A 0.0014 mm2 150 nW CMOS Temperature Sensor with Nonlinearity Characterization and Calibration for the −60 to +40 °C Measurement Range

This work presents a complementary metal–oxide–semiconductor (CMOS) ultra-low power temperature sensor chip for cold chain applications with temperatures down to −60 °C. The sensor chip is composed of a temperature-to-current converter to generate a current proportional to the absolute temperature (PTAT), a current controlled oscillator to convert the current to a frequency signal, and a counter as the frequency-to-digital converter. Unlike the conventional linear error calibration method, the nonlinear error of the PTAT current under the low temperature range is fully characterized based on the device model files provided by the foundry. Simulation has been performed, which clearly shows the nonlinear model is much more accurate than the linear model. A nonlinear error calibration method, which requires only two-point calibration, is then proposed. The temperature sensor chip has been designed and fabricated in a 0.13 μm CMOS process, with a total active die area of 0.0014 mm2. The sensor only draws a 140 nA current from a 1.1 V supply, with the key transistors working in the deep subthreshold region. Measurement results show that the proposed nonlinear calibration can decrease the measurement error from −0.9 to +1.1 °C for the measurement range of −60 to +40 °C, in comparison with the error of −1.8 to +5.3 °C using the conventional linear error calibration.


Introduction
Temperature monitoring is mandatory in cold chain applications for the production, storage, distribution, and transportation of perishable, but life-critical products, such as foods, blood products, and vaccines [1][2][3][4]. The temperature range for the food cold chain is commonly from −20 to +15 • C [1, 2]. For blood products, such as fresh frozen plasma, the cold chain needs to maintain a low temperature under −25 • C. Temperature control for the vaccine cold chain is even more stringent [3], and some special vaccines, such as the anti-Ebola vaccine, may need a low temperature of −60 • C [4].
The fundamental requirement of temperature monitoring in cold chains is a high accuracy. Low power consumption is another key requirement, since in many cold chain applications, the temperature monitoring function is incorporated into the extremely power-constrained wireless telemetry circuit powered by a miniature battery or a radio frequency identification (RFID) tag based on wireless power transfer [5]. In addition, cold chain applications need temperature monitoring for every truck and even every small product package, and such applications are usually cost sensitive [5][6][7][8]. In general, there is a great need to investigate low-cost, low-power, wide-range, and high-accuracy temperature sensors for cold chain applications [9]. The target of this work is to develop a technology to design such a temperature sensor using the complementary metal-oxide-semiconductor (CMOS) technology.

Temperature Sensor Circuit Architecture
The block diagram of the presented temperature sensor is shown in Figure 1. It includes three main functional blocks. The temperature-to-current converter generates a temperature-dependent current. Ideally, the output current is a PTAT current. A current controlled oscillator then converts the current to a frequency signal. The oscillation frequency is proportional to the PTAT current, and thus proportional to the absolute temperature. A counter serves as the frequency to digital converter, which digitizes the oscillator output frequency. In this work, the designed sensor is one function part of a system-on-a-chip (SoC), and it is powered by an on-chip 1.1 V low-dropout (LDO) regulator in the SoC or an external power supply. The design is optimized for the 1.1 V power supply.
The temperature-to-current converter is the key of this sensor. Figure 2 shows the circuit implementation of the temperature-to-current converter. M1 and M2 are working in the subthreshold region. The drain-source currents of M1 and M2 have the following relationship [20,21]: where µ is the mobility, C ox is the oxide capacitance, W/L is the transistor aspect ratio, V th1 and V th2 are the threshold voltages of M1 and M2, respectively, and n is the subthreshold gate coupling coefficient. V T = k T/q, in which k is the Boltzmann constant, T is the absolute temperature, and q is the electron charge. As shown in Figure 2, V 1 is the voltage across the poly resistor, R0, and V 2 is the gate voltage of both M1 and M2. In this design, the aspect ratio of M1 is 2 times that of M2.
where μ is the mobility, Cox is the oxide capacitance, W/L is the transistor aspect ratio, Vth1 and Vth2 are the threshold voltages of M1 and M2, respectively, and n is the subthreshold gate coupling coefficient. VT = k T/q , in which k is the Boltzmann constant, T is the absolute temperature, and q is the electron charge. As shown in Figure 2, V1 is the voltage across the poly resistor, R0, and V2 is the gate voltage of both M1 and M2. In this design, the aspect ratio of M1 is 2 times that of M2.  The current mirror formed by M3 and M4 is carefully matched. Ideally, M1 and M2 have the same current, which is denoted as I0: Substituting (1) and (2) into (3) leads to: where μ is the mobility, Cox is the oxide capacitance, W/L is the transistor aspect ratio, Vth1 and Vth2 are the threshold voltages of M1 and M2, respectively, and n is the subthreshold gate coupling coefficient. VT = k T/q , in which k is the Boltzmann constant, T is the absolute temperature, and q is the electron charge. As shown in Figure 2, V1 is the voltage across the poly resistor, R0, and V2 is the gate voltage of both M1 and M2. In this design, the aspect ratio of M1 is 2 times that of M2.  The current mirror formed by M3 and M4 is carefully matched. Ideally, M1 and M2 have the same current, which is denoted as I0: Substituting (1) and (2) into (3) leads to: The current mirror formed by M3 and M4 is carefully matched. Ideally, M1 and M2 have the same current, which is denoted as I 0 : Substituting (1) and (2) into (3) leads to: If V th1 = V th2 , then: If the resistance of R0 is constant, then the drain-source current of M1 is given by: Ideally, I 0 is a PTAT current [22][23][24], which is the basis of this type of temperature sensor [22]. A similar PTAT implementation in [24] showed the measured PTAT behavior.

