T1 Fundementals of Freq Referencee

ISSCC 2023 Tutorial
Fundamentals of Frequency References
Danielle Griffith
Texas Instruments
Live Q&A Session: Feb. 19th, 2023, 10:00am-3:00pm, PST

D. Griffith T1: Fundamentals of Frequency References 1 of 84
© 2023 IEEE International Solid-State Circuits Conference

About Danielle Griffith
 B.S.E.E. and M.Eng degrees from the Massachusetts
Institute of Technology, Cambridge
 2003 - Current, Texas Instruments in Dallas, Texas
 Fellow in the Low Power Connectivity business unit
 Focusing on efficient wireless systems
 Contributions to IEEE Solid-State Circuits Society
 >50 papers, book chapter, multiple workshops
 Technical Program Committees: RFIC (‘14 – ‘15),
ISSCC (‘15 – ‘19 & from ‘23), VLSI (‘19 – ‘22)
 IEEE SSCS Distinguished Lecturer (‘21 – ‘22)
 Associate editor of the IEEE Journal of Solid-State
Circuits (‘20 – current)

Outline
 Motivation
 Key metrics
 Frequency reference landscape
 Fully integrated silicon frequency references
 Discrete/multi-chip frequency references
 Ceramic
 MEMS Increasing accuracy
 Crystal
 TCXO, OCXO
 Chip scale atomic clock
 Summary & conclusions
 Papers to see this year
 Key references

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

Motivation
 Why do we need frequency references?
 Nearly every electronic system needs a clock
 But accuracy requirements vary
 Examples:
 MCU clock
 Accuracy: +/-1% - 1 second in 1.7 minutes
 USB interface for data transfer
 Accuracy: +/-500ppm – 1 second in 33 minutes Increasing
 Wireless node for machine health monitoring in a factory accuracy
 Accuracy: +/-20ppm – 1 second in 14 hours
 NASA’s deep space atomic clock
 Accuracy: 3·10-9 ppm - 1 second in 10 million years

Brief History of Timekeeping
 Calendar: At least 5000 years ago
 Sundial: ~1500 BC
 Accuracy ~1 hour
 Weight driven mechanical clock: 1283 Balance spring &
wheel, Huygens
 Accuracy ~30minutes/day
 Balance spring/wheel: 1657 (Huygens,
Hooke)
 Accuracy ~10s/day
 Quartz crystal oscillator: 1921 (Cady)
 Accuracy ~200s/year
 Atomic clock: 1955 (Essen)
 Accuracy ~30ns/year

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

Key Frequency Reference Metrics (1)
 Frequency accuracy
 How close the frequency reference is to ideal value
 Short term & long term frequency stability freq Cycle-to-cycle jitter
 How much frequency varies around the ideal value

 Quantified by time
 Phase noise / jitter Two sample deviation
freq
 Shift in instantaneous frequency
 Two sample deviation / Allan deviation τ
 Shift in average frequency over observation window τ time
 Frequency drift / aging freq Freq drift
 Shift over hours/days/years
time

Key Frequency Reference Metrics (2)
 Target frequency
 Most are kilohertz to gigahertz
 Frequency tuning range
Frequency
 Ability to adjust frequency to meet requirements
 Ability to compensate frequency shift from
environmental changes T=-40ºC
 Power consumption T=27ºC

 Tradeoff with stability Target
freq Tuning code
 Startup / stabilization time
 Can be less than 1µs to more than 1 hour
depending on type T=85ºC
 Environmental robustness
 Stable frequency during environmental changes

Environmental Impact on Frequency
 Time
 Temperature
 Load impedance
 Supply voltage / bias currents
 Acceleration / vibration / shock
 Stress / strain
 Humidity
 Ambient pressure
 Gas permeation
 Radiation
 And so on...

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

Frequency Reference Landscape
Integrated
in Silicon Ceramic
Crystals, MEMS
TCXO, MEMS
OCXO
Rubidium
Cesium

Frequency Reference Applications
Integrated
in Silicon
MCU Ceramic
USB
Crystals, MEMS
TCXO, MEMS
OCXO
Rubidium
Cesium

Integrated
in Silicon
MCU Ceramic
USB
Utility Crystals, MEMS
meters
TCXO, MEMS
OCXO
Rubidium
Cesium

Integrated
in Silicon
MCU Ceramic
USB
meters Bluetooth Wi-Fi,
TCXO, MEMS
cellular
OCXO
Rubidium
Cesium

Integrated
in Silicon
MCU Ceramic
USB
TCXO, MEMS
cellular
Test
Base
equipment OCXO
station
Rubidium
Cesium

Integrated
in Silicon
MCU Ceramic
USB
TCXO, MEMS
cellular
Test
Base
equipment OCXO
station
Aviation, radar
satellites,
Rubidium
space, military,
seismology
Cesium

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

Ring Oscillator as a Frequency Reference?
 Very area and power efficient
 0.05μW/MHz/gate for 180nm CMOS
with 1.8V supply
 Further power reduction at smaller
geometries

Ring Oscillator as a Frequency Reference?
 Very area and power efficient
 0.05μW/MHz/gate for 180nm CMOS
with 1.8V supply
 Further power reduction at smaller
geometries
 Large frequency variation
 Example frequency variation:
 2x variation over industrial
temperature range
 3x variation with process
 5x variation with supply
 Poor reference - how to make a
more stable frequency reference?

