Microwatt power management: challenges of on-chip energy harvesting

IoT devices become more and more popular which implies a growing interest in easily maintainable and battery-independent power sources, as wires and batteries are unpractical in application scenarios where billions of devices get deployed. To keep the costs low and to achieve the smallest possible form factor, SoC implementations with integrated energy harvesting and power management units are a welcome innovation. On-chip energy harvesting solutions are typically only capable of supplying power in the order of microwatts. A significant design challenge exists for the functional blocks of the IoT-SoC as well as for the power management unit itself as the harvested voltage has to be converted to a higher and more usable voltage. Simultaneously, the power management blocks have to be as efficient as possible with the lowest possible quiescent currents. In this paper, we provide a look at on-chip microwatt power management. Starting with the energy-harvesting from RF power or light, we then show state-of-the-art implementations of ultra-low power voltage references and ultra-low power low-dropout regulator (LDO) designs.


Introduction
With a predicted number of more than nineteen billion short-range IoT devices in 2025 [3], there is also an upcoming challenge regarding the power supply of these devices. Many of them get deployed in areas without easily-accessible wired power-supplies, and a constant replacement of batteries becomes more and more unpractical and environmentally unfriendly with rising device counts. This trend is also visible in the road map of 6G, which also aims for zero-energy devices, e.g. for environmental monitoring, structural integrity monitoring, or pervasive industrial process monitoring [24]. Such batteryless devices also have the potential to save costs due to the lack of removable parts, which also enables smaller form factors and integration into non-accessible areas.
IoT devices cover a broad range of applications, implying also specialized silicon implementations for optimized performance in relatively low quantities, e.g. for specialized medical applications or environmental monitoring. At the same time, the goal is to aim for the lowest possible product costs which leads to the reasonable consequence that mature CMOS technology nodes (like 130 nm or 180 nm) should be chosen instead of the most advanced technologies to hold the NRE (non-recurring engineering) costs low for this many different specialized applications with low quantities.
To enable such zero-energy IoT devices, power-efficient designs of the functional blocks are required, but the real door openers are the energy-harvesting techniques and the accompanying powermanagement units to convert the harvested energy with the lowest possible losses into the required supply rails.
In Sect. 2.1 an overview of possible on-chip RF-energy harvesting options is provided. Following, in Sect. 2.2 on-chip photovoltaics options are shown, both with exemplary circuit implementations. Further, Sect. 3.2 will detail the design of an ultra-low-power low-Februar 2021 138. Jahrgang In Sect. 4 we present an exemplary biomedical SoC which implements an RF-powered EEG sensor with wireless data transmission using most of the presented techniques. Finally, Sect. 5 summarizes and concludes the paper.

Energy-harvesting
Depending on the application and location of the IoT device, there are several energy sources available that can be harvested, e.g. motion with piezoelectric generators, heat differentials with thermoelectric generators, light with photovoltaic cells, or RF energy with RF-energy harvester. The focus of the following two sections is on the latter two, as they are of interest for a very wide spectrum of different applications.

