A low-power fast start-up crystal oscillator with an autonomous

dynamically adjusted load

Citation for published version (APA):
Ding, M., Liu, Y. H., Harpe, P., Bachmann, C., Philips, K., & van Roermund, A. (2019). A low-power fast start-up
crystal oscillator with an autonomous dynamically adjusted load. IEEE Transactions on Circuits and Systems I:
Regular Papers, 66(4), 1382-1392. Article 8544016. https://doi.org/10.1109/TCSI.2018.2880282

Document license:
TAVERNE

DOI:
10.1109/TCSI.2018.2880282

Document status and date:

Published: 01/04/2019

Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)

Please check the document version of this publication:

• A submitted manuscript is the version of the article upon submission and before peer-review. There can be
important differences between the submitted version and the official published version of record. People
interested in the research are advised to contact the author for the final version of the publication, or visit the
DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page
numbers.
Link to publication

General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.

If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please
follow below link for the End User Agreement:
www.tue.nl/taverne

Take down policy

If you believe that this document breaches copyright please contact us at:
[email protected]
providing details and we will investigate your claim.

Download date: 29. May. 2024

1382 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 66, NO. 4, APRIL 2019

A Low-Power Fast Start-Up Crystal Oscillator With

an Autonomous Dynamically Adjusted Load
Ming Ding , Member, IEEE, Yao-Hong Liu , Senior Member, IEEE, Pieter Harpe , Senior Member, IEEE,
Christian Bachmann, Kathleen Philips, and Arthur Van Roermund , Senior Member, IEEE

Abstract— An energy-efficient fast start-up method for crystal frequency error is several orders worse [2]. To improve the
oscillators is presented, which enables aggressive duty-cycled performance, the oscillator has to consume significantly more
operation of IoT radios to minimize overall power consumption. power (∼mW) [3], which is too much for an IoT radio [4].
A digitally controlled crystal oscillator using the proposed start-
up technique in 90-nm CMOS is presented. Thanks to the As a result, crystal oscillators (XO) have been widely used
dynamically adjusted load, the negative resistance is boosted, as the frequency reference in IoT radios thanks to their high
achieving a 13× start-up time reduction and an overall power quality factor (tens of thousands) [5], which enables high
of 95µW for a 24-MHz crystal oscillator at 1 V. A fully performance [6]–[18].
autonomous feedback loop detects the oscillators envelop and For IoT radios, low-power is essential to achieve long
adjusts the load capacitance at start-up. Thanks to the low-power
start-up circuits, both the start-up time and the start-up energy battery life time. To do that, the transceiver is only activated
are reduced. In addition, the robustness and versatility of the when transmitting or receiving, and disabled to save power
proposed method is verified by measuring quartz crystals with when not needed. In this way, the overall system power
different frequencies and quality factors, as well as measur- is reduced without degrading the performance of the radio.
ing against temperature, supply voltage, and load capacitance But this requires swift start-up behavior for the whole trans-
variations.
ceiver. This is relatively easier for the transmitter/receiver,
Index Terms— Crystal Oscillator, fast start-up, low-power, IoT, the Phase-Locked-Loop (PLL), and the Power-Management-
negative resistance, variation-tolerant. Unit (PMU), whose start-up time is up to a few μs [4]. Due
I. I NTRODUCTION to the high quality factor, the typical start-up time (Ts ) of an
XO is, however, relatively long (∼ms) [6]–[8], [19]–[23]. In

W IRELESS sensor nodes (WSN) in Internet-of-Things

(IoT) applications require an accurate frequency ref-
erence (∼MHz), which generates a stable reference clock
applications with a short off-time, the XO can therefore not be
switched off and on in time. Thus, it has to be always active,
thereby increasing the off-state power of the whole system
for the PLL to synthesize a carrier and to derive clocks significantly. In applications with a short on-time, the XO can
for all other parts of the transceiver SoC. To meet the be switched on and off, but the extra power due to the start-
requirements of IoT radios, high performance is desired for up process can not be neglected. In a duty-cycled radio with a
the frequency reference. For instance, low-phase noise is relatively short packet format (for instance the packet length
required for IoT radios to meet the requirement of adja- can be as short as 128μs in BLE), the power overhead due to
cent channel interference and also transmission modulation the XO start-up process can go beyond 25% [24]. Therefore,
quality. In addition, robust communication among IoT radios a reduction of the XO start-up time of the XO is necessary,
requires a small frequency error as well as high stability and at the same time, the energy overhead to enable a fast
over environmental variations. For example, a popular stan- start-up should be minimized in order to reduce the overall
dard such as Bluetooth-Low-Energy (BLE) requires that the energy consumption.
transmission frequency error should be within ±41ppm (parts The XO start-up time, Ts , not only depends on its quality
per million) of the desired frequency [1]. This leads to a factor and oscillation frequency, but also on its negative
great challenge for on-chip CMOS oscillators, of which the resistance (RN ) [25] and its internal noise [26] at start-up
Manuscript received June 18, 2018; revised September 26, 2018 and Octo- (Fig. 1). Prior-art tried to reduce the start-up time from the
ber 21, 2018; accepted November 5, 2018. Date of publication November 26, following aspects: increase the |RN | [27], or increase the
2018; date of current version March 15, 2019. This paper was recommended initial oscillation energy by injecting another signal at the
by Associate Editor N. Krishnapura. (Corresponding author: Ming Ding.)
M. Ding is with the imec-nl/Holst Centre, 5656 AE Eindhoven, The start-up [24], [28], or a combination of the two techniques
Netherlands, and also with the Eindhoven University of Technology, 5612 AZ [29]. Increasing the transconductance (gm ) of the amplifier by
Eindhoven, The Netherlands (e-mail: [email protected]). providing a higher bias current has been widely adapted to
Y.-H. Liu, C. Bachmann, and K. Philips are with the imec-nl/Holst Centre,
5656 AE Eindhoven, The Netherlands. boost |RN |, thus reducing Ts [27]. This method reduces the
P. Harpe and A. Van Roermund are with the Eindhoven University of start-up time, but at the sacrifice of increased start-up power
Technology, 5612 AZ Eindhoven, The Netherlands. consumption (Fig. 1(a)). Furthermore, the effectiveness of this
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. method is limited by the finite voltage headroom of the devices
Digital Object Identifier 10.1109/TCSI.2018.2880282 in the analog circuitry.
1549-8328 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
DING et al.: LOW-POWER FAST START-UP CRYSTAL OSCILLATOR WITH AN AUTONOMOUS DAL 1383

