A 2.2GHz Sub-Sampling PLL with 0.16ps
rms
Jitter and
-125dBc/Hz In-band Phase Noise at 700μW Loop-Components Power
Xiang Gao, Eric Klumperink, Gerard Socci*, Mounir Bohsali*, and Bram Nauta
University of Twente, Enschede, The Netherlands; *National Semiconductor, Santa Clara, California
+31-53-489-3811, X.Gao@utwente.nl, B.Nauta@utwente.nl
Abstract
A divider-less PLL exploits a phase detector that directly
samples the VCO with a reference clock. No VCO sampling
buffer is used while dummy samplers keep the VCO spur
<-56dBc. A modified inverter with low short-circuit current
acts as a power efficient reference clock buffer. The 2.2GHz
PLL in 0.18μm CMOS achieves -125dBc/Hz in-band phase
noise with only 700μW loop-components power.
Introduction
Clock multiplication PLLs with very low jitter have recently
been proposed based on sub-sampling [1,2] and injection
locking [3,4]. In a PLL, the VCO dominates the out-of-band
phase noise while the loop-components dominate the in-band
phase noise. The sub-sampling (SS) PLL [1,2] can achieve
very low in-band phase noise because: 1) divider noise is
eliminated; 2) the phase detector (PD) and charge pump (CP)
noise is not multiplied by N
2
. This paper describes a new
SSPLL design aiming to drastically reduce the
loop-components power while maintaining its superior
in-band phase noise performance.
Proposed Low Power SSPLL
Fig. 1(a) shows the low power SSPLL architecture. A
sub-sampling phase detector (SSPD) samples the VCO with a
reference clock Ref and converts VCO phase error into
sampled voltage variation. A CP converts the sampled voltage
to current. A Pulser controls the CP gain and simplifies the
SSPD design to a track-and-hold [1]. A frequency locked loop
ensures correct frequency locking and is disabled after locking
to save power. In a SSPLL the PD and CP noise contributions
are low and thus their power can be scaled down progressively.
The VCO and Ref buffers for the SSPD then become the
bottlenecks for low power. In [1], they account for 30% and
60% of the total loop-components power, respectively. In this
design, we propose two techniques to alleviate these
bottlenecks: 1) direct sampling of the VCO without buffer
while keeping the disturbance to the VCO low; 2) power
efficient Ref buffering with drastically reduced short-circuit
current.
Fig. 2 shows the LC VCO and SSPD schematic. Different
from [1], no buffer is used between the VCO and SSPD
samplers. This saves power as buffers running at f
VCO
are
power consuming. The samplers use PMOS switches since the
VCO DC level is high. A concern of this buffer-less direct
VCO sampling is the disturbance to the VCO operation. When
Ref turns on/off the sampling switch, the VCO is
loaded/un-loaded by the sampling capacitors C
sam
. The VCO
load and thus f
VCO
is changed resulting in binary frequency
shift keying (BFSK), causing spurs at integer multiples of f
ref
.
In order to reduce this effect, dummy samplers are added as
(a)
Frequency locked loop
PFD/CP
with dead zone
÷ N
VCOVCO
X
O
VCO
Low Power
Buffer
CP
Pulser
Ref
N
∑∑
-
+
F
LF
(s)
K
VCO
/s
++ ++ ++
++
nref ,
φ
[Ф
out
]
nCP
i
, nLF
v
, nVCO,
φ
[Ф
ref
]
K
CP
++
nsam
v
,
K
SSPD
(b)
SSPD
v
sam
For differential sampling, K
SSPD
=2A
VCO
Fig. 1. Sub-Sampling PLL (a) architecture, (b) phase domain model.
C
sam
=10fF
Ref
Ref
V
tune
VCO+
VCO-
dummy sampler
C
sam
V
sam+
dummy sampler
10fF
10fF
sampler
sampler
3b Cap Array
C
sam
=10fF
9nH
V
sam-
dummy CP
Fig. 2. Schematic of the VCO and SSPD.
shown in Fig. 2, which are controlled by the inverted Ref. A
transmission gate (not shown in the figure) compensates the
inverter delay. Due to the complementary switching of the
sampler and its dummy, the VCO load does not change over
time and the BFSK effect is compensated. In reality, the
compensation is not perfect due to capacitor mismatch ΔC
sam
between the sampler and its dummy. Since ΔC
sam
scales with
the value of C
sam
, it is desirable to have a small C
sam
for a low
spur level. However, a smaller C
sam
means more sampler noise.
