op-amp input. The noise
currents are usually neglected for MOSFET op-amps.
 Typically these three noise sources are uncorrelated.

Values for all three noise sources

depend on op-amp implementation.

T. Dickson © 2020
 Noise Analysis – Common Source
The three noise sources are uncorrelated,
so we can analyze by superposition.
-

For simplicity, assume RD << ro

Also, MOSFET input capacitance noted by Cin
(accounts for Cgs and Miller-multiplied Cgd)
Step 1: Identify noise sources
2
Vns2 ( f ) = 4kTRs Vnd2 ( f ) = 4kTRD 2
I nd ( f ) = 4kT gm
3
(ignore 1/f noise) Step 2: Compute output-referred noise
due to each noise source
Output noise due to Rs:
Vno2 1 ( f ) = 4kTRs ( g m RD )
2 1
1 + (ωRs Cin )
2 d.
Output noise due to RD: Vno2 2 ( f ) = 4kTRD
2
Vno2 3 ( f ) = 4kT g m (RD )
2
Output noise due to M1:
3
T. Dickson © 2020
Noise Analysis – Common Source
Step 3: Find total output noise
Vno2 ( f ) = Vno2 1 ( f ) + Vno2 2 ( f ) + Vno2 3 ( f )
 g m2 Rs 1 2 
V ( f ) = 4kTR 
2 2
+ + gm 
1 + (ωRs Cin ) RD 3 
no D 2

Step 4: Refer back to the input

Vno2 ( f ) Vno2 ( f )
V 2
ni (f )= 2
=
(g m RD )2
1 + ([
ω R C
s in )2
]
A( f )
  1 2 1  2 
V ( f ) = 4kT  Rs +  2
2
ni +  1 + (ωRs Cin )  [ ]
  g m RD 3 g m  
Note: input referred noise INCREASES with
frequency due to reduced amplifier gain at high
frequencies.
T. Dickson © 2020
2-Stage Op-Amp Noise

Assume Q1/Q2 are matched,

Q3/Q4 are matched

T. Dickson © 2020
2-Stage Op-Amp Noise
Analyze by superposition –
find the output noise due to
each noise source.
of Qi r2

gain
Opamp

Vno2 1 ( f ) = ( g m1 R o ) Vn21 ( f )
2
-

Vno2 2 ( f ) = ( g m 2 R o ) Vn22 ( f ) = ( g m1 R o ) Vn21 ( f )

2 2

Vno2 3 ( f ) = Vno2 4 ( f ) = (g m3 Ro ) Vn23 ( f )

2

2
 g 
(Output noise due to Q5 found in Vno2 5 ( f ) =  m5  Vn25 ( f )
manner similar to common-mode  2 g m3 
gain analysis of the stage.)

T. Dickson © 2020
 's Cfl
Vno found
By approx
was
.

T.tdgiiz.IE.vn#-vns-*gng.gLm-
¥5 Ins
V
#A -
-

9ms Uns

I
WILL Inz

19M¥ )
%
"
A
vii. sis ) = Un's Cs) *

Qs is a diode
connected transistor
 2-Stage Op-Amp Noise
Total output noise is found through summation -

Vno2 ( f ) = Vno2 1 ( f ) + Vno2 2 ( f ) + Vno2 3 ( f ) + Vno2 4 ( f ) + Vno2 5 ( f )

2
 g 
Vno2 ( f ) = 2(g m1Ro ) Vn21 ( f ) + 2(g m 3 Ro ) Vn23 ( f ) +  m 5  Vn25 ( f )
2 2

 2 gm3 
Refer back to the input by dividing by the gain.

Vno2 ( f )
V (f )=
2
ni
(g m1R o )2

2 2
g   g m5  2
Vni2 ( f ) = 2Vn21 ( f ) + 2 m 3  Vn23 ( f ) +   Vn 5 ( f )
 m1 
g  m1 m 3 o 
2 g g R

T. Dickson © 2020
 2-Stage Op-Amp Noise
2 2
g   g m5  2
Vni2 ( f ) = 2Vn21 ( f ) + 2 m3  Vn23 ( f ) +   Vn 5 ( f )
small & be
 g m1   2 g m1 g m 3 Ro  can

Notes: I neglected .

