1. Introduction
The development of low-power applications, such as the Internet of Things [
1], Energy Harvest [
2], Intra-Body Communication systems [
3], and the Wireless Sensor Network (WSN) [
4,
5], has spurred the research on low-power design. Among these applications, WSN systems and various potential applications, such as health monitoring, location, and monitoring of hazardous areas [
6], have received increasing attention. In order to reduce the area of sensor nodes, WSNs require stringent limits on the size and weight of the battery. Therefore, low-power design is essential for saving the battery size and prolonging the battery life. Despite the recent advancements, the WSN system lifetime is still limited by the large power consumption of its radio, especially the phase-locked loop (PLL) that performs as a local oscillator and provides high-frequency accuracy and low phase noise. In recent years, all-digital PLLs (ADPLLs) [
7,
8,
9] have been preferred over their analog counterparts, i.e., charge pump PLLs, owing to their better flexibility, smaller area, and lower power consumption.
As a key sub-block of an ADPLL, a digitally controlled oscillator (DCO) consumes the most power to generate local frequency. To evaluate the overall performance of the DCOs, a figure of merit (FoM) [
10], which includes the phase noise, power consumption, and carrier frequency, is used.
where PN is the phase noise, f
0 is the carrier frequency, and Δf is the frequency offset from the carrier. Furthermore, a DCO dominates the out-band phase noise of an ADPLL. The phase noise spectrum at the ADPLL RF output due to the DCO quantization effect is
where f
LSB is the resolution of DCO, Δf is the frequency offset from the carrier, and f
ref is the reference frequency of ADPLL. The resolution of a DCO dominates the phase noise performance. To satisfy the demands of systems, DCOs have been widely researched over the past decade. Ring oscillators [
11,
12,
13] have a simple structure and wide tuning range, but their noise performance is weaker than that of LC oscillators. To reach the performance of the counterparts in CPPLLs, i.e., the voltage-controlled oscillators (VCOs), the oscillator in [
14] used a nine-bit digital-to-analog converter (DAC) to convert the digital frequency control words (FCW) to analog signals that were fed to a VCO. Although it had a very high resolution, the added DAC multiplied the burdens of power consumption and area. Some works [
15,
16,
17,
18] have been based on metal oxide semiconductor (MOS) varactors with a small capacitance to improve the resolution and lower the power. However, unit variable capacitance is not small enough and it is very sensitive to the parasitic effect when scaled to an aF level. Therefore, ΔΣ modulators are applied to MOS varactors to further improve the time-average resolution through high-frequency dithering [
19,
20]. However, the ΔΣ modulators always work at a high frequency to achieve oversampling, which results in the power rising and phase noise deterioration due to ΔΣ quantization. Class-F DCOs [
21,
22] based on a transformer-feedback structure can provide passive voltage gains to adapt to low-voltage operation. However, the gates are separated from drains via the transformers, which results in a very low-frequency pushing.
According to
[
23], reducing the supply voltage is one the most efficient methods for realizing low-power implementations. This paper presents a low-power LC-DCO at near-threshold supplies (0.5~0.7 V) conforming to the Zigbee standard. As a sub-block of an ADPLL, the proposed DCO was used to generate the local frequency for a quadrature receiver for WSN applications, as shown in
Figure 1. In this quadrature receiver, it was not necessary to use a quadrature local oscillator that generates a double carrier frequency. The power consumption was thereby reduced [
24]. The LC-DCO based on a cross-coupled structure achieved a more superior phase noise performance than ring oscillators. The cross-coupled pair used the current-reuse technique to lower the power. A multi-stage capacitance shrinking (MACK) technique was proposed to form the LC tank, which largely improved the resolution of the ΔΣ-less DCO without increasing the power consumption of the chip. The rest of this paper is organized as follows.
Section 2 analyzes the model of the proposed LC tank based on the MACK technique and describes the implementation details. In
Section 3, the experimental results of the prototype and the relevant discussions are demonstrated, followed by the conclusions in
Section 4. Finally, the detailed theoretical analysis of models and derivation of equations in
Section 2 are explained in
Appendix A.