Fig. 2. Illustration of typical BLE advertising (a) and the fast start-up
facilitated BLE advertising (b).

has been ubiquitously deployed and easily accessible through

Fig. 1. Illustration of start-up time reduction techniques: increasing gm (a) battery-powered mobile devices. Small duty-cycling ratio is
and frequency injection (b). essential to assure a long battery life time, considering the
relatively high power consumption of the radios (∼mW [4]).
The oscillators presented in [24], [28], and [30] increase In a typical BLE advertising event, the main radio is asleep
their initial oscillation energy to reduce Ts by injecting either most of the time, and only activated when broadcasting.
a single-tone [28], [30], or a dithered signal [24]. They During each advertising event, three packets are transmitted
require a separate RC oscillator, and its frequency has to be over three channels CH37, CH38, and CH39 consecutively,
calibrated sufficiently close to the resonant frequency of the which are located at frequencies of 2402MHz, 2426MHz, and
crystal oscillator (Fig. 1(b)). This is not convenient because 2480MHz respectively (Fig. 2(a)). The shortest packet has
the frequency of an RC oscillator is vulnerable to PVT a short format of 128μs. The event interval between two
(process, supply voltage, and temperature) variations. [30] is consecutive advertising events is between 10ms to 10.24s,
able to overcome most issues, but requires a very good PVT- during which the BLE transceiver has sufficient time to be
compensated design in order to guarantee the performance. switched off and on to save power. However, within each
To make the frequency injection method less sensitive to PVT advertising event, the interval between each packet is no more
variations, Chirp Injection (CI) is introduced in [23] and [29] than 10ms, which is in the same order as typical start-up
by sweeping the frequency of the injection oscillator over a time of crystal oscillators. For this reason, the XO can not
certain range. However, the improvement in reduction of start- be switched off and has to be always on, consuming power
up time is then limited. while the BLE transceiver is in off-state. To minimize the
In this work, to reduce the start-up time while minimiz- power consumption, the time interval between two packets is
ing the overhead in energy, area and system complexity, a kept short. However, this increases the chance that two or three
dynamically-adjusted load (DAL) method is proposed [31], advertisement packets will collide with one interference and
[32]. The proposed DAL technique reuses the XO clock once the whole advertisement event will be blocked.
it is available to achieve an automatic control as shown later. Alternatively, if the XO start-up time can be reduced suffi-
The proposed technique achieves a short start-up time (200μs) ciently in a power-efficient way, the XO can also be switched
and low start-up energy (<40nJ). off between two advertising packets (Fig. 2(b)). In this way, the
Section II introduces an application which requests aggres- power consumption of the XO can be reduced with a factor of
sive duty-cycling and thus requires an energy-efficient start- more than 10×, leading to significant overall power consump-
up method for the crystal oscillator. Section III introduces the tion reduction. Moreover, thanks to the short XO start-up time,
design considerations of a crystal oscillator. In section IV the the interval between two advertising packets can be increased
implementation of the crystal oscillator is shown. The mea- without introducing power overhead of the XO. In this way,
surement results will be presented in section V and conclusions the chance that all the three advertising packets collide with a
will be drawn in section VI. strong interference is reduce significantly, leading to a better
communication quality.
II. FAST S TART-U P FOR BLE A DVERTISING The essential to enable the proposed BLE advertising is a
As mentioned, WSNs rely on a small duty-cycling ratio power-efficient start-up approach for crystal oscillator, which
to reduce its average system power consumption. BLE radios will be introduced in the next section.
1384 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 66, NO. 4, APRIL 2019

Fig. 4. Comparison of gm method and DAL method.