1. The last term (noise from Q5) is small compared with the first
two terms and will be neglected.
2. Noise from 2nd stage of a 2-stage op-amp will be divided by the
gain of the 1st stage when referring to the input, hence it will
also be small and will be neglected.
I

2
 g m3  2
V ( f ) ≈ 2V ( f ) + 2
2
ni
2
n1
 Vn3 ( f )
 g m1 

T. Dickson © 2020
2-Stage Op-Amp Noise
Thermal Noise
2
g 
V ( f ) ≈ 2V ( f ) + 2 m3  Vn23 ( f )
2
ni
2
n1
 g m1 
 2  1   2  1 
Vn21 ( f ) = 4kT   Vn23 ( f ) = 4kT   
 3  g m1   3  g m 3 
  2  1   g m 3    2  1  16  1  16  g m 3 
2 2
 1 
Vni2 ( f ) ≈ 2 4kT    + 2 g   4 kT    = 3 kT  g  + 3 kT  g   
  3  m1   m1 
g   3   m 3 
g  m1   m1   m3 
g

Observation: Make gm1 as large as

possible to keep the thermal noise low.

This will also help with flicker (1/f)

noise.

T. Dickson © 2020
2-Stage Op-Amp Noise
Thermal Noise
• Sometimes the input-referred noise is
equated to an equivalent noise
resistance, by finding the resistance
value that would result in the same
input referred noise.
Thermal noise and resistance

interchangeably
noise are

value
Resistor of this will give

eg
.
noise
produced by opamp .

To
16  1    g m 3   4  1    g m 3  
V ( f ) = kT 
2
 1 +   = 4kTRn Rn =    1 +  
 3  g m1    g m1 
ni
3  g m1    g m1 
T. Dickson © 2020
Linearity and Distortion

T. Dickson © 2020
 Common-Source w/ Active Load
Large-signal characteristic:
In amplifier design, we make
a ‘linear’ approximation valid
for small input amplitudes
superimposed around a
certain dc input level. High gain

region

However, the complete input/output transfer

characteristic is nonlinear.

T. Dickson © 2020
 Nonlinearities/Distortion
vO
“DC operating point”
vI vO vo
VO We super impose AC
vi signal on top DC
(assuming no operating point .

memory effects) vI
VI

Input-Output relationship can be described by a polynomial (Taylor series)…

where the DC component is
VO + vo = ao + a1vi + a2 vi2 + a3vi3 + ... ao = VO
↳ Taylor series
around specific
For small inputs, we approximate this as a linear system. point
In other words, we neglect the higher-order powers such that
-

operating
it
changes based
⑦
vo ≈ a1vi on the op .
point .

For larger input signals, the linear approximation is less accurate and
=
we must consider the impact of higher order terms.
T. Dickson © 2020
Harmonic Distortion
Ai Az, have
Consider the nonlinear system described by Az , . . .

different units .

vo (t ) = a1vi (t ) + a2vi2 (t ) + a3vi3 (t ) + a4 vi4 (t ) + ...

Let the input signal be vi (t ) = ViA cos(ω0t ) value of

shift in the
Harmonics terms are Even terms give
signal
M the

ViA2 ViA3
vi (t ) = ViA cos (ω0t ) =
2 2 2
[cos(2ω0t ) + 1] v (t ) = V cos (ω0t ) =
3
i
3
iA
3
[cos(3ω0t ) + 3 cos(ω0t )]
2 D-
4
ViA4
v (t ) = V cos (ω0t ) =
4
i
4
iA
4
[cos(4ω0t ) + 4 cos(2ω0t ) + 3] Dc term ( shift average
8 -

value @ output
Collecting terms with the same frequencies, we can now express
the output in terms of harmonics of the input frequency.
vo (t ) = b0 + b1 cos(ω0t ) + b2 cos(2ω0t ) + b3 cos(3ω0t ) + b4 cos(4ω0t ) + ...