RF-energy-harvesting
There are two possibilities to harvest incidental electromagnetic (EM) waves in the MHz-up to the GHz-range. First, the environmentally available EM waves, where the transmitter of the potential energy source could be located several kilometers away from the energyharvesting system itself. In this case, the transmitter needs to have a very high RF output power in the range of kilo-watts, because of the strong decay of the EM far-field. In practice, this is best fulfilled by digital TV broadcast stations. Experimental tests show that it is possible to harvest 15 to 17 µW of power with the TV broadcast station 6.3 km away [23]. However, this test was carried out with a discrete RF charge-pump and a relatively large log-periodic antenna.
The other possibility is to use a dedicated local RF-power transmitter to energize several surrounding mobile nodes in a range of a few meters. The most practical frequencies are in the UHF ISM band (902 to 928 MHz) and the SRD band (863 to 870 MHz) because they give a good compromise between antenna size and path losses. These bands are available worldwide, and allow (at least) a maximum allowed transmitted power of 2 W ERP [4], which corresponds to a maximum available power at the receiving node of approx. 25 µW at a distance of 10 m [17].
For an on-chip RF-energy harvester, this means for both cases that the voltage level at the input is smaller than 100 mVpp (assuming 50 antenna impedance), which has to be rectified and multiplied to a higher voltage level at the same time. The design challenge of such RF-to-DC rectifiers/voltage multipliers (RVMs) is to achieve the highest possible power conversion efficiency (PCE). In general, there are RVM topologies that use diodes or diode-connected transistors in differential full-wave [5,26] or single-ended half-wave rectifiers [10,16,17]. Using Schottky diodes [11]-because of their lower turn-on voltage-is another possibility, but might not be available in a chosen technology. The main impact on the PCE are the threshold voltages of the used rectifying devices, which makes special compensating measures necessary. In [6] a static compensation scheme is shown which uses the higher voltage of the following stages for diode-connecting MOSFETs. While achieving good results, this method can't compensate the threshold voltage precisely. There also exist compensation techniques, which use auxiliary charge-pumps exactly subtracting the threshold voltage of the rectifying devices [15]. The more elaborate the compensation circuits get, the larger the parasitics become which degrade the PCE as well. Therefore, a promising approach is to use zero-V th MOS-FETs as rectifying device, keeping the circuit simple and rendering additional compensation measures unnecessary. A drawback is the large process sensitivity of this threshold voltage and therefore the risk of current back-flow due to the limited reverse isolation of the used MOSFETs, discharging the energy storage element. The work in [17] demonstrates how the concept of a simple zero-V th 5-stage Dickson charge-pump can be combined with an ultra-low-power diode (ULPD) [9] in the last stage to effectively prevent current from flowing from the output buffer capacitor back into the circuit, as shown in Fig. 1. As in most technologies only n-type zero-V th MOStransistors exist, the PMOS in the ULPD is built with standard-V th MOSFET, compensated with the principle of [15]. The implemented circuit is shown in Fig. 1 and is capable of generating 5 µW of dc power at 1.285 V output voltage from −12.3 dBm RF input power at a frequency of 868 MHz resulting in a PCE of 8.5% including matching losses. The associated plot of the dc output voltage versus the RF input power is shown in Fig. 2.

On-chip photovoltaic cell
Another interesting possibility to implement on-chip energy harvesting is the integration of a photovoltaic cell in a standard CMOS process. The idea is the same as with dedicated poly-silicon photovoltaic cells, namely a photo-diode with a relatively large area. Besides the theoretical solar-efficiency limit of 28.6% for crystalline sil-32 heft 1.2021 icon [14], in a standard CMOS process, more parameters (e.g. missing anti-reflective coatings and non-ideal doping concentrations) are sub-optimal for peak efficiencies. Additionally, the effective area of on-chip photovoltaic cells is comparatively small, and so the expectable power output under direct sunlight is in the order of microwatts.
In Fig. 3a the structure of a typical on-chip photovoltaic cell is shown. Theoretically, every p-n junction in a standard CMOS process can function as a photo-diode, but there are differences in efficiency. The absorption depth of light in silicon depends on the wavelength, which means that highly energetic ultraviolet and blue light gets absorbed near the surface, while lower energetic red and infrared light progresses into deeper regions. Theoretically, this would mean that shallow p-n junctions are able to harvest the most energy from the high energetic photons, but the crystal defects at the surface result in a higher recombination probability and therefore lost charge carriers. Therefore, the deeper p-n junctions tend to have higher efficiency, as has been shown in [21], and this also coincides with our own measurements. Figure 3b shows a photovoltaic cell fabricated in 180 nm CMOS, which implements the diode D 1 pictured in Fig. 3a between the p-substrate and a deep n-well. The cell consists of 25 unity cells with a total area of 0.267 mm 2 which is able to provide 2.78 µW of power at a voltage of −375 mV at direct sunlight on a sunny day (158 W/m 2 ) in its maximum power point, which results in an efficiency of 6.6% under this lighting conditions. Related works utilizing a comparable process in [8] and [1] report a maximum efficiency of 2.1% and 16.7% respectively, both measured with halogen light sources. The comparability is relatively difficult because of different spectral compositions of different light sources. Figure 4 shows the characteristic curve of the implemented photovoltaic cell under different levels of incident lighting. Thereby, the incident power of the sunlight was measured with a pyranometer. As the generated voltage is negative and too low for most on-chip applications, special charge-pump circuits are necessary, exemplary shown in the work of [21], as the diodes are not stackable, since the anode is tied to the common p-substrate.