3. Experimental Results and Discussions
The proposed DCO was designed in the Cadence IC software package and was fabricated using a 130-nm 1P8M CMOS technology. As shown in the
Figure 9, the chip occupied 0.9 mm
2, including a buffer circuit occupying 0.12 mm
2. The shape of the chip was too narrow and the ratio of the length and width was up to 5.7. The chip must accommodate in order to contain many other circuits in the whole block. It affects the DCO performance more or less due to the difficult routing and signal attenuation. It will be improved in the further work.
The proposed DCO core worked at 0.5~0.7-V supplies, while the buffer circuit worked at a 1.2-V supply. The DCO circuit consumed 0.8 mA, 1.1 mA, and 1.7 mA at a supply of 0.5 V, 0.6 V, and 0.7 V, respectively. A 7-mW power was consumed by the buffer circuit used for measurements only.
In this work, the total power was mainly consumed by the output buffer and the cross-coupled pair. The unit MOS varactor was composed of two PMOS pairs that were inversely connected in parallel, as shown in
Figure 6. Obviously, the varactor banks scarcely contributed to the power consumption. In order to apply this approach to different applications, a higher- (or lower-) stage MACK-based tank can be easily formed by cascading more (or less) varactor banks to achieve a higher (or lower) resolution, which scarcely increases the power consumption of the chip. However, increasing the stage number will result in a larger chip area and higher production cost.
The frequency range of the DCO is shown in
Figure 10a. The DCO had the widest tuning range of 130 MHz when it worked at a 0.5-V supply, ranging from 2.382 GHz to 2.512GHz. The tuning ranges were from 2.385 GHz to 2.5GHz and from 2.385 GHz to 2.494 GHz at a supply of 0.6 V and 0.7 V, respectively.
It is visible that the tuning range narrowed gradually with the rising of the supply. Additionally, the lowest frequencies in the three cases were almost the same, while the highest frequency was inversely proportional to the voltage supply used. The main reason for this is that the three cases had the same control voltage (0 V) when FCW = 0, while the control voltages were different (0.5~0.7 V) when FCW = 1. As shown in
Figure 7, the C-V curve (the red line in
Figure 7) was not flat in the high-voltage region, where the capacitance became larger as the control voltage rose. Therefore, the capacitance was smallest at a 0.5-V supply (FCW = 0.5 V), while the capacitance was largest at a 0.7-V supply (FCW = 0.7 V). On the other hand, the capacitances were absolutely the same at different supplies when FCW = 0 due to the same control voltage (0 V). Therefore, the frequency range was widest at a 0.5-V supply, while it was narrowest at a 0.7-V supply. All of the three cases covered the frequency range of the Zigbee requirement (2.4 GHz~2.4835 GHz).
Figure 10b shows the results for FCW [5:0], which changed from 0 to 63 at a 0.6-V supply. The average resolution at 2.4 GHz was 7.3 kHz. The step size increased as the frequency rose. The overall average resolution was about 8 kHz. The variations of resolution mainly resulted from the mismatch of varactors. There are two reasons for the mismatch. First, the variations of voltage supply caused the variations of ΔC
int through the C-V, characteristic of the MOS varactor. Second, the parasitic capacitance degraded the accuracy of the unit variable capacitance, and the linearity of the resolution was thereby deteriorated.