to start-up is proportional to Rm (2ωCL )2 [25], indicating that

less power is needed for crystal oscillators with smaller CL
for start-up. Therefore, in the start-up phase, a smaller CL is
desired to reduce both start-up time and power consumption.
In the stable phase, the frequency pulling factor (∼ 2CpC+2C
m
L
)
[25] indicates the difference between the real oscillation fre-
quency and the intrinsic resonant frequency of a quartz crystal.
Fig. 3. Start-up behavior illustration (a) and simplified block diagram of a
Pierce crystal oscillator with lumped model (b). A large CL is preferred for a quartz crystal as it results in
a small frequency pulling factor and thus a small frequency
III. DYNAMICALLY-A DJUSTED L OAD error, which is desired for radio. In addition, a large CL also
M ETHOD FOR FAST S TART-U P improves stability over environmental variations.
Reference [17] proposes to increase RN by isolating the
For a crystal oscillator in a duty-cycled wireless sensor parasitic shunt capacitor Cp of the quartz crystal during start-
node, once enabled, the process can be roughly divided into up. Reference [18] proposes to increase RN by programming
two phases: start-up phase and stable phase Fig. 3(a). In the the load capacitance to one small value and to another large
start-up phase, the amplitude of the oscillator ramps up and value after a delay, but did not provide insight how to optimize
reaches a stable value in the stable state, in which the crystal start-up time and energy and lacks the feedback loop to
oscillator clock can be used by other building blocks of the automatically increment the load capacitor CL . In practice,
chip. In the stable phase, the performance of the crystal the relation between RN and gm is more complicated (eq. 1),
oscillator should be sufficient for the RF blocks. In addition, which is illustrated in Fig. 4. The required gm of a crystal
the overhead of the start-up method in energy, area and system oscillator is proportional to (ωCL )2 [25]. For example, for a
complexity should be minimized. crystal oscillator with 10pF load capacitance, gm has to be
Fig. 3(b) shows the simplified block diagram of a popular relatively high for fast start-up. Compared to the gm method,
Pierce crystal oscillator. The crystal quartz can be lumped the DAL method temporarily sets the load capacitance to a
with Lm , Cm , Rm and Cp as the motional inductor, capacitor, smaller value such that during start-up a much smaller gm is
resistor and parallel parasitic capacitor. In order for a crystal necessary to achieve the same |RN |. In this way, the DAL
oscillator to start-up, the loss of a quartz crystal Rm should method is more energy-efficient compared to the gm method.
be compensated by a negative resistance RN , which is given For each load CL , |RN | increases with gm within a certain
by [25] range. However, if the gm is larger than its value where |RN |
gm C1 C2 is optimal, |RN | will decrease due to the gm -boosted effect on
|RN | = , (1)
(gm Cp )2 + ω2 (C1 C2 + C2 Cp + C1 Cp )2 Cp , which is not desired (Fig. 4). It is worthwhile to mention
that a larger CL leads to a larger theoretical peak |RN | (Fig. 5),
where C1(2) = 2CL . In the start-up phase, the key to reduce but at the cost of much higher power as mentioned before.
the start-up time is to increase |RN | as much as possible. For In addition, to quantitatively analyze the start-up time Ts
a crystal oscillator, C1(2)  Cp , and RN can be approximated and start-up energy Es for a given crystal oscillator, they are
g
by − (2ωCm )2 , where ω is the oscillation angular frequency and computed and simulated versus the transconductance gm and
L
load capacitance CL equals 0.5C1(2). It can be seen that RN load capacitance CL in Matlab. The start-up time for a given
is approximately quadratic to 1/CL , so reducing CL is very crystal oscillator can be estimated as [26], [29]:
effective to obtain a higher |RN | at start-up. Considering that  
gm can be limited by the voltage headroom of the analog cir- 2L m 0.9ωC T V D D
Ts = ln , (2)
cuits [27], reducing CL can overcome the bottleneck of voltage |R N | − Rm |I M(0) |
headroom to boost RN . This gives an additional advantage for
low supply voltage designs in modern CMOS technologies. where C T is defined as CC11+C
C2
2
+C p , V D D is the supply voltage
In addition, the required minimum gm for a crystal oscillator and I M(0) is the amplitude of the current in the node containing
DING et al.: LOW-POWER FAST START-UP CRYSTAL OSCILLATOR WITH AN AUTONOMOUS DAL 1385

Fig. 5. Simulated |RN | versus gm and load capacitance CL .

the quartz crystal. Ts is the time to reach steady-state, and

2L m
|R N |−Rm equals the time constant, τ , of the amplitude expo-
nential growth [26]. The relationship between the start-up time
Ts and gm and C L is calculated and shown in Fig. 6. For each
load capacitance, there is an optimum gm,opt that results in a
minimum start-up time Ts . This is because |R N | is maximal Fig. 6. Simulated start-up time Ts versus gm and load capacitance CL .
at gm,opt . Ts increases at a smaller or larger gm due to the L m = 5.8mH, Rm = 20, ω = 2π × 24M rad/s, V D D = 1V, I M(0) = 5nA.
decrease of the |R N |. When |R N | approaches Rm , Ts goes up
rapidly. Once |R N | drops below Rm , the crystal oscillator will
fail to start up. The optimum value for gm , gm,opt , is expressed
as [25]:
 
C1 C2
gm,opt = ω C1 + C2 + . (3)
Cp
The equation shows that gm,opt is proportional to its load
capacitance C1(2) (Eq. 3). Therefore, for crystal oscillators
with smaller load capacitance C L , a smaller gm is needed.
The start-up time at gm,opt can be expressed as:
 
2L m 0.9ωC T V D D
Ts,opt = ln .
1/(2ωC p (1+ (C1 +C2 ) C p )) − Rm
C1 C2
|I M(0) |
(4)
The theoretical maximum |R N | for crystal oscillators with
smaller C L is lower than that with higher C L (Fig. 4 and
Fig. 5). As a result, the Ts,opt for crystal oscillators with
smaller C L is larger than that with larger C L (Fig. 6). The Fig. 7. Simulated start-up energy Es versus gm and load capacitance CL .
measured start-up time of this realized crystal oscillator (two L m = 5.8mH, Rm = 20, ω = 2π × 24M rad/s, V D D = 1V, I M(0) = 5nA.
points as shown in Fig. 6) is close to the calculated optimum
point for C L = 2pF. be observed that if the circuit is optimized for startup-time
With the calculated start-up time, the total energy during Ts , this leads to a near-optimal start-up energy E s directly.
the start-up is further calculated as follows: However, a circuit optimized for start-up energy E s is not
E s = V D D × Id × Ts necessarily achieving the optimum start-up time Ts . As gm
  further increases, due to the |R N | degradation, see Fig. 4,
VD D L m gm 0.9ωC T V D D
= ln , (5) the start-up time will increase, and thus the start-up energy will
VO V |R N | − Rm |I M(0) |
increase. The minimum start-up energy for a crystal oscillator
where VO V is the overdrive voltage of the transistor that gener- with C L = 2pF is approximately 200× lower than that with
ates the negative resistance and Id is the current consumption C L = 50pF. This gives an indication that start-up with smaller
of the crystal oscillator, which can be estimated as gm VO V /2 C L can save energy significantly, which is also shown in Fig. 7.
assuming VO V is constant. As shown in Fig. 7, for each To optimize both the start-up performance and the fre-
load capacitance C L , there is an optimum range for gm that quency stability of the stable state, in this work, an XO with
results in a minimum start-up energy. Within this range, as gm a dynamically-adjusted load (DAL) technique is proposed,
increases, start-up time Ts decreases, and the start-up energy minimizing the CL in the start-up phase at first for fast start-
remains the same. When combining Fig. 6 and Fig. 7, it can up and thereafter incrementing CL for stable operation in the
1386 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 66, NO. 4, APRIL 2019