Power management
Energy-harvesting solutions are not able to directly generate the needed voltage rails at exactly the time when the energy is required. The usual way is therefore to design the energy-harvesting system in a manner so that it generates a higher voltage than needed and stores it in an energy buffer. This buffer is-in most cases-too big for monolithic integration and has to be placed off-chip. Depending on the requirements, two buffer techniques are promising. Either a simple capacitor, which has the disadvantage of linear dependency between stored charge and voltage, can be used. Or, better yet, a rechargeable solid-state SMD battery (e.g. TDK CeraCharge™ [22]), available in small form factors and providing a capacity of 100 µA h while maintaining a battery-typical relatively constant voltage of ≈ 1.4 V, is a good fit for IoT applications. As the latter mentioned SMD battery has an internal resistance of ≈ 200 , appropriate decoupling capacitors have to be used to buffer high peak-currents, e.g. from radio-transmitters. For many applications decoupling in the order of 100 pF is sufficient, which is possible to integrate onchip.
The missing link in the power management is the energy-efficient regulation of the buffered, higher voltage to the constant application voltage. Consequently, an ultra-low-power, supply-voltageindependent reference voltage generator, and an ultra-low-power LDO or DC/DC-converter is needed. The latter one has as advantage the capability to use all the energy stored in the buffer-capacitor, while a LDO simply converts the excess voltage into heat. Drawback of DC/DC converters working at such low power levels is their poor conversion efficiency. To reach efficiencies > 60%, comparably complex and area consuming designs are necessary [7]. Additionally, the advantage of DC/DC-converter vanishes, when the voltage on the buffer-capacitor is near the regulated voltage, exactly the region, where highest efficiency is the most important. Therefore the following focus is on simple LDO designs.

Low-power voltage-reference design
The requirements of voltage-reference circuits for ultra-low-power IoT SoCs are quite challenging, as for many analog sensing applications a precise and stable voltage reference is inevitable. Most critical is the minimization of the consumed power, but also the line-regulation capabilities are very important, as the supply voltage from the energy buffer can vary substantially. Low noise and good temperature-stabilization are also desirable properties, but here compromises can be accepted depending on the specific use case, as, for instance, temperature compensation might not be as In [20] an ingenious circuit, named 2-Transistor (2T) voltage reference, has been shown, which has a 100× to 1000× better energy efficiency than other nanowatt voltage references, while offering a better line-sensitivity and temperature coefficient in less area, without the need of additional circuitry including operational amplifiers, biasing-or startup-circuits. Figure 5 shows an implementation of this 2T voltage reference circuit in 180 nm CMOS technology. M 1 /M 2 and M 3 /M 4 are series-connected maximum-length transistors, respectively. The functional principle is based on subthreshold-biased transistors with distinct V th levels, in this case, a standard-V th thickoxide device (M 3 /M 4 ) and a native-V th thick-oxide device (M 1 /M 2 ). C 1 is used for improving the noise performance. The implemented design uses 20 branches as shown in Fig. 5 in parallel to further improve the noise performance, with a total area of 0.027 mm 2 and consumes only 311 pA at a nominal voltage of 1 V, while the circuit is operable down to 0.4 V, determined by simulation as the voltage reference is not directly measurable, but only in combination with the LDO. The generated voltage is approx. 278 mV with a nominal line sensitivity of 28.4 µV/V, a temperature sensitivity of 1.67 µV/ • C and an integral noise of 7.1 µV rms in 4096 Hz bandwidth.
An important note for this particular type of reference structure is to connect the generated reference voltage only to gates of thick-oxide devices, as the gate-leakage current of thin-oxide devices could disturb the reference because of its low quiescent current.