Nonlinearity Characterization for a Wide Measurement Range
The validation of approximating I 0 to a PTAT current relies on the assumptions that M1 and M2 have the same threshold voltages, and the resistance of R0 is not temperature dependent. These assumptions are approximately true for a modest temperature range [24]. However, when extending the measurement range to very low temperature, i.e., −60 • C, such an approximation actually introduces quite large nonlinearity. In this section, the nonlinearity of the temperature-to-current converter for a wide temperature range will be characterized.
Firstly, the transistor threshold voltage is temperature dependent. Based on the BSIM model [25][26][27][28], the threshold voltage of a MOS transistor is written as: where K T1 , K tl1 , and K T2 are process-dependent parameters, L eff is the effective channel length, and V bseff is the body-source voltage. Note that M1 and M2 have different source voltages, and the body-source voltage difference is just −V 1 . Taking the body effect into consideration [25,26,29,30], there exists a small difference between the threshold voltages of M1 and M2, which is given as: Substituting (8) into (4) leads to: in which K T2 = −K T2 /T 0 . On the other hand, the temperature dependent resistance, R 0 (T), at the temperature, T, is given by: where R 0 is the resistance at the reference temperature, T 0 , and TC 1 and TC 2 are the temperature coefficients. Based on the foundry design kit (FDK), TC 1 is on the order of 10 −5 , and TC 2 is on the order of 10 −7 . The current, I 0 , through R 0 is derived as: In the concerned temperature range of −60 to +40 • C, (K T2 + TC 1 ) · (T − T 0 ) is much larger than (TC 2 + K T2 · TC 1 ) · (T − T 0 ) 2 and K T2 · TC 2 · (T − T 0 ) 3 . Equation (11) can be simplified by ignoring the high order terms in the denominator: in which K TC = K T2 + TC 1 is a process dependent constant. The calculation based on the CMOS process technology file provided by the foundry indicates that K TC = 7.10 × 10 −5 if T 0 = 300 K. In the coefficient, K TC , the contribution of K T2 due to the transistor threshold voltage temperature dependence and that of TC 1 due to the resistor temperature dependence is about 70% and 30%, respectively.
It can be concluded that with the temperature dependence of the transistor threshold voltage and the resistance temperature dependence, the drain-source current of M1 cannot simply be treated as a PTAT current. Equation (12) indicates that the current-temperature curve has a hyperbolic shape. Though the FDK available only provides device models down to −40 • C, Equation (12) works for much lower temperatures since the BSIM model is valid for quite a wide temperature range [25].
Equation (12) also indicates that if the measurement range is small, then the measured temperature is close to the reference temperature, T 0 ; K TC · (T − T 0 ) is small; and the denominator, 1 + K TC · (T − T 0 ), degenerates to 1. In that case, Equation (12) degenerates to a linear relationship between I 0 and T, resulting in a PTAT current as expected. However, for the large measurement range targeted in this work, K TC · (T − T 0 ) is not negligible, and the linear approximation will introduce quite large error, which requires calibration. Figure 3 shows the simulated current, I 0 , for the temperature range of −40 • C to 0 • C (the lower simulation boundary is limited by the FDK), in contrast to the conventional linear approximation and the hyperbolic prediction using (12). The black line is the simulated current, while the blue dashed line is a straight line by connecting two points (−20 • C and 0 • C) on the black line, and the red dashed line is the hyperbolic fitting of these two points. The difference between the simulated current and the straight line reaches 0.07 nA for the temperature of −40 • C, which means that the measurement error will be about 1.2 • C for the −40 • C point with conventional linear approximation. In contrast, the difference between the simulated current and the nonlinear prediction is only 0.02 nA at −40 • C, which corresponds to only a 0.34 • C measurement error. Obviously, the simulated current, I 0 , is closer to the hyperbolic line as predicted by (12) rather than the straight line. It can also be roughly calculated from Figure 3 that the measurement error using conventional linear approximation will reach 2.8 • C when the temperature is down to −60 • C. threshold voltage temperature dependence and that of TC1 due to the resistor temperature dependence is about 70% and 30%, respectively. It can be concluded that with the temperature dependence of the transistor threshold voltage and the resistance temperature dependence, the drain-source current of M1 cannot simply be treated as a PTAT current. Equation (12) indicates that the current-temperature curve has a hyperbolic shape. Though the FDK available only provides device models down to −40 °C, Equation (12) works for much lower temperatures since the BSIM model is valid for quite a wide temperature range [25].
Equation (12) also indicates that if the measurement range is small, then the measured temperature is close to the reference temperature, T0; KTC • (T − T0) is small; and the denominator, 1 + KTC • (T − T0), degenerates to 1. In that case, Equation (12) degenerates to a linear relationship between I0 and T, resulting in a PTAT current as expected. However, for the large measurement range targeted in this work, KTC • (T − T0) is not negligible, and the linear approximation will introduce quite large error, which requires calibration. Figure 3 shows the simulated current, I0, for the temperature range of −40 °C to 0 °C (the lower simulation boundary is limited by the FDK), in contrast to the conventional linear approximation and the hyperbolic prediction using (12). The black line is the simulated current, while the blue dashed line is a straight line by connecting two points (−20 °C and 0 °C) on the black line, and the red dashed line is the hyperbolic fitting of these two points. The difference between the simulated current and the straight line reaches 0.07 nA for the temperature of −40 °C, which means that the measurement error will be about 1.2 °C for the −40 °C point with conventional linear approximation. In contrast, the difference between the simulated current and the nonlinear prediction is only 0.02 nA at −40 °C, which corresponds to only a 0.34 °C measurement error. Obviously, the simulated current, I0, is closer to the hyperbolic line as predicted by (12) rather than the straight line. It can also be roughly calculated from Figure 3 that the measurement error using conventional linear approximation will reach 2.8 °C when the temperature is down to −60 °C.