According to Equation (6), the unit variable capacitance ΔC
fra was not a constant when the DCO worked at different frequencies. Therefore, the frequency resolution of the DCO was also a variable value. A lower frequency leads to a smaller ΔC
fra and a higher-frequency resolution, which may influence the operation of ADPLL. On power-up, only the auto-frequency calibration (AFC), DCO, and 1/8 divider were active. AFC was used to detect the frequency difference between
f0 and the target frequency and changed FCW [17:6] accordingly. FCW [5:0] = 32 and remained constant. The fluctuation of ΔC
fra did not disturb frequency locking because the frequency locking did not depend on ΔC
fra and ADPLL was a robust negative feedback loop in this step. Once AFC finished frequency locking, it froze FCW [17:6] and activated the time-to-digital converter (TDC), digital loop filter (DLF), and Fractional-N divider. TDC and DLF drove the third-stage bank of DCO, i.e., changed FCW [5:0] to finish phase locking. Because the frequency resolution (or DCO gain) affected the loop gain, bandwidth, and stability of ADPLL, these loop parameters thereby fluctuated when DCO worked at different frequencies. The effect on frequency locking was negligible, but was remarkable on phase locking, especially for the ADPLL with wide tuning range. In order to avoid fluctuations of the loop performance or even fail locking, parameters should be carefully chosen based on the DCO output frequency. A good choice is using a programmable digital loop filter to compensate for the fluctuation of the frequency resolution. When phase locking was finished, FCW [17:6] remained constant, and only FCW [5:0] was variable. Therefore, the range of ΔC
fra was limited, although ΔC
fra was a variable value according to the analysis in
Section 2. For example, when the ADPLL worked at 2.4 GHz, FCW [17:12] = 7 and FCW [11:6] = 32. Therefore, C
0 = 1162 fF and C
1 = 1262 fF. Assuming that FCW [5:0] had the largest change range, i.e., from 0 to 63, C
2 accordingly changed from 1134 fF to 1386 fF (note that the minimum capacitance of the bank is not zero). The calculating ΔC
fra changed from 5.1 aF to 6.3 aF, and the range was only 1.2 aF. In fact, the actual change range of ΔC
fra was smaller because FCW [5:0] did not change from 0 to 63 when the ADPLL was locked. Therefore, the fluctuation of the frequency resolution was negligible after the ADPLL was locked.
The low- and high-frequency output spectrums are depicted in
Figure 11. The output power was −1.93 dBm at 2.4 GHz while it was −1.40 dBm at 2.5 GHz. This translates into a 0.25-V output at 2.4 GHz and a 0.27-V output at 2.5 GHz. The measurements are exhibited at a 0.6-V supply.
As stated above, the resolution was about 7.3 kHz at a carrier of 2.4 GHz. The resulting phase noise was about −131 dBc/Hz at 1-MHz offset when a 50-MHz reference is used, according to Equation (2). When the DCO works at 2.5 GHz, substituting a 9 kHz resolution into Equation (2), we can obtain
−129 dBc/Hz at 1-MHz offset. The measured phase-noise plots of the proposed DCO operating at a low- and high-frequency are shown in
Figure 12. As shown in
Figure 12a,b (at a 0.6-V supply), at a 2.4-GHz carrier, the measured results were −57.0974 dBc/Hz at 10 kHz offset, −81.11 dBc/Hz at 100 kHz offset, −126.27 dBc/Hz at 1-MHz offset, and −135.9425 dBc/Hz at 3-MHz offset. At a 2.5-GHz carrier, the measured results were −52.4966 dBc/Hz at 10 kHz offset, −77.1384 dBc/Hz at 100 kHz offset, −125.9480 dBc/Hz at 1-MHz offset, and −136.6574 dBc/Hz at 3-MHz offset. The phase noise at a 2.5-GHz carrier was slightly weaker than that at 2.4 GHz, except at 3-MHz offset. As shown in
Figure 12c,d (at a 0.5-V supply), at a 2.4-GHz carrier, the measured results were −52.1115 dBc/Hz at 10 kHz offset, −76.0526 dBc/Hz at 100 kHz offset, −122.2067 dBc/Hz at 1-MHz offset, and −130.8525 dBc/Hz at 3-MHz offset. At a 2.5-GHz carrier, the measured results were −46.5848 dBc/Hz at 10 kHz offset, −71.0387 dBc/Hz at 100 kHz offset, −120.7571 dBc/Hz at 1-MHz offset, and −130.0356 dBc/Hz at 3-MHz offset. Compared with the phase noise at a 0.6-V supply, the phase noise at a 0.5-V supply was deteriorated by an average of 5 dB due to the lower current and lower signal amplitude.
The scatter diagram shown in
Figure 13 summarizes the phase noise performance at a 1-MHz offset for the whole tuning range at the supplies of 0.5~0.7 V. The measured results ranged from −122.3 dBc/Hz to −120.7 dBc/Hz, from −126.7 dBc/Hz to −125.95 dBc/Hz and from −128.2 dBc/Hz to −127.5 dBc/Hz at a supply of 0.5 V, 0.6 V, and 0.7 V, respectively. The DCO had the best phase noise performance when VDD = 0.7 V at the cost of the most power, which approximately doubled the consumption when VDD = 0.6 V. Although the power consumption was only 0.4 mW at a 0.5-V supply, the phase noise in this case was greatly deteriorated and the FoM was thereby the worst among the three cases (see below). However, it provides a definite tendency for our future work, in which we will attempt to design circuits at a 0.5-V or even subthreshold supply to further reduce the power on the premise of good phase noise.