monitoring the amplitude of the oscillator, and will decide if

the crystal oscillator has a sufficient output swing to provide
a clock for the digital circuit. If so, a finite state machine
(FSM) will automatically increase C1a(2a) to the targeted value.
During the start-up, the frequency of the crystal oscillator
deviates from the nominal frequency (maximum hundreds
of ppm as shown later) because of the frequency pulling effect
due to the varying capacitor value (C1a(2a) ). This frequency
offset error is not a problem because it is still sufficient for
operation of the digital circuits in the feedback loop. Besides,
once C1a(2a) is set, the oscillator frequency settles to the correct
value automatically after a few μs delay. After this latency the
RF circuits can be switched on, as the oscillator is fully settled
then. Note that the whole start-up process is fully autonomous
without requiring an extra oscillator or sequence. This makes
it possible to implement the DAL method with only an enable
signal, which is very convenient for many IoT applications.
In addition, it is worthwhile to mention that, thanks to the
switched capacitor bank, the frequency of the oscillator can be
trimmed by configuring the switched capacitor banks. In this
way, the proposed DAL is compatible with the implementation
of a digitally-controlled XO (DCXO), which is desired for
many WSN [33]. In those applications, cheap quartz crystals
are preferred, however, they usually have larger frequency
variations and require frequency trimming.
Fig. 8. Comparison of various start-up approaches with respect to load
capacitance (a) and XO frequency (b). A. Oscillator Core
The Pierce oscillator core uses an NMOS transistor to
stable phase. As a result, both the start-up time and the start- implement the negative gm for negative resistance. The bias
up energy can be reduced. In the stable phase, the operation current circuit provides a reconfigurable current for the oscil-
condition (e.g, load capacitance) of the crystal oscillator is lator from a 20μA bias current input to compensate the varia-
correctly configured, and the performance in the stable phase tions due to process spread, environmental fluctuation, or dif-
is not sacrificed as shown later. ferent quartz crystals. Since the noise of the crystal oscillator
In the proposed method, the load capacitance at the start- will be up-converted when it is used as the reference clock for
up remains similar regardless of the CL used in the stable a PLL, the noise of the crystal oscillator should be optimized.
phase of different quartz crystal (Fig. 8(a)). This allows a The ratio of the current mirror is minimized to reduce the noise
larger CL to be employed to improve the tolerance to the contribution from the current mirror devices. The resistor Rb
parasitics changes without increasing the start-up time. In addi- is set to 40kohm to bias the NMOS while avoiding loading
tion, for the frequency injection method, when the frequency the crystal quartz. The trip-point of the inverter-based buffer is
of the crystal oscillator is changed (for example, due to a optimized to equal the DC level of the crystal core, so that the
different quartz frequency), the injection oscillator has to be gain of the inverter is maximized. As a result, it can amplify
re-designed, otherwise it will not be effective in reducing the the signal of the oscillator core to rail-to-rail so that the input-
start-up time (Fig. 8(b)). The proposed method is less sensitive referred noise of the inverter is minimized.
to such frequency variations since no frequency injection is The load capacitors C1(2) , equal to 2CL , consist of two parts:
required. on-chip switched capacitor banks C1a(2a) and the parasitics
IV. I MPLEMENTATION OF THE C RYSTAL O SCILLATOR C1b(2b) due to on-chip routing, bonding and packaging, and
PCB routing. C1a(2a) is programmable up to 16pF, and is
Fig. 9 shows the block diagram of the fast start-up XO with
implemented with a 6b switched capacitor bank (Fig. 10).
the proposed DAL as well as an illustration of its start-up
C1b(2b) is about 4pF, resulting in a maximum (16pF+4pF)/2 =
behavior in the time domain. It consists of a Pierce oscillator
10pF load capacitance. At start-up, the capacitor bank is
as its core, a reconfigurable bias current circuitry, two switched
switched dynamically, which disturbs the oscillation amplitude
capacitor banks (C1a(2a) ), and a feedback loop to implement
of the crystal oscillator. Assuming the charge on the capacitor
the DAL. The proposed dynamically-adjusted load method
stays the same when switching a capacitor, the voltage vari-
is implemented through the feedback loop, which works as
ation V at nodes XOP and XON due to switching of the
follows (Fig. 9(b)): C1a(2a) is reduced to minimum to boost the
capacitor can be estimated as:
negative resistance RN , facilitating a fast start-up. Therefore,
the amplitude of the oscillator is able to ramp up in a short V0
V = , (6)
time. A clock detector in the feedback loop is continuously 1 + C0 /Cu
DING et al.: LOW-POWER FAST START-UP CRYSTAL OSCILLATOR WITH AN AUTONOMOUS DAL 1387

Fig. 9. Block diagram of the crystal oscillator with autonomous dynamically-adjusted load (DAL) (a) and a waveform illustration (b).

The switches of the capacitor bank are carefully sized to

trade-off between the quality factor of the capacitor bank and
the induced parasitics. In this work, the switches are sized
relatively large (15μm/0.1μm) to assure sufficient quality
factor (700). This is at the cost of an increase of the off-state
parasitic capacitance (1.26pF in total). Each time at start-up,
by configuring the final settled value of C1(2) , the frequency
of the oscillator in the stable phase is trimmed automatically.
The tuning accuracy in this work is <7ppm with a tuning
range larger than ±50ppm, which is sufficient for many IoT
applications, for example, BLE or IEEE802.15.4.