Open Loop Relaxation Oscillator
 Relaxation oscillator I I
 Current charges capacitor Vin +

 Comparator resets capacitor when Vref −
voltage Vin reaches Vref = I·R
C R
 Frequency of Vclk = 1/RC
 To first order, frequency is Vclk
dependent only on R & C. Vref
 R and C are on-chip resistors and Vin
capacitors & can vary ~±20% over
process corners Vclk
 One-time production trimming
required
time

Accuracy Limitations
 R & C can vary over temperature I I
 Typically 50ppm/ºC - 3000ppm/ºC Vin +

 0.5% to 30% frequency variation over Vref −
industrial temperature range
C R
 Frequency can be impacted by
 Temperature coefficient of R & C Vclk
 Supply voltage variation Vref

 Comparator offset & delay variation Vin
 Current mismatch
Vclk
time

Chopped Relaxation Oscillator
Paidimarri,
 Improved stability by adding ISSCC ‘13 I I
chopping at comparator input V1 + φ
 Impact of input noise and current & V2 −
comparator mismatch reduced φ φ φ φ
C φ R C φ φ

Chopped Relaxation Oscillator
Paidimarri,
 Improved stability by adding ISSCC ‘13 I I
chopping at comparator input V1 + φ
 Impact of input noise and current & V2 −
comparator mismatch reduced φ φ φ φ
 Frequency F given by
1 C φ R C φ φ
𝐹𝐹 =
2𝑅𝑅𝑅𝑅 + 2𝑡𝑡𝑑𝑑
 where td is comparator delay
 Comparator delay variation impacts
frequency stability
 Sets lower limit to power consumption

Remove Impact of Comparator Delay
+ Griffith, ISSCC’14 Voltage V1
I ≈3/2∙Vbias
−
Vbias
Vbias Inverter trip point
VSW≈1/2∙Vbias
V1 Vclk Ground
R C ≈−1/2∙Vbias
Time
Delay through
inverters
 Bias voltage Vbias tracks inverter  Still sensitive to R & C temperature

transistors’ threshold voltages coefficients (TC)
 Inverter delay becomes constant  Requires low-TC or composite resistor
over temperature & supply voltage
→ more stable frequency

Closed Loop Ring Oscillator
 Ring oscillator:
 Power efficient, unstable
 Relaxation oscillator:
 Less power efficient, more stable

Voltage controlled
Vref ring oscillator
 Ring oscillator: + CLK
 Power efficient, unstable −
 Relaxation oscillator: 1
𝐹𝐹𝐶𝐶𝐶𝐶𝐶𝐶 ∝
 Less power efficient, more stable F→V
÷N 𝑅𝑅𝐹𝐹 𝐶𝐶𝐹𝐹
converter
 Stabilize ring oscillator with feedback –
Frequency locked loop (FLL) Abidi, JSSC 1987
 Power efficient, stable

Voltage controlled
Vref ring oscillator
 Ring oscillator: + CLK
 Power efficient, unstable −
 Relaxation oscillator: 1
𝐹𝐹𝐶𝐶𝐶𝐶𝐶𝐶 ∝
 Less power efficient, more stable F→V
÷N 𝑅𝑅𝐹𝐹 𝐶𝐶𝐹𝐹
converter
 Stabilize ring oscillator with feedback –
Frequency locked loop (FLL) Abidi, JSSC 1987
VDD
 Power efficient, stable
VDD
 Implementation:
 Feedback sets switched phi1
phi2
impedance of Cf & Rf to Vref +
 Frequency proportional to 1/RFCF CF RF
 Amplifier has low BW, low power - CLK CLK phi1
 Frequency still sensitive to R & C Gen
phi2
temperature coefficients Vre f

Pulse Density Modulated Resistors
VDD
 Resistors R1 and R2 have VDD
different temperature
phi1
coefficients phi2
+ A. Khashaba, JSSC 2022
 Pulse density modulate CF
resistors to compensate - CLK CLK phi1
resistor temperature coefficient Gen
phi2
 SW closed: Equivalent Vre f
R1 R2
resistance → R
 SW open: Equivalent SW1 SW2
resistance → ∞
1/Req
 Switch resistors based on a
lookup table from 2-point trim 1/R2
1/R1
 Create equivalent resistance
with ~zero TC → stable freq.
temperature

Frequency Drift in Integrated Oscillators
 Oscillator frequency is proportional to
1/RC
 P-type poly resistor is commonly used
due to high density
Measured frequency, MHz

 Weak Si-H bonds at grain boundaries
can be broken with thermal stress,
creating traps that increase resistance.
 Drift is logarithmic over time, strongly
dependent on temperature
 Up to 1-2% frequency change over
lifetime
 Dominant error source for state-of-the-
art integrated oscillators
 Previously rarely discussed in literature
 See paper 3.5 at this year’s ISSCC Aging time, hours

Long Term Stability: Allan Deviation
D. Allan, 1966
 Allan deviation is a statistical
measurement of frequency
instability
 Differences between successive
readings of the frequency
deviation sampled over the
averaging period τ
Clock under test
yn+1
yn
τ τ
Reference clock

Long Term Stability: Allan Deviation
D. Allan, 1966
 Allan deviation is a statistical  Allan deviation is plotted vs.
measurement of frequency the averaging time τ.
instability  High Allan deviation →
 Differences between successive frequency varies a lot around
readings of the frequency the nominal frequency over
deviation sampled over the measurement period τ.
averaging period τ
Allan deviation σΑ(τ), ppm

Clock under test
yn+1
-1 +1
yn
-0.5 +0.5
Quantization Correlated
White Flicker Bias Random Frequency
noise, white noise,
noise, angle noise instability rate walk drift
phase noise, sinusoidal
τ τ flicker phase
random
walk
noise,
Reference clock Averaging time τ


Allan Deviation & Calibration Power
 Allan deviation of two 32kHz oscillators
shown
 Same power consumption (~200nW)
 Similar Allan deviation floor (3-5ppm)
 In the system, oscillators are periodically 100ppm
calibrated to a more accurate, high
power, clock source
 To calibrate to 100ppm accuracy
 Blue requires 1.5ms
 Red requires 40ms
 Much higher power penalty to calibrate
red vs. blue
 Allan deviation floor is not the only
important FOM

Integrated Reference Example Use Case
Battery
radio SoC
ADC
Power  All wireless
management
SoCs use
modem
integrated
DAC ROM NVM oscillators for
multiple non-
MCU
UART, radio functions
PFD
Output I2C, SPI
divider
 32kHz + MHz
Feedback common
divider
PLL Integrated  Often
oscillators calibrated to
Temperature Real Time
crystal
Sensor Sensor
Clock Interface
ADC Sensor oscillator