Table 2 summarizes the performance comparison among similar works. Reference [
2] had the lowest supply (0.4 V), widest frequency range (800 MHz), and the highest resolution (1.3 kHz). Reference [
8] had the lowest power (only 0.26 mW) and achieved a wider frequency range (470 MHz) than this work, so it achieved the best FoM
T, up to 199 dB, among the works in
Table 2.
When VDD = 0.6 V, the proposed DCO achieved a phase noise of -126.27 dBc/Hz at a 1-MHz offset at a carrier of 2.4 GHz. The power consumption was 0.66 mW. The calculated FoM was 195.68 dB. The calculated FoM was 193.78dB when VDD = 0.5 V and it was 194.8 dB when VDD = 0.7 V. Among the three supplies (0.5, 0.6, and 0.7 V), the best result was obtained at a 0.6-V supply, which is shown in
Table 2.
However, the DCO based on the proposed method improved the FoM at the cost of a narrow frequency range. As shown in
Table 2, the frequency range of the proposed DCO was the narrowest (115 MHz) among the other state-of-the-art works (200~800 MHz), which limited the applications of the prototype to a narrowband system. The FoM
T was only 189.12dB when VDD = 0.6 V. The frequency range can be widened by increasing more FCWs of every varactor bank. However, the penalty is a larger chip area and higher production cost, although it will not result in higher power in this work. Therefore, a tradeoff between the frequency range and area exists in this work. In addition, properly reducing the frequency overlap between adjacent bands on the premise of no gap is also an option, although it will weaken the tolerance against PVT variations.
4. Conclusions
We proposed a MACK technique for low-voltage and low-power DCOs. The model of the MACK was analyzed based on the analysis of a one-stage model. The unit variable capacitance of the LC tank was largely reduced by the serial bridging capacitor and the increased number of stages (exponential modulation). A particular case and preconditions of each model were also provided to guide the implementation.
The proposed LC-DCO core was designed in Cadence. A current-reuse structure at a near-threshold supply was used for reducing the power. The LC tank based on the MACK technique contained three six-bit varactor banks with the same structure. A MOS varactor, comprising two PMOS pairs that were inversely connected in parallel, was used as the unit of the three banks to achieve a 4-fF ΔCint. All the design parameters and the design flow were provided. A 4.7-aF~7 aF ΔCfra and an 8-kHz resolution were achieved in this work.
Based on the analysis and design, an LC-DCO based on MACK was fabricated in a 130-nm CMOS technology with multiple supplies (0.5~0.7 V for the core, 1.2 V for the buffer). The DCO core consumed 0.4, 0.66, and 1.2 mW at a supply of 0.5, 0.6, and 0.7 V, respectively. The frequency range was from 2.382 GHz to 2.512 GHz at a 0.5-V supply, and the highest frequency was reduced gradually with the rising of the supply. The measured results of phase noise at a 0.6-V supply were −126.27 dBc/Hz at 1 MHz and −125.9480 dBc/Hz at 1 MHz at the carriers of 2.4 GHz and 2.5 GHz, respectively. The best FoM of 195.68 was obtained when VDD = 0.6 V. Thanks to the current-reuse technique, the DCO achieved almost the same noise performance as the traditional cross-coupled structure that theoretically consumes double the current. Furthermore, thanks to the proposed MACK technique, a high resolution was achieved without any additional power consumption. Therefore, the DCO achieved good phase noise and low power simultaneously.
In order to apply this system to different applications, a higher- (or lower-) stage MACK-based tank can be easily formed by cascading more (or less) varactor banks to achieve a higher (or lower) resolution, which scarcely increases the power consumption of the chip. Additionally, the MACK technique can be used not only in LC DCOs, but also in ring oscillators at the cost of the area. The drawback of this work is the narrow frequency range (115 MHz), which limits the applications of the prototype. In future work, the frequency range can be improved by increasing the bits of FCW at the cost of the area or by increasing ΔCint at the cost of a worse resolution and phase noise.