B. Dynamically-Adjusted Load Circuits

Fig. 10. The 6b thermometer-encoded capacitor bank. A feedback loop, which implements the dynamically-
adjusted load method, performs the detection function to
monitor the oscillator amplitude and determines the moment to
where V0 and C0 are the voltage and capacitance of the increase the load capacitance. It consists of a clock detector
capacitor before switching the capacitor, and Cu is the added circuit and an FSM (Fig. 9(a)). The clock detection circuit
capacitance. Depending on the switching moment, C0 can includes two parts (Fig. 11): an amplitude detector which
be any value between 4pF and 20pF, while V0 should be is composed of an envelope detector (ED) in series with a
around 0.5V D D because the capacitor is switched at the edge comparator, and a digital clock detector. The envelope detector
of the clock. The switching induced voltage variation V and the comparator are continuously monitoring the amplitude
V0
starts with a relatively larger value ( 1+(C1b +C 2b )/C u
) at the of the crystal oscillator, by comparing with a reference voltage.
beginning and then gradually reduces to a smaller value Once the oscillator amplitude exceeds this threshold, the
V0
( 1+(C1a +C2a +C 1b +C 2b )/C u
) at the later stage of the start-up. comparator output will trigger the FSM to generate a 6b tuning
Theoretically the unit capacitor should be as small as possible code that gradually ramps up the switched capacitor bank to
to minimize the voltage variation. However, this increases the the final value.
complexity of the capacitor bank and therefore the parasitics. The envelope detector extracts the amplitude information
Therefore, in this work, a 6b capacitor bank is chosen and of the sinewave of the crystal oscillator. It uses a pseudo-
this results in a 250fF LSB size. In addition, the capacitors differential architecture as shown in Fig. 11(a) to suppress the
are switched at the zero crossing point of XOP and XON common-mode variations. The resistor Re and the capacitor Ce
when the differential charge in the load capacitors is minimum, are set to 55kohm and 2pF respectively, resulting in a 700kHz
so that the differential switching induced voltage variation is bandwidth, which effectively attenuates the high frequency
also reduced. The unit capacitor cell is implemented using components. This generates a relatively smooth output of the
a MIM (Metal-Insulator-Metal) capacitor in series with a envelope detector, with an acceptable delay in the DAL loop
complementary switch (Fig. 10). (a few μs).
1388 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 66, NO. 4, APRIL 2019

Fig. 12. Chip photo.

Fig. 11. The amplitude detection circuitry (a) and the digital detection
circuitry (b).

The comparator is continuously comparing the ED output

with a reference voltage. It consists of a three-stage op-
amp and a latch. The input stage consumes only 1μA, and
the second stage 0.5μA. The bandwidth is set as 4MHz,
resulting in a negligible delay compared to that of the envelope
detector. The third stage is an inverter, providing a rail-to-rail
output. The voltage Vedo at the comparator input has ripples
due to the limited high frequency attenuation of the envelope
detector, resulting in glitches at the comparator output, which
is not desired. In order to overcome this issue, a digital
Schmidtt trigger is implemented using a cross-coupled latch
(Fig. 11(a)) [34]. The hysteresis of the latch is designed
to be higher than the ripple at the ED output to avoid
oscillation at the comparator output. The reference voltage of
the comparator uses a resistor-ladder based voltage divider
from the bias circuits.
After the comparator indicates sufficient oscillation ampli- Fig. 13. Measured start-up behavior in two scenarios: without, with DAL.
tude, a digital clock detection circuit is further used to ensure
that the clock quality is sufficient before the load adaptation is has to be proper in both modes. In this work, at the start-up of
initiated (Fig. 11(b)). The logic filters out accidental errors of the oscillator, a relative short transient time (∼10μs) is used to
the comparator, and will only proceed if M correct cycles are compromise between delay and possibility of glitches. After
counted. The counter operates at a lower frequency (divided the start-up of the oscillator, the crystal oscillator clock will
by 4), resulting in an acceptable delay (∼1μs). Once the RDY be used for the other circuits of the chip.
becomes active, the FSM starts to function. It generates a In practice, the circuits that are used for fast start-up
6-bit tuning code DT which increments from 0 to the desired (including the ED, the comparator, and the reference voltage)
value, to control the capacitor banks. The operation frequency can be disabled once the output of the FSM increments to the
of the FSM is kept relatively low (divided by 4), slowing expected value. This is not implemented in this chip.
down the switching transient meanwhile with an acceptable
delay (maximum 10μs). Moreover, the capacitor bank is
thermometer-encoded instead of binary-encoded. In this way, V. M EASUREMENTS
the voltage fluctuations of the crystal oscillator during this To verify the functionality of the proposed DAL method,
transient are minimized, avoiding as much as possible clock a prototype of the test chip is implemented in 90nm LP CMOS
glitches. In a dual-mode crystal oscillator [9], a long transition as shown in Fig. 12. The core area is 0.072mm2, including
time (130μs) is used during switching between two modes. the load capacitor (84%) and the feedback loop to implement
This is to assure that no glitches will happen because the clock the DAL method (10%). A 24MHz quartz crystal, of which
DING et al.: LOW-POWER FAST START-UP CRYSTAL OSCILLATOR WITH AN AUTONOMOUS DAL 1389

Fig. 14. Measured start-up time with respect to temperature, supply voltage
and load capacitance variations.

Fig. 16. Measured current profile without boosting technique (a), and with
DAL (b).

Fig. 15. Measured hang-up of the output clock during startup.

the estimated quality factor is 54800, is used to verify the

functionality of the chip. At 1V, the chip consumes 95μW in
the steady state and the circuits to implement the DAL method
consume 9μW extra at the start-up.
The start-up time is measured as follows: the CLK− OUT
is measured utilizing the amplitude/frequency demodulation
feature of the FSV Rohde&Schwarz signal analyzer, because
it is not desired to measure the amplitude of the oscillator Fig. 17. Measured frequency trimming performance.
directly by probing due to the large capacitive load (namely
tens of pF). In this way, the amplitude and frequency of TABLE I
the output clock can be measured without disturbing the S TART-U P M EASUREMENTS U SING T HREE D IFFERENT
load capacitance of the crystal oscillator. Fig. 13 shows the Q UARTZ C RYSTALS
measured amplitude (left part) and frequency (right part) of the
XO output CLK− OUT during the start-up, which comprises
three phases. Before the oscillator starts up, the amplitude of
CLK− OUT is low and the demodulated frequency is very
noisy. Once the XO starts up, CLK− OUT will be detected
but its frequency is not yet accurate and stable. After the
load capacitor C1(2) is correctly settled, the frequency will the frequency can settle in 200μs, resulting in 13.3× Ts
be stable. The start-up time is defined as the duration for the reduction. If only the gm method is applied, it gives roughly
frequency to settle within ±20ppm from the target frequency, 2× start-up time reduction. The gm can be further increased
which is well within the requirement of many IoT standards, by increasing the power consumption, but that is not desired.
e.g. BLE (±41ppm) and IEEE802.15.4 (±40ppm) [1]. Without As shown in Fig. 15, a temporary hang-up of the output
any start-up technique, the oscillator takes 2.66ms to start clock CLK− OUT occurs during the startup of the oscillator.
up. With the DAL technique only, the frequency can settle This is caused by the switching of the capacitor bank, which
within 375μs, resulting in a 7× Ts reduction. By slightly produces a temporary internal common mode voltage drop on
increasing the power to 146μW for a higher gm of 1mS, nodes XOP and XON. The oscillator core is still running
1390 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 66, NO. 4, APRIL 2019