State-of-the-Art in Integrated Oscillators
Integrated RC oscillators
 Recent advancements:
Temperature coefficient of frequency,

published 2016-2023
 Lower temperature
coefficient of frequency
(TCF) with improved energy
efficiency
 Only same-polarity TC
resistors required
 No need for zero-TC or
positive + negative TC
resistors
ppm/ºC
 Improved performance with
1-point trim instead of 2+
 Aging compensation Energy efficiency, pJ/cycle

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

Ceramic Resonators
 Ceramic resonators
 Passive three terminal element of piezoelectric ceramic
 Resonance frequency variation of <0.5%
 Quality factor ~1000
 Useful for applications such as
 Microcontrollers that need better accuracy than what is
achievable with integrated oscillators over their lifetime
(limited by frequency drift/aging to ~1%)
 Temperature compensated oscillators over wide temp ranges
 Oscillators made from ceramic resonators are
 Lower performance, lower cost than quartz crystals
oscillators
 Higher stability, higher cost than integrated oscillators

Modified Butterworth Van-Dyke (MBVD)
 MBVD model is electrical equivalent

model for ceramic (and crystal)
resonators Electrical Equivalent Model:
Modified Butterworth Van-Dyke (MBVD)
 LM=motional inductance
 CM=motional capacitance Rm Cm Lm
 RM=equivalent loss
 C0=parasitic capacitance from CO
package leads
 Inductive between series and 1 𝐿𝐿𝑚𝑚
𝑄𝑄 =
parallel resonance frequencies 𝑅𝑅𝑚𝑚 𝐶𝐶𝑚𝑚
 High impedance at DC

Quality Factor, Q
 fS = series resonance freq
 fP = parallel resonance freq 𝑓𝑓𝑃𝑃
𝑄𝑄 =
 Q = 2π*energy stored / ∆𝑓𝑓3𝑑𝑑𝑑𝑑
energy dissipated per cycle
 Higher Q 1 𝐿𝐿𝑚𝑚
 Lower loss 𝑄𝑄 =
fs fp 𝑅𝑅𝑚𝑚 𝐶𝐶𝑚𝑚
 Lower power consumption to
sustain oscillation
 Better stability
 Lower phase noise
 Longer startup
1 1 𝐶𝐶0 + 𝐶𝐶𝑚𝑚
𝑓𝑓𝑠𝑠 = 𝑓𝑓𝑝𝑝 =
2𝜋𝜋 𝐿𝐿𝑀𝑀 𝐶𝐶𝑚𝑚 2𝜋𝜋 𝐶𝐶0 𝐶𝐶𝑚𝑚 𝐿𝐿

Quality Factor, Q
 fS = series resonance freq
 fP = parallel resonance freq 𝑓𝑓𝑃𝑃
𝑄𝑄 =
 Q = 2π*energy stored / ∆𝑓𝑓3𝑑𝑑𝑑𝑑
energy dissipated per cycle
 Higher Q 1 𝐿𝐿𝑚𝑚
 Lower loss 𝑄𝑄 =
fs fp 𝑅𝑅𝑚𝑚 𝐶𝐶𝑚𝑚
 Lower power consumption to
sustain oscillation
 Better stability
 Lower phase noise Ceramic resonator
 Longer startup Q ~1000
1 1 𝐶𝐶0 + 𝐶𝐶𝑚𝑚
𝑓𝑓𝑠𝑠 = 𝑓𝑓𝑝𝑝 =
2𝜋𝜋 𝐿𝐿𝑀𝑀 𝐶𝐶𝑚𝑚 2𝜋𝜋 𝐶𝐶0 𝐶𝐶𝑚𝑚 𝐿𝐿

Two Single-Ended Oscillator Topologies
Pierce oscillator Colpitts oscillator
I I
ceramic X2
resonator/
crystal Rfb
C1
X1 X1
M1 M1
C1 C2
C2
ceramic Vbias
resonator/
crystal
 Two pin oscillator  One pin oscillator

 Pseudo-differential signal on  Single-ended signal on
pins (lower spurs/pulling) pin (high spurs/pulling)
 Commonly used

Two Single-Ended Oscillator Topologies
Pierce oscillator Colpitts oscillator
I I
ceramic X2
resonator/ Other single-ended
crystal Rfb topologies:
C1
Butler, Modified Butler,
X1 X1 Clapp.
M1 M1
C1 C2
C2 Differential topologies
ceramic Vbias
resonator/ occasionally used
crystal
 Two pin oscillator  One pin oscillator

 Pseudo-differential signal on  Single-ended signal on
pins (lower spurs/pulling) pin (high spurs/pulling)
 Commonly used

Ceramic Resonator Example Use Case
 Low noise 30ksps 24b ADC
 Supports 4-8MHz ceramic resonators or crystals
 Ceramic resonator advantages vs crystal: ADS1255
 Low cost
 Higher reliability
 Higher ESD
tolerance
 Faster start due to
lower Q than crystal

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

Microelectromechanical Systems (MEMS)
 Electromechanical structures that vibrate (kHz – GHz)
 High Q possible: 1000+
 Enables low power oscillators with low phase noise
 MEMS oscillator: electromechanical resonator + active circuitry
 Requires active + passive temperature compensation
 Possible to achieve:
 kHz: <100ppm freq variation, <3µW
 GHz: <20ppm freq variation, <1mW
 Small size, can be co-packaged with CMOS
 More robust in harsh environments (shock, vibration) compared to quartz
crystals

Examples of 32kHz MEMS Resonators
H. Barrow
D. Serrano
2012 Freq Control
2011 Freq Control
Symp.
Symp.
 Fewer mask steps compared to
standard CMOS (~10 vs. 30+)
 High drive voltage (Vp) required  Tuning voltage applied to adjust
(>3V) to excite structure into oscillator frequency
resonant mode