T. Dickson © 2020
Harmonic Distortion
vo (t ) = b0 + b1 cos(ω0t ) + b2 cos(2ω0t ) + b3 cos(3ω0t ) + b4 cos(4ω0t ) + ...
The bi coefficients are functions of the nonlinear
coefficents (ai) and the input amplitude ViA

DC term b0 =
1 3
a2ViA2 + a4ViA4 + ... (shifts operating point)
2 8
3 (expands or compresses
ω0t term b1 = a1ViA + a3ViA3 + ...
4 the “linear” gain)
Harmonic terms
1 1 1 1
b2 = a2ViA2 + a4ViA4 + ... b3 = a3ViA3 + ... b4 = a4ViA4 + ...
2 2 4 8
For small ViA
a2 2 a3 3 a4 4
b1 ≈ a1ViA b2 ≈ ViA b3 ≈ ViA b4 ≈ ViA
2 4 8

T. Dickson © 2020
 Harmonic Distortion
Total Harmonic distortion
Harmonic distortion coefficients
→ unit less of a
system .

b2 1 a2 b3 1 a3 2
HD2 ≡ ≈ ViA HD3 ≡ ≈ ViA
b1 2 a1 b1 4 a1

Total harmonic distortion…

b22 + b32 + b42 + ...
THD = the
b1 ← Normalized
to

fundamental
output of the

What does the output spectrum look like? THD depends on VIA
Input Output

ω ω
ω0 ω0 2ω0 3ω0
T. Dickson © 2020
 Distortion in Differential Systems
Each box
Again consider our nonlinear system… is

described by same

vo (t ) = a1vi (t ) + a2vi2 (t ) + a3vi3 (t ) + ... Polynomial .

Consider a fully differential implementation…

tf flefentiaf amp
÷÷±÷:÷n::÷÷:÷÷÷:÷÷
v p (t ) = a1vi (t ) + a2 vi2 (t ) + a3vi3 (t ) + a4 vi4 (t ) + ...

If
.v (t) i
+ ,
+
- vdiff (t ) = v p (t ) − vn (t ) = 2a1vi (t ) + 2a3vi3 (t ) + 2a5vi5 (t ) + ...
-vi(t)
Fully diff
¢

,£[¥ft-
-

otherof diff
.
vn (t ) = −a1vi (t ) + a2 vi2 (t ) − a3vi3 (t ) + a4 vi4 (t ) + ...
half In a fully differential system, rinn
amp
.

vine
even-order distortion terms cancel.
Note: Remember, it was the even-order terms that were
responsible for the shift in the DC operating point… &
T. Dickson © 2020
Example: MOS Differential Amplifier
Large-signal characteristic:

‘Average’ or ‘dc’ level is set by the current source and the drain
resistor. This operating point is independent of the input level
(to a certain degree, assuming the FETs operate in saturation)

T. Dickson © 2020
Gain Compression
Plotting the output power at the fundamental frequency
as a function of the applied input power

Output Power
Transfer
The amount of gain compression
character .
that can be tolerated depends on
1dB the application.
OP1dB
Non -

Linearity
Gain The “1dB compression point”
compression
is widely used, and represents a
10% reduction in the gain.

a IP1dB
Input Power

Linear gain .

T. Dickson © 2020
dB is OP -

IP
in
gain not oP/£p
 Estimating 1dB Compression
For a nonlinear system dominated by 3rd order harmonics
-

vo (t ) ≈ a1vi (t ) + a3vi3 (t )
Consider a single-tone input signal vi (t ) = ViA cos(ω0t )
From previous discussion, we know that the output
including impact of nonlinearities will be
 3  (focusing only on the
vo (t ) = b1 cos(ω0t ) + ... =  a1ViA + a3ViA3  cos(ω0t ) + ...
 4  fundamental tone)

We can estimate the 1dB compression point by comparing

the output power to that predicted by the small-signal gain
 3 
20 log a1ViA,1dB + a3ViA3,1dB  = 20 log(a1ViA,1dB ) − 1dB

-
as coefficient
 4 

ViA,1dB = 0.145 1
a
↳ FIFTY
"

"feFewgE%gLaViH
Produces compression .

✓ a3

+ as expansion
↳
T. Dickson © 2020

,
•
Linearity of Band-limited Systems

Lowpass ed
Filter :#
after
*
HD2 H
D3
ω ω 2ω 3ω

 Note that when measured with input tones near the circuit’s passband
edge, distortion terms are filtered out
 With a general (broadband) input pattern, such high-frequency
distortions may mix with lower-frequency tones in-band
 Single-tone measurements may not accurately capture the linearity
performance of a circuit at all frequencies
 Multi-tone inputs will show non-linearities…
T. Dickson © 2020
Two-tone testing
W, & we are pretty close
to each other .
Like it .

diff .