TABLE II
P ERFORMANCE S UMMARY AND C OMPARISON W ITH S TATE - OF - THE -A RT

correctly, but the common-mode shift moves the oscillator

signal temporarily out of the trip-point of the inverter buffer,
thus resulting in a hang-up of the output clock CLK− OUT
for approximately 6 cycles. This is not a problem because this
only happens at the start-up when switching capacitors. After
that, the capacitor digital control bits are static and no hang-
up will occur. In addition, the start-up time measurements are
repeated with temperature variations, supply voltage variations
and load variations (Fig. 14). With the temperature swept
from −40◦ C to 90◦C, the measured start-up time variation
is 13% and 26.6% without/with DAL. No frequency trimming
is performed and the frequency drifting is less than 20ppm.
With a swept supply voltage from 0.9V to 1.2V, the measured
start-up time variation is 66%. The performance fluctuation is
mainly due to the gm variation. Moreover, with the large load
capacitance varying from 5pF to 10pF, a very stable start-up
performance (3%) is achieved. This matches with the expected
behavior as shown in Fig. 8(a).
To measure the energy during the start-up process, the real-
time current profile of the supply voltage is captured using
Fig. 18. Measured crystal oscillator start-up variations over temperature for
Keysight current analyzer N6705B LXI, as shown Fig. 16. three different crystal devices.
The measured total energy is calculated based on the measured
real-time current. With DAL enabled, the measured total start-
up energy is reduced from 255nJ to 38nJ. Thanks to the low- The frequency tuning range is measured using the Agi-
power circuit for start-up, the energy overhead is minimized, lent 53230A frequency counter. Each frequency is averaged
leading to an approximately 6.7× start-up energy reduction over 1000 measurements. The measured frequency tuning
with a 7× start-up time reduction. When increasing gm and range is roughly [−48ppm, +440ppm] (Fig. 17), among
DAL are both applied at start-up, the start-up time reduction is which [−48ppm, +102ppm] is suggested to be used for
further reduced to 13×. But as mentioned before, increasing frequency trimming, because further reducing the load capac-
only the gm reduces the start-up time at a higher power itance will degrade the performance of the XO. The averaged
consumption. Thus, the start-up energy reduction is similar as frequency trimming accuracy is 5ppm, which is sufficient
that with DAL applied only (6.9×). The energy of the voltage for many IoT applications, such as BLE or IEEE802.15.4.
reference generation is not included because it is located in a It achieves −141dBc/Hz measured phase noise at 100kHz fre-
separate power domain with a separate supply voltage. With quency offset, which is good enough for most IoT applications,
it included, the total start-up energy increases by only 1.5nJ such as BLE or IEEE802.15.4.
and 0.8nJ in the scenario with DAL only and with both DAL To verify the validity of the proposed method, the mea-
and gm . surement is repeated using three different quartz crystals with
DING et al.: LOW-POWER FAST START-UP CRYSTAL OSCILLATOR WITH AN AUTONOMOUS DAL 1391