With the phase domain model in Fig. 1(b), the in-band phase
noise due to the samplers can be derived as
).log(
,
refVCOsam
SSPDbandin
fAC
kT
⋅⋅
=
−
2
2
10L
With f
ref
=55MHz and VCO amplitude A
VCO
=0.4V, C
sam
is
chosen to be 10fF resulting in -136dBc/Hz, 10% of the
targeted -126dBc/Hz of [1].
In order to properly sample the GHz VCO, Ref should have a
steep sampling edge with a slew rate (SR) higher than the VCO
SR. In most applications, Ref is derived from a sine wave
crystal oscillator (XO) which often has a much lower SR than
the VCO since f
ref
<< f
VCO
. A buffer converting the sine XO
into a square wave Ref is thus needed. In the 10s-of-MHz
frequency range, a CMOS inverter buffer is more power
efficient than a CML buffer as it mainly consumes dynamic
power. Noise on Ref is critical for in-band phase noise as it is
still multiplied by N
2
when transferred to the SSPLL output;
see Fig. 1(b). Thus large inverters need to be used at the
expense of power. As the input SR is low and output SR high,
power is wasted due to the “short-circuit” current caused by
simultaneous conduction of the NMOS and PMOS transistors
during switching.
In a sampling process, only one of the two clock edges is
used as the sampling edge. In this SSPD design (Fig. 2), the
Ref rising edge is the sampling edge. For low noise sampling,
the Ref sampling edge is highly critical and needs to be clean
while the other Ref edge is not relevant. Fig. 3 shows the
proposed Ref buffer, which exploits this property to drastically
reduce power. A similar circuit has been used in [2] to control
the Ref duty cycle. Here we exploit it to achieve low power.
The idea is to directly convey the critical edge and re-position
the other non-critical edge at a convenient place to avoid the
short-circuit current. The buffer core is an inverter with an
NMOS N1 and a PMOS P1. N1 is directly connected to XO as
in a conventional inverter, while a timing control circuit (TCC)
is inserted between P1 and XO. The TCC consists of two delay
cells Δt
1
and Δt
2
and a few standard logic gates. It generates a
narrow pulse V
GP
from the XO and controls the gate of P1. As
shown in Fig. 3, Δt
1
and Δt
2
are set such that the time when V
GP
is low (P1 conducts) and the time when XO is higher than the
threshold of N1 (N1 conducts) are non-overlapping. Since f
ref
is low, this timing plan is easy to achieve. In this way, N1 and
P1 will not conduct simultaneously thereby eliminating the
short-circuit current. Since the Ref rising edge is the critical
sampling edge, the size of N1 is kept big to maintain a low
sampling edge noise, while the TCC and P1 use small sizes to
save power as they only add noise to the non-critical edge. The
first block Inv1 in the TCC is a conventional inverter and has
the slow XO as its input. It thus still has short-circuit current,
but the contribution to the total buffer power is negligible as its
size is small. The proposed buffer thus greatly reduces power
while maintaining the critical edge’s noise performance.
Experimental results
The 2.2GHz PLL was fabricated in standard 1.8V 0.18-µm
CMOS with an active area of 0.4 x 0.5 mm
2
(Fig. 4). Measured
in-package with a 1.8V
p-p
55MHz XO as input, the in-band
phase noise
£
in-band
at 200kHz is -125dBc/Hz as shown in Fig.
4. The jitter integrated from 10kHz to 100MHz is 0.16ps
rms
.
The PLL loop-components consume 0.7mW and the VCO
1.8mW. The worst case reference spur measured from 20 chips
while changing Ref duty cycle is -56dBc. Fig. 5 summarizes
the PLL performance and benchmarks it to low jitter PLLs.
This design has the best PLL FOM. Note that we directly used
a 55MHz sine-wave XO as the PLL input while [3] used a
50MHz square wave and [4] used a 1GHz sine wave.
Compared with [1], the loop-components power is 8x lower
while
£
in-band
is only 1dB worse. Compared with [2], the
loop-components power is 3x lower while
£
in-band
is 4dB better.
References
[1] X. Gao, et al., “A 2.2GHz 7.6mW sub-sampling PLL with
-126dBc/Hz in-band phase noise and 0.15ps
rms
jitter in 0.18μm
CMOS,” ISSCC, pp. 392 - 393, Feb. 2009.