Lowpass
Filter

HD3
ω1 ω2

 Two-tone measurements address this problem by measuring 3rd-order

distortion terms looking at the intermodulation of two closely-spaced
⑧

tones
 Two-tone testing is especially important for bandpass systems where
single-tone harmonics always lie out-of-band
 Let’s look at the theory behind this distortion…
T. Dickson © 2020
 Intermodulation Distortion
For the same nonlinear system described by
vo (t ) = a1vi (t ) + a2 vi2 (t ) + a3vi3 (t ) + ...
Consider a 2-tone input signal vi (t ) = V1 A cos(ω1t ) + V2 A cos(ω2t )
The output is

vo (t ) = a1 [V1 A cos(ω1t ) + V2 A cos(ω2t )] + a2 [V1 A cos(ω1t ) + V2 A cos(ω2t )] + ...

2

T
a3 [V1 A cos(ω1t ) + V2 A cos(ω 2t )]
3

Expand the 2nd term…

[ ]
a2 [V1 A cos(ω1t ) + V2 A cos(ω2t )] = a2 V12A cos 2 (ω1t ) + V 22A cos 2 (ω2t ) + 2V1 AV2 A cos(ω1t ) cos(ω2t )
2

Already know these terms Will yield sum and difference

yield harmonics at 2ω1 and 2ω2 components
1 1
2nd order intermod products cos(ω1 + ω2 )t + cos(ω1 − ω2 )t
2 2
T. Dickson © 2020
 Intermodulation Distortion
Define “fundamental” and 2nd-order intermod distortion levels as

I D1 ≈ a1V1 A I D 2 ≈ a2V1 AV2 A

2nd order intermodulation value is ratio of these two terms
do
I inter:L?
aV V a 2 order
-

IM 2 ≈ 2 1 A 2 A = 2 V2 A = 2HD2 -

a1V1 A a1
Similarly, you can show that expanding the 3rd order terms IDs amplitude of

will yield terms with frequencies at ω1, 3ω1, ω2, 3ω2, (2ω2-ω1), 3rd order inter mod

(2ω2+ω1), (2ω1-ω2), and (2ω1+ω2) distortion product

3rd-order intermod distortion level can be shown to be

3 (depending on which term
I D 3 ≈ a3V1 AV22A or I D 3 ≈ 3 a3V2 AV12A
4 4 we’re considering) 3 3
Usually for testing, we just set 1 A
V = V2A = ViA in which case I D 3 ≈ a 3ViA
4

0
I 3 a
3rd order intermodulation ratio is IM 3 ≡ D 3 ≈ 3
ViA2
I D1 4 a1
T. Dickson © 2020 3rd -
order
Inter modulation
 Third-Order Intercept Point
IP3 is the point
where the power I D1 ≈ a1ViA
level of the
fundamental
equals the power
3
I D 3 ≈ a3ViA3
4
3g
level of the third- Assuming
linear growth
order intermod
products.

It cannot be
directly measured,
as the amplifier
reaches
compression before
that point – thus
you have to
extrapolate.

T. Dickson © 2020
 Estimating IP3 3rd order non - lineaiities in
systems
are
normally the dominant non linearity
-

Simply set ID1 = ID3

3
I D1 = I D 3 ≈ a1ViA, IP 3 = a3ViA3, IP 3
4
4 a1
ViA, IP 3 ≈
3 a3
Recall that 1dB
compression occurs at
a1
ViA,1dB ≈ 0.145
a3
ViA, IP 3 4
ViA,1dB
≈
0.435
≈ 9.6dB IIP3 is 9.6dB higher than IP1dB
T. Dickson © 2020
 How to measure IP3
“Two-tone” test: apply vi (t ) = ViA cos(2πf1t ) + ViA cos(2πf 2t )
where f1 and f2 are close together (~ 1% difference).
→

Example: f1=1GHz f2 = 1.01GHz

Hence 2f1-f2 = 0.99 GHz, 2f2 – f1 = 1.02 GHz

Measure output power levels on spectrum analyzer for different ViA

Pout (dBm) 3rd order tones will
increase much more
rapidly than fundamental
tones as ViA is increased
(3dB for every 1dB).
f (GHz)
0.99 1.0 1.01 1.02