Commercial Low Freq MEMS Oscillators
Examples at 32.768kHz:
Freq Stability Size, Cost Resonator

Manufacturer Part Number Current
over Temp mm2 (Digikey) /Oscillator
SiTime SiT1532 1.4μA ±100ppm 1.5x0.8 $0.89 Oscillator
Vectron/
MT-9560A ~1μA ±10ppm 1.5x0.8 - Oscillator
Microchip
WMRAG32K76
Murata CS1C00R0 N/A ±20ppm 0.95x0.6 $0.70 Resonator
(discontinued)
 Limited number of companies producing low frequency MEMS

 32.768kHz crystal cost is ~$0.30, MHz ceramic resonators ~$0.12 in similar
volumes
 MEMS more expensive, but useful for small size & harsh environments

Examples of MHz MEMS
 SiTime  Berkeley Sensor &
 Rings in extensional stress, Actuator Center
center anchors for low loss  20µm radius
 f = 47MHz, Q = 147k  f=167MHz, Q=4536
M. H. Roshan, JSSC 2017 Q. Xie, 2020 IEEE IUS

GHz MEMS
 Piezoelectric material surrounded by electrodes
 Piezoelectric converts electrical ↔ mechanical energy
 Surface Acoustic Wave (SAW) resonator → mechanical resonant
behavior in the surface of the material
 BAW (Bulk Acoustic Wave) resonator → mechanical resonant behavior
in bulk of material
Resonant frequency, fR dependent on

piezoelectric film thickness t and
acoustic velocity vL to first order
𝑣𝑣𝐿𝐿
𝑓𝑓𝑅𝑅 ≈
2𝑡𝑡

BAW Resonator
 Acoustic mirror: Alternating high- and
low-acoustic impedance layers
 Mirror prevents 5µm
 Energy from leaking into substrate or
package
 Contamination → low frequency
drift/aging
 Moisture and helium impacting
resonator → No vacuum cavities or
hermetic packaging required
BAW resonator
 First BAW oscillator published >40 BAW resonator
cross section
years ago 600µm
 BAW oscillators used in commercial
products since 2019 BAW stacked
on CMOS die 400µm

Passive Temperature Compensation
 MEMS resonators expand as
temperature increases
 Temperature coefficient of
frequency (TCF) is -25ppm/°C.
 >3000ppm Δfreq over industrial
temperature range
 SiO2 has positive TCF
 Adding an SiO2 layer can reduce
effective TCF to ±0.5ppm/°C
 Manufacturing tolerances limit
passive temperature
compensation to <300ppm
 Not sufficient for all
communications protocols

Active Temperature Compensation
 Uncompensated BAW resonator: ∆F vs.
T ~3000ppm
 BAW resonator with passive temp.
compensation: ∆F vs. T ~300ppm Look up Temp
 Still insufficient for many wireless Passively
table Sensor
communication applications compensated Temperature
 BAW resonator with passive + active BAW resonator, correction
300ppm
temp. compensation: ∆F vs. T ~20ppm
 Compensation limited by temperature Compensated
sensor accuracy Sustaining clock, 20ppm
 Compensation can be continuous or amplifier
periodic

Oscillator Design Constraints
 BAW resonator stacked on and 1.7V
wire-bonded to CMOS die
 Four modes possible
 Latched if Cd too small Cd
 Parasitic relaxation oscillation if Ibias
Cd>nCL/2 [Thirunarayanan, M2 M1
Rm LBW
2011]
Cm CO
 Desired oscillation (2-3GHz)
 Parasitic oscillation at fB (5- BAW+ BAW- CL
8GHz) Lm
Equivalent
M4 M3 oscillator
1 capacitance
𝑓𝑓𝐵𝐵 =
𝐶𝐶𝐿𝐿 𝐶𝐶𝑂𝑂 Cd
2𝜋𝜋 𝐿𝐿𝐵𝐵𝐵𝐵
𝐶𝐶𝐿𝐿 + 𝐶𝐶𝑂𝑂
Thirunarayanan, 2011

Active Aging & Frequency Drift
 Aging: the long-term
oscillator frequency drift
 Contamination &
outgassing cause freq drift
 Oscillator operated
continuously and frequency
drift measured
 Logarithmic aging observed
 Resonator stress
relaxation
 CMOS transistor aging
 More frequency drift at
higher temperatures

Mechanical Shock
TEST STANDARD METHOD DESCRIPTION
Mechanical MIL-STD-883H 2002.5, Acceleration peak 1,500g | Pulse duration 0.5ms
Shock (QSS 009-119) Level B | 3 perpendicular axes (x, y, z) | 5 shocks
Guide rods Shock table

DUT
Impact block Signal source
analyzer
Accelerometer
Power supply
Shock table controller

Measured Results - Shock
 SoC frequency
SoC with integrated BAW SoC with external crystal
measured with both 8 8 8 8 8
(DUT1-Down) (DUT2-Down) (DUT3-Down) (DUT1-Down) (DUT2-Down)

integrated BAW and 6 6 6 6 6
external crystal 4 4 4 4 4
 SoC with integrated Impact
Frequency Stability [ ppm ]
Frequency Stability [ ppm ]

2 2 2 2 2
BAW shows 4x
better shock
0 0 0 0 0
immunity than with -2

8
-2
8
-2
8
-2
8
-2
8
(DUT1-Up) (DUT2-Up) (DUT3-Up) (DUT1-Up) (DUT2-Up)

external crystal 6 z6 6 6 6
 BAW resonator less 4 x 4 y 4 4 4
affected by shock 2 2 2 2 2
than crystal due to

lower mass
0 0 0 0 0
-2 -2 -2 -2 -2
0.021 0.023 0.025 0.021 0.023 0.025 0.021 0.023 0.025 0.021 0.023 0.025 0.021 0.023 0.025
Time [ sec ] Time [ sec ]


MEMS Reference Example Use Case
 Bluetooth Low Energy radio
used with BAW resonator as a
frequency reference for radio
PLL
 BAW resonator co-packaged
with SoC
 <+/-40ppm frequency
accuracy achieved, including
10 year aging, sufficient for
radio operation
 Space saved by removing
crystal allows more sensors to
be used in the same footprint