[2] X. Gao, et al., “Spur-Reduction Techniques for PLLs Using
Sub-Sampling Phase Detection,” ISSCC, pp. 474-475, Feb. 2010.
[3] B. Helal, et al., “A low jitter programmable clock multiplier based
on a pulse injection-locked oscillator with a highly-digital tuning
loop,” J. Solid-State Circuits, pp.1391–1400, May 2009.
[4] J. Lee and H. Wang, “Study of subharmonically injection-locked
PLLs,” J. Solid-State Circuits, pp.1539–1553, May 2009.
XO
V
GP
&
Inv1 toggle
V
th
of N1
V
GP
810/0.3
20/0.18
Δt
1
Δt
2
)
.
/
.
./.
(
18001
18041
N1 on
Δt
1
Timing Control Circuit (TCC)
N1
P1
Inv1
P1 on
Δt
2
Inv
out
XO
Inv
out
Ref
Fig. 3. Schematic and timing diagram of the low power buffer.
VCO
Loop
Filter
0.4
mm
0.5 mm
PD
CP
Fig. 4. Measured PLL output phase noise.
This Work
[1] [2] [3] [4-chip A] [4-chip B]
f
out
(GHz) 2.21 2.21 2.21
3.2 20 20
f
ref
(MHz)
55.25 55.25 55.25 50 1000 2500
RMS jitter
σ
t
(ps)
0.16
(10k-100M)
0.15
(10k-40M)
0.3
(10k-100M)
0.13
(100-40M)
0.11
(50k-80M)
0.048
(50k-80M)
In-band phase
noise (dBc/Hz)
-125
@200kHz
-126
@200kHz
-121
@200kHz
-127
@1MHz
-113
@1MHz
-123
@1MHz
Ref Spur (dBc)
(# of sample)
-56
(#=20)
-46
(#=1)
-80
(#=20)
-64
(#=1)
-46
(#=1)
-55
(#=1)
PLL Power P (mW)
2.5
7.6
3.8
28.6 38 105
Loop-Components
Power (mW)
0.7
5.8 2 - - -
PLL FOM (dB)
-252 -248 -244 -243
-243 -246
Active area (mm
2
) 0.20 0.18 0.20 0.40
<0.45 <0.32
Technology (CMOS)
0.18-μm 0.18-μm 0.18-μm 0.13-μm 90-nm 90-nm
This Work
[1] [2] [3] [4-chip A] [4-chip B]
f
out
(GHz) 2.21 2.21 2.21
3.2 20 20
f
ref
(MHz)
55.25 55.25 55.25 50 1000 2500
RMS jitter
σ
t
(ps)
0.16
(10k-100M)
0.15
(10k-40M)
0.3
(10k-100M)
0.13
(100-40M)
0.11
(50k-80M)
0.048
(50k-80M)
In-band phase
noise (dBc/Hz)
-125
@200kHz
-126
@200kHz
-121
@200kHz
-127
@1MHz
-113
@1MHz
-123
@1MHz
Ref Spur (dBc)
(# of sample)
-56
(#=20)
-46
(#=1)
-80
(#=20)
-64
(#=1)
-46
(#=1)
-55
(#=1)
PLL Power P (mW)
2.5
7.6
3.8
28.6 38 105
Loop-Components
Power (mW)
0.7
5.8 2 - - -
PLL FOM (dB)
-252 -248 -244 -243
-243 -246
Active area (mm
2
) 0.20 0.18 0.20 0.40
<0.45 <0.32
Technology (CMOS)
0.18-μm 0.18-μm 0.18-μm 0.13-μm 90-nm 90-nm
PLL Power (mW)
10
0
10
1
10
2
10
3
10
0
Jitter Variance (ps
2
)
10
-1
10
-2
10
-3
07_17.1
06_32.5
00_12.5
FOM
=
-
2
6
0
d
B
F
O
M
=
-
2
3
0
d
B
F
O
M
=
-
2
4
0
d
B
08_19.1
04_5.5
03_10.3
F
O
M
=
-
2
5
0
d
B
This work
[1]
[3]
[4-A]
[4-B]
]
1
)
1
log[(10
2
m
W
P
s
FOM
t
⋅=
σ
[2]
Fig. 5. Performance summary and comparison. The un-referenced
ones are the PLL designs with the best FOM in last 10 years’ ISSCC,
marked with “Year_PaperNumber”.