Low-power LDO design
Like the reference voltage generation circuits, also the voltage regulator circuits should be as power-efficient as possible for micro-watt energy harvesting SoCs. Interesting implementations are for example shown in [12,13,25]. The key for adequate performance with very-low power-drain is (again) simple circuit design and the avoidance of as much periphery as possible. Figure 6 shows an LDO circuit, implemented in 180 nm CMOS, consisting of a simple error amplifier in a symmetrical OTA topology, utilizing a 8.7 M resistor for biasing and therefore avoiding separate bias current generation as there is no stable supply voltage available. Additionally, this biasing method offers a wide supply voltage range. At a nominal supply voltage of 1.2 V and a regulated output voltage of 1 V, the simulated current consumption of the error amplifier is 214 nA. Although the average power consumption of the application is very low, it surges high peak-currents, which implies requirements on the dynamic settling behavior and therefore requiring a certain gain and therefore current. The pass-transistor is realized by a native-V th NMOS transistor, which is the best choice for this application case, as the NMOS offers a more stable behavior through the intrinsic load regulation of the pass-device itself. As an NMOS pass-device needs at least a drop-out voltage of V th , a native type (with a threshold voltage close to zero) is chosen to ensure very low required drop-out voltages. The feedback resistors R 2 and R 3 are built out of 84 × 54 k resistors in series where the bottom 20× resistors are digitally programmable with thermometer-coded shunt transistors, enabling an output voltage range of 0.74 to 1.42 V. Figure 7 shows the measured output voltage versus the input (supply) voltage with a constant 10 M load at different temperatures and an output voltage versus temperature plot. The determined temperature coefficient is −522 µV/ • C, equaling −502 ppm/ • C. Compared to the value of [20] between 54 ppm/ • C and 176 ppm/ • C for their 180 nm 2T-reference our result is a bit worse, but it also contains the influence of the LDO. The line regulation is 7 mV/V, and the load regulation is 0.68 mV/µA.

Application example
As an exemplary application of the presented techniques an RFpowered, wireless EEG electrode is briefly discussed. This highlyintegrated sensor SoC as shown in Fig. 8 and Fig. 9 is powered by the RF-harvester described in Sect. 2.1. The harvested energy is stored in an off-chip 10 µF-sized capacitor and further regulated to 1 V using the LDO presented in Sect. 3.2. Instead of a fullblown bandgap circuit the compact 2T reference voltage generator in Fig. 3.1 is used to stabilize the supply voltage. In this SoC a complete EEG signal acquisition system is realized, consisting of a pre-amplifier [19] and an ultra-low-power SAR ADC [18]. After decimation, the ADC data is protected by forward error correction and sent via an ultra-low-power impulse-radio ultra-wideband (IR-UWB) transmitter [17] with a data rate of 5.12 kbps using a double pulse interval coded alphabet. We could obtain an indoor range of ≈ 12 m with a custom built base-station. The whole system is intended to continuously operate without a battery, and fits into a power envelope of < 5 µW, enabling a buffer-time of ≈ 1.4 s assuming an usable buffer voltage overhead of 0.7 V. The shown SoC is demonstrating the potential of well-designed micro-watt power management.

Conclusion
This paper gives an overview of the challenges of on-chip power harvesting and power management and presents possible solutions. State-of-the-art examples in the different categories are given, complemented with own implementations. Finally, the interaction of the individual implemented blocks is shown exemplary with a wireless EEG electrode SoC.