different package size, frequency or quality factor as shown [12] K.-J. Hsiao, “A 1.89 nW/0.15 V self-charged XO for real-time
in Table I. Crystal oscillators with higher quality factor quartz clock generation,” in IEEE ISSCC Dig. Tech. Papers, Feb. 2014,
pp. 298–299.
have longer start-up time. But the proposed DAL method can [13] D. Yoon, T. Jang, D. Sylvester, and D. Blaauw, “A 5.58 nW crystal
effectively reduce the start-up time for all the three crystal oscillator using pulsed driver for real-time clocks,” IEEE J. Solid-State
oscillators, and with both DAL and gm applied, the start-up Circuits, vol. 51, no. 2, pp. 509–522, Feb. 2016.
[14] S. Iguchi, A. Saito, Y. Zheng, K. Watanabe, T. Sakurai, and
time can be further reduced. In addition, the start-up time M. Takamiya, “93% power reduction by automatic self power gating
measurements against temperature variations are also repeated (ASPG) and multistage inverter for negative resistance (MINR) in 0.7 V,
for all three crystal oscillator as shown in Fig. 18. It can 9.2 μW, 39 MHz crystal oscillator,” in Proc. Symp. VLSI Circuits,
Jun. 2013, pp. C142–C143.
be seen that the measured start-up time is also stable over [15] Y. Rajavi, M. M. Ghahramani, A. Khalili, A. Kavousian, B. Kim,
temperature variations. and M. Flynn, “A 48-MHz differential crystal oscillator with 168-fs
The performance of this work is summarized in Table II. jitter in 28-nm CMOS,” IEEE J. Solid-State Circuits, vol. 52, no. 10,
pp. 2735–2745, Oct. 2017.
Note that the absolute start-up time/energy can be influ- [16] C. Elgaard and L. Sundström, “A 91.52 MHz 840 μW crystal oscillator
enced by many factors such as the crystal quality factor in 28 nm FD-SOI CMOS for 5G applications,” in Proc. ESSCIRC,
and package size. Nonetheless, the proposed method can Sep. 2017, pp. 247–250.
[17] A. Shrivastava, R. Yadav, and P. K. Rana, “Fast start-up crystal oscilla-
improve both the start-up time and start-up energy thanks to tor,” U.S. Patent 8 120 439, Feb. 21, 2012.
its low-power start-up circuit. In addition, the start-up circuits [18] D. N. Tran, “Crystal oscillator circuit having fast start-up and method
are fully autonomously operated and insensitive to crystal therefor,” U.S. Patent 7 639 097, Dec. 12, 2009.
[19] J. Lin, “A low-phase-noise 0.004-ppm/step DCXO with guaranteed
oscillator frequency, making them suitable for future IoT monotonicity in the 90-nm CMOS process,” IEEE J. Solid-State Circuits,
applications. vol. 40, no. 12, pp. 2726–2734, Dec. 2005.
[20] D. Griffith, F. Dülger, G. Feygin, A. N. Mohieldin, and P. Vallur,
VI. C ONCLUSION “A 65 nm CMOS DCXO system for generating 38.4 MHz and
a real time clock from a single crystal in 0.09 mm2 ,” in IEEE
A fast start-up crystal oscillator with a dynamically-adjusted Radio Freq. Integr. Circuits Conf. Dig. Tech. Papers, May 2010,
load in 90nm CMOS is demonstrated. The proposed method pp. 321–324.
[21] Y. Chang, J. Leete, Z. Zhou, M. Vadipour, Y.-T. Chang, and H. Darabi,
features a fully autonomous operation, and has negligible “A differential digitally controlled crystal oscillator with a 14-bit tun-
overhead in power. Thanks to the DAL method, this work ing resolution and sine wave outputs for cellular applications,” IEEE
improves both the start-up time and the start-up energy, which J. Solid-State Circuits, vol. 47, no. 2, pp. 421–434, Feb. 2012.
[22] S. Zaliasl et al., “A 3 ppm 1.5×0.8 mm 2 1.0 μA 32.768 kHz
is very attractive for future IoT applications. MEMS-based oscillator,” IEEE J. Solid-State Circuits, vol. 50, no. 1,
pp. 291–302, Jan. 2015.
R EFERENCES [23] K.-M. Lei, P.-I. Mak, M.-K. Law, and R. P. Martins, “A regulation-
free sub-0.5-V 16-/24-MHz crystal oscillator with 14.2-nJ startup energy
[1] Bluetooth Special Interest Group. (Dec. 2016) Specification of the Blue- and 31.8-μW steady-state power,” IEEE J. Solid-State Circuits, vol. 53,
tooth System, Core Package Version 5.0, Bluetooth Specifications, Blue- no. 9, pp. 2624–2635, Sep. 2018.
tooth Special Interest Group (SIG). [Online]. Available: https://www. [24] D. Griffith, J. Murdock, and P. T. Røine, “A 24 MHz crystal oscillator
bluetooth.org/en-us/specification with robust fast start-up using dithered injection,” in IEEE ISSCC Dig.
[2] Y. Cao, P. Leroux, W. De Cock, and M. Steyaert, “A 63,000 Q-factor Tech. Papers, Jan./Feb. 2016, pp. 104–105.
relaxation oscillator with switched-capacitor integrated error feedback,” [25] E. A. Vittoz, M. G. R. Degrauwe, and S. Bitz, “High-performance crystal
in IEEE ISSCC Dig. Tech. Papers, Feb. 2013, pp. 186–187. oscillator circuits: Theory and application,” IEEE J. Solid-State Circuits,
[3] Y. Satoh, H. Kobayashi, T. Miyaba, and S. Kousai, “A 2.9 mW, vol. SSC-23, no. 3, pp. 774–783, Jun. 1988.
+/− 85 ppm accuracy reference clock generator based on RC oscillator [26] A. Rusznyak, “Start-up time of CMOS oscillators,” IEEE Trans. Circuits
with on-chip temperature calibration,” in Symp. VLSI Circuits Dig. Tech. Syst., vol. CAS-34, no. 3, pp. 259–268, Mar. 1987.
Papers, Jun. 2014, pp. 1–2. [27] U. Karthaus, “A differential two-pin crystal oscillator—Concept, analy-
[4] M. Ding et al., “A 0.8 V 0.8 mm2 bluetooth 5/BLE digital-intensive sis, and implementation,” IEEE Trans. Circuits Syst. II, Exp. Briefs,
transceiver with a 2.3 mW phase-tracking RX utilizing a hybrid loop vol. 53, no. 10, pp. 1073–1077, Oct. 2006.
filter for interference resilience in 40 nm CMOS,” in IEEE ISSCC Dig. [28] Y. I. Kwon, S.-G. Park, T.-J. Park, K.-S. Cho, and H.-Y. Lee, “An ultra
Tech. Papers, Feb. 2018, pp. 446–447. low-power CMOS transceiver using various low-power techniques for
[5] W. Sansen, Analog Design Essentials. New York, NY, USA: Springer, LR-WPAN applications,” IEEE Trans. Circuits Syst. I, Reg. Papers,
2006. vol. 59, no. 2, pp. 324–336, Feb. 2012.
[6] S. Farahvash, C. Quek, and M. Mak, “A temperature-compensated [29] S. Iguchi, H. Fuketa, T. Sakurai, and M. Takamiya, “Variation-tolerant
digitally-controlled crystal pierce oscillator for wireless applications,” quick-start-up CMOS crystal oscillator with chirp injection and neg-
in IEEE ISSCC Dig. Tech. Papers, Feb. 2008, pp. 352–353. ative resistance booster,” IEEE J. Solid-State Circuits, vol. 51, no. 2,
[7] M.-D. Tsai, C.-W. Yeh, Y.-H. Cho, L.-W. Ke, P.-W. Chen, and pp. 496–508, Feb. 2016.
G.-K. Dehng, “A temperature-compensated low-noise digitally- [30] H. Esmaeelzadeh and S. Pamarti, “A quick startup technique for high-Q
controlled crystal oscillator for multi-standard applications,” in IEEE oscillators using precisely timed energy injection,” IEEE J. Solid-State
Radio Freq. Integr. Circuits Conf. Dig. Tech. Papers, Jun. 2008, Circuits, vol. 53, no. 3, pp. 692–702, Mar. 2018.
pp. 533–536. [31] M. Ding et al., “A 95 μW 24 MHz digitally controlled crystal oscillator
[8] R. van Langevelde, M. van Elzakker, D. van Goor, H. Termeer, J. Moss, for IoT applications with 36 nJ start-up energy and >13× start-up
and A. J. Davie, “An ultra-low-power 868/915 MHz RF transceiver time reduction using a fully-autonomous dynamically-adjusted load,” in
for wireless sensor network applications,” in IEEE Radio Freq. Integr. IEEE ISSCC Dig. Tech. Papers, Feb. 2017, pp. 90–91.
Circuits Conf. Dig. Tech. Papers, Jun. 2009, pp. 113–116. [32] M. Ding et al., “Ultra-low-power clock generation for IoT radios,” in
[9] D. Griffith, J. Murdock, P. T. Røine, and T. Murphy, “A 37 μW dual- Proc. Adv. Analog Circuit Design Workshop, 2018.
mode crystal oscillator for single-crystal radios,” in IEEE ISSCC Dig. [33] D. Ruffieux et al., “A 3.2 × 1.5 × 0.8 mm3 240 nA 1.25-to-5.5 V
Tech. Papers, Feb. 2015, pp. 104–105. 32 kHz-DTCXO RTC module with an overall accuracy of μ1ppm and
[10] S. Iguchi, T. Sakurai, and M. Takamiya, “A low-power CMOS crystal an all-digital 0.1 ppm compensation-resolution scheme at 1 Hz,” in IEEE
oscillator using a stacked-amplifier architecture,” IEEE J. Solid-State ISSCC Dig. Tech. Papers, Jan./Feb. 2016, pp. 208–209.
Circuits, vol. 52, no. 11, pp. 3006–3017, Nov. 2017. [34] A. Paidimarri, D. Griffith, A. Wang, G. Burra, and A. P. Chandrakasan,
[11] A. Shrivastava, D. A. Kamakshi, and B. H. Calhoun, “A 1.5 nW, “An RC oscillator with comparator offset cancellation,”
32.768 kHz XTAL oscillator operational from a 0.3 V supply,” IEEE IEEE J. Solid-State Circuits, vol. 51, no. 8, pp. 1866–1877,
J. Solid-State Circuits, vol. 51, no. 3, pp. 689–696, Mar. 2016. Aug. 2016.
1392 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS–I: REGULAR PAPERS, VOL. 66, NO. 4, APRIL 2019