State-of-the-Art MEMS Oscillators
 GHz resonator/oscillator  kHz resonator/
performance: oscillator performance
 FOM at 1kHz offset better  <±100ppm stability with passive
than -200dB temperature compensation
 Q>1500  <±3ppm stability with active +
 <±20ppm stability (including passive temperature
aging) with active + passive compensation
temperature compensation
 MEMS size:
 MHz resonator/oscillator  Packaged <1.5x1.0mm2
performance:  Resonator <0.5x0.5mm2
 Q>30,000
 ~±20ppm stability with only 𝑓𝑓𝑜𝑜 𝑃𝑃𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
𝐹𝐹𝐹𝐹𝐹𝐹 = 𝐿𝐿 𝑓𝑓𝑚𝑚 − 20 log10 + 10 log10
passive temperature compensation 𝑓𝑓𝑚𝑚 1𝑚𝑚𝑚𝑚
B. Bahr, RFIC 2022 W. Chen, IDL 2021 S. Zaliasl, JSSC 2015

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

Quartz Crystals - MHz
 Quartz crystals have stable resonance
frequency vs. temperature with proper cut
 Cubic (s-shape) with «AT-cut» most
commonly used for 8MHz - 80MHz
 ±20ppm to ±40ppm vs. over operating
temperature -40 to 85ºC
 Up to 300MHz crystals in development
 A quartz crystal is specified as frequency X
with capacitive load Y (e.g. 48MHz at
CL=9pF)
 CL impacts the startup time, power
consumption, and pulling sensitivity
 Packaged size is often between 1.0x0.8mm2
to 3.5x2.5mm2.

Quartz Crystals - kHz
 Below 1MHz, tuning forks are dominant
 32.768kHz is often used for timing
 Quadratic temperature coefficient of
frequency (TCF) Vibration motion
 ±100ppm from -40 to +85°C for 32kHz
crystals
 Variations in cut cause variations in TCF
 Why 32.768kHz?
 32.768kHz =215
 Precise 1s period with 15 stage binary
counter (low power, low complexity)

Crystal Oscillators
 Quartz crystal resonator + active circuitry =
crystal oscillator
 Advantages
 Mature technology (100 years old in 2021)
 Quality Factor>30,000: Low power consumption
(<10μW/MHz)
 Stable frequency – <100ppm variation -40°C to 85°C
 Disadvantages
 Cost (equivalent to 1-2mm2 of silicon)
 Limits integration, size 2.0x1.6mm2 is common
 Degraded frequency stability at temperature extremes
 Relatively slow startup time
 Sensitivity to shock and vibration
 Frequent supply issues (12-26 week lead time is common)

Classic Crystal Oscillator Architecture
Vittoz, JSSC 1988

VDD
Iout
Vin
Oscillator Amplitude regulator


 Oscillator needs to provide enough “negative resistance” (Rn) to
overcome losses in the crystal (ESR)
 Rn > 5*ESR to give reliable oscillator startup. Vittoz, JSSC 1988
VDD
Iout
Vin


VDD
 However, to sustain oscillation
Rn>ESR only. Iout
Vin


VDD
Rn>ESR only. Iout
 Higher Rn requires higher power
consumption
Vin


VDD
Rn>ESR only. Iout
 Higher Rn requires higher power
consumption
 Add amplitude control loop
 High power at startup, lower in
steady state Vin
 Vittoz architecture used extensively
for both kHz and MHz oscillators


Crystal Oscillator Negative Resistance
 Need negative resistance C0
Rn>>ESR to ensure reliable
startup (typically Rn>5∙ESR)
Lm Cm Rm
crystal model
 Need Rn>ESR to sustain
oscillation
gm
 Need high gm for robust operation 2CL 2CL
but also want low current
consumption Rfb
 Maximum gm/I is in subthreshold
where gm=I/(n0VT) −𝑔𝑔𝑚𝑚 𝐶𝐶𝑜𝑜
2
 Subthreshold slope factor n0 ∼ 1.3 𝑅𝑅𝑛𝑛 = 𝐸𝐸𝐸𝐸𝐸𝐸 = 𝑅𝑅𝑚𝑚 1+

2𝜋𝜋𝜋𝜋 2 2𝐶𝐶𝐿𝐿 2
𝐶𝐶𝐿𝐿
 Thermal voltage VT=kT/q=25mV
at room temperature

Crystal Oscillator Frequency Tuning
 Frequency tuning by adjusting C0
of capacitive loading on the
crystal.
Lm Cm Rm
crystal model
 Capacitive loading is series
combination of C1 and C2 in
gm
Pierce and Colpitts topologies. C1 C2
 If C1=C2, CLeq=C1/2=C2/2
 CLeq must equal the crystal- Rfb
specified CL to achieve the
correct oscillation frequency.
𝐶𝐶𝑀𝑀 𝐶𝐶𝑀𝑀
 CLeq too small, frequency high. 𝑓𝑓𝐿𝐿 = 𝑓𝑓𝑠𝑠 ⋅ 1 + ≈ 𝑓𝑓𝑠𝑠 ⋅ 1 +
𝐶𝐶0 + 𝐶𝐶𝐿𝐿𝑒𝑒𝑒𝑒 2 𝐶𝐶0 + 𝐶𝐶𝐿𝐿𝑒𝑒𝑒𝑒
 CLeq too large, frequency low.