Extrapolate to find IP3 (see previous slides). Applied levels should

be well below compression levels.
T. Dickson © 2020
 Third-Order Intercept Point
We can estimate OIP3 (and IIP3) if we know a single ID1 point and its
corresponding intermod ratio IM3. Recall that OIP3 is defined as ID1 for which IM3
= 0dB. Also note that for every 1dB change in ID1, ID3 changes by 3 dB and IM3
changes by 2dB.
IM 3
Leg Loge OIP3 = I D1 −
2
scales
ID1 (OIP3, ID1 in dBm
IM3 in dB)

I D1 ≈ a1ViA
3
IM3 I D 3 ≈ a3ViA3
4

⑦
I D 3 3 a3 2
IM 3 ≡ ≈ ViA
I D1 4 a1

as:*.no#E.e:n...m.am-sire..t*..+n T. Dickson © 2020

IF not ,
data is been taken close to the compression point
.
 Non
linocuts
-

come from
I D2
&
ID ,

Remember :

Sm
=2I÷= IT
↳ Taylor Series

↳ expansion

This amp dont 2nd order linearity but 3rd order

 so
Be 00

I
More
linearity
@
expenses
of
loosing gain .
Dynamic Range
Signal swing
Max signal amplitude limited
by the tolerable nonlinearity

SNR improves

Less distortion

Dynamic
Range
Min signal amplitude limited by
the tolerable SNR

T. Dickson © 2020
Quantifying Dynamic Range
The ratio is
 Signal-to-noise ratio (SNR) → to
Vat
 Ratio of the output signal power to the
output noise power (usually for a
sinusoidal input).
 Signal-to-noise-and-distortion ratio
(SNDR or SINAD)
 Ratio of the output signal power to the
output noise and distortion power (again,
usually measured for a sinuoidal input).

T. Dickson © 2020
 Signal to noise and distortion ratio (SNDR)
 Defined as the ratio of the signal power to the total
power in all noise and distortion components

The maximum
achievable SNDR is a
popular metric for
Fig. 9.26 quantifying the
dynamic range of an
analog circuit
•
Vf
'

is the •
Viz this
,
. .
.
.

Power of the are the power

T. Dickson © 2020
fundamental of the output
of V' out signal
signal harmonics
 is unwanted tone like harmonics
spurious any

Spurious-Free Dynamic Range

The spurious-free dynamic range
(SFDR) is defined as the signal-
to-noise ratio when the 3rd order
intermod distortion power is
equal to the noise power.

=
SFDR = I D* 1 − N O = I D* 1 − I D* 3
This assumes that 3rd order terms f
of
dominated, i.e. in fully-
differential circuits. You may slope of

have to consider 2nd order terms slop}

depending on the circuit.

This SFDR definition is used by

amplifier or filter designers. In
ADCs, a slightly different (and
less stringent) definition is used.

T. Dickson © 2020
 Rate Hz
Sampling zoom

•
!

ins
-

Df
.
 in
SFDR and OIP3
Bs
Yomafree (all quantities in dB or dBm)
age II
*
.
wi fig
SFDR = I * D1 − N O = I D* 1 − I D* 3
other
I 1-
③

IM 3 = I D 3 − I D1 / DB

OIP3 = I D1 −
IM 3 (I − I )
= I D1 − D 3 D1
2 2
3 1 Holds before
OIP3 = I D1 − I D 3 of
pointthe
2 2 compression .

At I D1 ,
*
I D* 3 = N o and

I D* 1 = SFDR + N o
therefore
2
3 1 3 1
OIP3 = I D* 1 − I D* 3 = (SFDR + N o ) − N o or SFDR = (OIP3 − N o )
2 2 2 2 3
T. Dickson © 2020
 Spurious-Free Dynamic Range

An alternative definition of
SFDR is the ratio of the signal
power to the power of the
largest spurious spectral Largest
distortion
component (usually within a ←
that we
defined bandwidth). see .

This definition is commonly

used for data converters.

(From M. Zhang, IEEE

JSSC Dec 2019)

T. Dickson © 2020
© 2020 T. Dickson
For student use in ELEN E6316
Unauthorized distribution is prohibited

T. Dickson © 2020
•
Do BJT noise problem