Ming Ding received the B.E. degree from the Christian Bachmann received the Ph.D. degree
Huazhong University of Science and Technology, in electrical engineering in 2011. He is the Pro-
China, in 2009, and the M.Sc. degree from the Eind- gram Manager of imec’s ULP Wireless Systems
hoven University of Technology, The Netherlands, Program. He joined imec in 2011, working on ultra-
in 2011. In 2011, he started as a Researcher at the low-power wireless communication systems, digi-
imec/Holst Centre, The Netherlands. Since then, he tal baseband processing, and hardware/software co-
has been working on data converter and ultra-low- design. In his previous work, he has covered various
power wireless research and design. wireless communication solutions for 802.11ah Wi-
Fi, Bluetooth LE, 802.15.4 (Zigbee), ultra-wideband
impulse radio, and others. Prior to joining imec,
he has been researching hardware-accelerated power
estimation for VLSI systems both with Infineon Technologies and Graz
University of Technology, Graz, Austria.

Kathleen Philips received the Ph.D. degree in elec-

trical engineering. He is the Program Director of
Yao-Hong Liu received the Ph.D. degree from imec for the IoT. He has 20 years of experience
National Taiwan University, Taiwan, in 2009. Since in the domain of low power mixed-signal, RF, and
2010, he joined imec, The Netherlands. His current integrated system design. She started her career
position is the Principal Membership of Technical at Philips Research, The Netherlands, working on
Staff, and he is leading the development of the analog and mixed-signal circuits. She joined the
ultra-low power RFIC design. His research focuses imec/Holst Centre in 2007, where she is currently
are energy-efficient wireless transceivers and RF leading a program on ultra-low power wireless,
sensing for IoT. He currently serves as a Technical processing and sensing systems. This is a part of
Program Committee of the IEEE ISSCC and RFIC open innovation collaboration, together with local
Symposium. and international industry. The research targets applications in the area of
infrastructure and person-centric IoT systems, with a focus on ultra-low power
communication and sensing.

Arthur Van Roermund was born in Delft, The

Netherlands, in 1951. He received the M.Sc. degree
in EE from the Delft University of Technology
in 1975 and the Ph.D. degree in applied sciences
Pieter Harpe (SM’15) received the M.Sc. and from K U Leuven, Belgium, in 1987. From 1975 to
Ph.D. degrees from the Eindhoven University of 1992, he was with Philips Research Laboratories,
Technology, The Netherlands, in 2004 and 2010, Eindhoven. From 1992 to 1999, he was a Full
respectively. In 2008, he started as a Researcher at Professor at the EE Department, Delft University
the imec/Holst Centre, The Netherlands. Since then, of Technology. In 1999, he joined the Eindhoven
he has been working on ultralow-power wireless University of Technology, where he was the Chair-
transceivers, with a main focus on ADC research and man of the Mixed-Signal Microelectronics Group
design. In 2011, he joined the Eindhoven University until 2017, and the Director of the Research of the TU/e-EE Department
of Technology, where he is currently an Associate from 2002 to 2012. He co-organized the yearly workshop on Advanced
Professor on low-power mixed-signal circuits. He is Analog Circuit Design from 2001 to 2012. Since 2017, he has been an
a co-organizer of the yearly workshop on advances Emeritus Professor. He was a member of three international assessment panels:
in analog circuit design and the Analog Subcommittee Chair for the ESSCIRC Politecnico di Milano, Italy, in 2007; Aalto University, Finland, in 2009; KTH,
Conference. He also served as an ISSCC ITPC Member and the IEEE SSCS Stockholm, Sweden, in 2012. He has published over 700 articles and 30 books.
Distinguished Lecturer and is a recipient of the ISSCC 2015 Distinguished He is a co-editor of another 12 books. In 2004, he achieved the Simon Stevin
Technical Paper Award. Meester Award for his scientific and technological achievements.