Crystal Oscillator Startup Time
 Startup time constant is
−2 − 2Lm
τ = =
(Rn + Rm ) ⋅ (ω 2Cm ) Rm + Rn
 Stable amplitude takes 7-15τ
 Example parameters from 24MHz crystal:
Crystal size Lm Cm Example startup time Pulling sensitivity at CL=7pF
3.2x2.5mm 12.6mH 3.49fF 176µs -27ppm/pF
2.0x1.6mm 22.5mH 1.95fF 315µs -15ppm/pF
 Lower pulling sensitivity → less sensitivity to CL changes →

slower startup time → more energy for startup
 Tradeoff between frequency stability & startup energy

XO Startup Time Reduction
 Smaller crystal package size → Higher
Rm (loss)+larger Lm → slower XO
startup time → higher startup energy
 Lowest cost crystal size drops over crystal
time oscillator
injection
 Increasing energy consumption with oscillator
each new generation.
 One solution: kick-start crystal
oscillator
 Must be robust
 Systems must assume worst case
startup time → an XO that only starts
up quickly sometimes is not useful

XO Fast Startup Techniques
Examples (not a complete list):
Fast Startup Method Limitation / Comments

1. Use higher bias current on initial Has been used for >35 years, but there is a limit on how much
startup, lower in steady state improvement is possible
2. Use lower load capacitance, CL Crystals with only certain values of CL are manufactured.
Larger frequency shift from parasitics, environmental changes
3. Use a higher frequency crystal Higher steady state power consumption.
4. Inject a chirp frequency into Simple but limited effectiveness
crystal
5. Inject crystal with a tone at the Need a separate oscillator calibrated to better than ±0.5%
resonance frequency accuracy or with dithered injection frequency
6. Synchronized/self-timed injection Need to detect peak. Very fast startup time possible
7. Injection with phase error Need 1% accurate clock at much higher frequency than crystal
detection
See K.-M. Lei, IEEE TCAS-II 2021 for more details

Phase Noise
 Ratio of noise power to carrier
power
 Frequency domain
representation of noise
Phase noise, dBc/Hz

 Units are dBc/Hz at offset freq
 Ex: -80dBc/Hz at 100kHz
offset
 Higher Q tank → lower phase
noise
 High phase noise → Oscillator
energy spread out over wider
frequency range → less stable High Q tank
f0-100kHz f0 f0+100kHz

Phase Noise
 Ratio of noise power to carrier
power
 Frequency domain
representation of noise Low Q tank
Phase noise, dBc/Hz

 Units are dBc/Hz at offset freq
 Ex: -80dBc/Hz at 100kHz
offset
 Higher Q tank → lower phase
noise
 High phase noise → Oscillator
energy spread out over wider
frequency range → less stable High Q tank
f0-100kHz f0 f0+100kHz

Jitter
𝑓𝑓2
1 𝐿𝐿(𝑓𝑓)
 Jitter: Deviation of clock edge 𝜎𝜎𝑗𝑗 = � 10 10 𝑑𝑑𝑑𝑑
from the ideal value 2𝜋𝜋𝑓𝑓𝑐𝑐 𝑓𝑓1
 Cycle-to-cycle jitter: deviation of

period from one clock cycle to the
next
 Period jitter: deviation of the
period from the ideal value
 Long term jitter (N-cycle jitter):
deviation of multiple (N) periods
from the ideal value
 Integrated jitter: Phase noise
integrated over specified frequency
bandwidth and converted to jitter Phase noise L(f)→ Integrated jitter σj
7.6ps from 1kHz – 1MHz

Long Term Stability Comparison
 Both oscillators consume a
few hundred nanowatts 1.E-03
32kHz RC
 RC oscillator requires no 1.E-04 Oscillator
external components
Allan deviation σy
1.E-05
 Crystal oscillator achieves
lower Allan deviation. 1.E-06
32.768kHz
 Higher Q of crystal gives Crystal
1.E-07
lower noise, lower frequency Oscillator
variation, lower Allan 1.E-08
deviation
1.E-09
0.0001 0.001 0.01 0.1 1 10 100
Averaging time τ (s)

32kHz Crystal Oscillator Power Reduction
 New crystal oscillator
architectures are still being
proposed, even though the
first XO was made >100
years ago!
K.-J. Hsiao, ISSCC 2014


32kHz Crystal Oscillator Power Reduction
 New crystal oscillator
architectures are still being
proposed, even though the
first XO was made >100
years ago!
 Example: pulse injection
 Supply energy to the
crystal during short pulses
rather than having a
continually biased
oscillator core.
 First published in 2014
 Cuts power by ~100x
K.-J. Hsiao, ISSCC 2014

Crystal Oscillator Failure Modes
 There are many!
 Just a few examples: Fundamental
jX
 Oscillation at overtone instead 3rd overtone
of desired fundamental mode 5th overtone
 Phase noise degraded due to
spurious modes
Reactance
 Activity dips 0
frequency
 Failure to oscillate due to high
humidity / low impedance
between pins
 Breakage due to shock
-jX
 Improper startup Spurious responses

Crystal Oscillator Example Use Case
 Wi-Fi network processor with 40MHz
crystal oscillator
Wi-Fi network
 Frequency accuracy and phase noise processor
specifications set by RF regulatory
requirements and Wi-Fi standard.
 Phase noise at 40MHz
WLAN
 -125dBc/Hz at 1kHz offset PLL
 -139dBc/Hz at 10kHz offset 40MHz
 -143dBc/Hz at 100kHz offset
 ±20ppm over -40ºC to 85ºC
 Frequency drift <0.16ppm/ms
 Can be challenging with self-heating
due to on-chip power amplifier 2.4GHz / 5GHz

State-of-the-Art XO Designs
 MHz oscillator startup:  32kHz oscillator
 100-1000 cycles power consumption
 Versus 10,000-100,000 cycles for  ~1nW
a traditional design  Versus ~200nW for a traditional
 10x – 1000x improvement design
 MHz oscillator power  MHz and kHz oscillator silicon

consumption: area:
 1-2µW/MHz for design in  ~0.03mm2
production for Bluetooth radio  Includes area for load capacitance
(See W. Kruiskamp, ESSCIRC  Dependent on technology node and
2022) crystal parameters.
 Power is strongly dependent on
crystal parameters & phase noise
specification

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

“Software” TCXO
>+45ppm at 105ºC
Frequency offset, ppm

40 ∆F = 65ppm total
 Crystals have too much frequency variation vs.
temperature for some applications (example:
20
narrowband wireless standards, GPS)
 Need Temperature Compensated Crystal 0
Oscillator (TCXO)
 Can implement a low performance TCXO in -20
software Uncompensated
48MHz Compensated after
RF Frequency
Frequency calibration
Frequency offset, ppm

Synthesizer
40
Temp
Sensor
Firmware 20 <±10ppm
Stored room
computation of possible
estimated
temp freq
crystal 0
measurement
frequency
error
Stored fitting -20
coefficients -40 0 40 80
Temperature, ºC

TCXOs
 Crystal + oscillator + compensation circuitry
packaged together
Crystal
Temperature Fout
 Temperature sensitive reactance dependent Oscillator
compensates TCF of crystal reactance
 Compensation can be analog or digital

 TCXOs compared to crystals:
Compensating
 Cost 3x more Uncompensated response
frequency
 Have ~10-20x better stability vs. temp.
Frequency
 Consume 4x more power as a crystal osc. Compensated
frequency
 Have ~4x longer start-up time
Temp
 Only used when frequency stability
requirements demand it

Types of TCXO Output Clocks
 Clipped sine
 Lower spur level than CMOS
 Requires coupling capacitor and on- VDD1
VDD2
chip buffer to convert to square Clipped
wave sine
TCXO
CMOS level
 VDD1 does not have to equal VDD2
Rbias
 CMOS (rail-to-rail) VDD1
VDD2
 CMOS output can cause worse CMOS

output
spurs than clipped sine TCXO CMOS level
 VDD1 must equal VDD2 PCB SoC
 No coupling capacitor required

TCXO Example Use Case
 IoT standards with narrow
bandwidth require very stable
reference clock
 Example packet error rate vs.
frequency error shows RX sensitivity
degrading beyond ±10ppm
 Crystal oscillators will exceed this
frequency variation at temperature
extremes – TCXO needed
TCXO
Receiver
900MHz carrier,
TX 5kHz deviation

Oven Controlled Crystal Oscillator
 “SC-cut” crystal heated to the
flattest point of frequency vs. “AT-cut” crystal
temperature
 Usually 80-100ºC
 OCXO typical specifications: “SC-cut” crystal
 9mm x 7mm – 70mm x 38mm
 0.5 – 200ppb freq. stability vs temp.
 Warm up time ~ 1 minute
Oven
 Power consumption ~ 1W at 85ºC, Crystal
~4W at 25ºC oscillator OCXO clock
output
 Applications:
 GPS, clock synchronization in IEEE
Oven
1588 networks, radar, control Temperature
instrumentation sensor

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

Atomic Frequency Reference Principles
 Atomic clock frequency determined by
resonant frequency of atoms, a physical
constant
 Unlike oscillators where resonance frequency
is set by properties of a fabricated object
 Electrons in “physics cell” transition
Voltage controlled
between different energy levels when crystal oscillator
excited by the frequency, f0 Microwave f0
Frequency
 Detector and feedback loop adjust f0 to the Synthesizer
frequency with maximum electron Physics
transition Cell
 f0 set by atom control detector

 Rubidium → 6.8GHz
 Cesium → 9.2GHz

Chip-Scale Atomic Clock (CSAC)
 Cesium atoms sealed inside of
optically transparent “physics cells” &
heated to 75ºC to produce a gas
 Laser excites cesium, electron
transition measured by photodetector
 Microsemi SA.45s CSAC
 Power consumption: <120mW
 Aging <0.01ppm/year
 Allan deviation 1x10-11 for τ=1000s
 35grams
 <17cm3
 $2600/each in quantities of 250+ as
of Nov. 2022. https://www.microsemi.com/product-directory/clocks-
frequency-references/3824-chip-scale-atomic-clock-csac

Atomic Clock Example Use Case
 Undersea sensing
 Reflection seismology used for oil
exploration
 An array of sensors on ocean
floor measures reflections from a
sonic pulse Sonic pulse
 Sensor results post-processed to
map layers beneath the ocean
floor Sensors
 Timing accuracy determines
mapping accuracy
 Lower power & higher accuracy of
CSAC vs. OCXO enables smaller
batteries & more sensors → more
detailed maps

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

Frequency Reference Comparison
Allan deviation
Integrated RC Stability vs.
averaging
time window
Crystal
Rubidium
Cesium See also John Vig
crystal tutorial
hour day month


Frequency Reference Summary
Typical values:
Cesium
Integrated
Ceramic MEMS Crystal TCXO OCXO Rubidium (chip
silicon
scale)
Frequency kHz - kHz - kHz -
Hz to GHz MHz MHz MHz MHz
range GHz 300MHz MHz
Size, cm3 1.E-06 0.2 5.E-05 2.E-03 6.E-03 1 50 17
Stability 1.E-02 1.E-03 2.E-05 2.E-05 1.E-06 1.E-07 1.E-09 1.E-11
Power
consumption, 1.E-04 5.E-04 1.E-03 1.E-03 5.E-03 1 10 0.12
W
Cost, US$ 2.00E-04 0.07 0.35 0.15 0.5 25 2300 2600

Conclusions
 Nearly every electronic system needs a clock
 Accuracy, cost, power requirements vary
 Many types of frequency references exist, such as
 Silicon integrated
 Ceramic
 MEMS
 Crystal
 TCXO, OCXO
 Wide spectrum of performance
 8 orders of magnitude in power consumption
 14 orders of magnitude in frequency stability
 Fundamental concepts applicable to all of them

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

Papers to See This Year
 3.4 A 0.01mm2 10MHz RC Frequency Reference with a 1-Point
On-Chip-Trimmed Inaccuracy of ±0.28% from −45°C to 125°C in
0.18μm CMOS
 3.5 A 1.4μW/MHz 100MHz RC Oscillator with ±1030ppm Inaccuracy
from -40°C to 85°C After Accelerated Aging for 500 Hours at 125°C
 3.6 A 12/13.56MHz Crystal Oscillator with Binary-Search-Assisted Two-
Step Injection Achieving 5.0nJ Startup Energy and 45.8μs Startup Time
 3.7 A 16MHz XO with 17.5μs Startup Time Under 104 ppm-ΔF Injection
Using Automatic Phase-Error Correction Technique
 3.8 A 0.954nW 32kHz Crystal Oscillator in 22nm CMOS with Gm-C-
Based Current Injection Control
 3.9 A 0.5-to-400MHz Programmable BAW Oscillator with Fractional
Output Divider Achieving 4ppm Frequency Stability over Temperature
and <95fs Jitter

Outline
 Motivation
 Key metrics
 Ceramic
 Crystal
 TCXO, OCXO
 Key references

References
 A. Paidmarri, et al., “A 120nW 18.5kHz RC oscillator with comparator offset
cancellation for ±0.25% temperature stability,” ISSCC 2013.
 D. Griffith, et al., “A 190nW 33kHz RC oscillator with ±0.21% temperature stability
and 4ppm long term stability,” ISSCC 2014.
 A. A. Abidi, “Linearization of voltage-controlled oscillators using switched capacitor
feedback,” JSSC 1987.
 A. Khashaba, et al, “A 32-MHz, 34-μW temperature-compensated RC oscillator using
pulse density modulated resistors,” JSSC 2022.
 D. Allan, “Statistics of atomic frequency standards,” Proc. of the IEEE 1966
 H. Barrow, et al, “A real-time 32.768-kHz clock oscillator using a 0.0154-mm2
micromechanical resonator frequency-setting element,” 2012 IEEE Int. Freq. Control
Symp.
 D. Serrano, et al, “Tunable piezoelectric MEMS resonators for real-time clock,” 2011
IEEE Int. Freq. Control Symp.
 M. H. Roshan, et al, “A MEMS-Assisted Temperature Sensor With 20- μK Resolution,
Conversion Rate of 200 S/s, and FOM of 0.04 pJK2,” IEEE JSSC 2017

References
 Q. Xie, et al, “167-MHz AlN Capacitive-Piezoelectric Oscillator,” 2020 IEEE IUS
 R. Thirunarayanan, et. al, “Complementary BAW oscillator for ultra-low power
consumption and low phase noise,” IEEE New Circuits and Systems Conf. 2011
 B. Bahr, et al, “Class-C BAW Oscillator Achieving a Close-in FOM of 206.5dB at 1kHz
with Optimal Tuning for Narrowband Wireless Systems,” 2022 IEEE RFIC
 W. Chen, et al, “A Temperature-Stable and Low Impedance Piezoelectric MEMS
Resonator for Drop-in Replacement of Quartz Crystals,” IEEE Electron Device Letters,
2021
 S. Zaliasl, et al, “A 3 ppm 1.5 × 0.8 mm 2 1.0 μA 32.768 kHz MEMS-Based
Oscillator,” IEEE JSSC 2015
 E. Vittoz, et al., “High-performance crystal oscillator circuits: theory and application,”
JSSC, Vol. 23, issue 3, pp. 774-783, 1988.
 K.-M. Lei, et al., “Startup Time and Energy-Reduction Techniques for Crystal
Oscillators in the IoT Era,” IEEE Trans. on Circuits and Systems II, January 2021.
 K.-J. Hsiao, “A 1.89nW/0.15V self-charged XO for real-time clock generation”, ISSCC
2014

References
 W. Kruiskamp, “A Fully Differential 40 MHz Switched-Capacitor Crystal Oscillator
with Fast Start-Up,” ESSCIRC 2022
 https://www.microsemi.com/product-directory/clocks-frequency-
references/3824-chip-scale-atomic-clock-csac
 http://beckelec.com/john-vig-crystal-tutorial.pdf

T1 Fundementals of Freq Referencee

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

T1 Fundementals of Freq Referencee

Uploaded by

Copyright:

Available Formats

ISSCC 2023 Tutorial

Fundamentals of Frequency References

Live Q&A Session: Feb. 19th, 2023, 10:00am-3:00pm, PST

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 2 of 84

© 2023 IEEE International Solid-State Circuits Conference

© 2023 IEEE International Solid-State Circuits Conference

© 2023 IEEE International Solid-State Circuits Conference

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 6 of 84

© 2023 IEEE International Solid-State Circuits Conference

© 2023 IEEE International Solid-State Circuits Conference

 How much frequency varies around the ideal value

© 2023 IEEE International Solid-State Circuits Conference

 Power consumption T=27ºC

D. Griffith T1: Fundamentals of Frequency References 9 of 84

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 10 of 84

© 2023 IEEE International Solid-State Circuits Conference

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 12 of 84

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 13 of 84

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 13 of 84

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 13 of 84

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 13 of 84

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 13 of 84

© 2023 IEEE International Solid-State Circuits Conference

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 15 of 84

© 2023 IEEE International Solid-State Circuits Conference

© 2023 IEEE International Solid-State Circuits Conference

 Current charges capacitor Vin +

© 2023 IEEE International Solid-State Circuits Conference

 Typically 50ppm/ºC - 3000ppm/ºC Vin +

 Supply voltage variation Vref

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 18 of 84

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 18 of 84

© 2023 IEEE International Solid-State Circuits Conference

 Bias voltage Vbias tracks inverter  Still sensitive to R & C temperature

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 20 of 84

© 2023 IEEE International Solid-State Circuits Conference

 Power efficient, stable

D. Griffith T1: Fundamentals of Frequency References 20 of 84

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 20 of 84

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 21 of 84

© 2023 IEEE International Solid-State Circuits Conference

Measured frequency, MHz

D. Griffith T1: Fundamentals of Frequency References 22 of 84

© 2023 IEEE International Solid-State Circuits Conference

© 2023 IEEE International Solid-State Circuits Conference

Allan deviation σΑ(τ), ppm

Reference clock Averaging time τ

© 2023 IEEE International Solid-State Circuits Conference

D. Griffith T1: Fundamentals of Frequency References 24 of 84