Digitization and Calibration
The current controlled oscillator and the frequency-to-digital converter (counter) in Figure 1 are used to digitize the temperature dependent current, I0. As shown in Figure 4, the current controlled oscillator is implemented as a relaxation oscillator, taking advantage of the small area and low power consumption [31]. In the oscillator, the current, I0, is duplicated and amplified by p times to serve as the charging current. To ensure a linear relationship between the oscillation frequency and the charging current, p • I0, all the unwanted delays in the oscillator feedback loop are carefully optimized Current/nA

Digitization and Calibration
The current controlled oscillator and the frequency-to-digital converter (counter) in Figure 1 are used to digitize the temperature dependent current, I 0 . As shown in Figure 4, the current controlled oscillator is implemented as a relaxation oscillator, taking advantage of the small area and low power consumption [31]. In the oscillator, the current, I 0 , is duplicated and amplified by p times to serve as the charging current. To ensure a linear relationship between the oscillation frequency and the charging current, p · I 0 , all the unwanted delays in the oscillator feedback loop are carefully optimized and reduced. An edge-to-pulse generator is inserted in the loop to shortly turn on the switch transistor, M5, and reset the sawtooth waveform at the charge/discharge node periodically.  Ideally, the oscillator output clock cycle period, τ, is calculated as: where C0 is the timing capacitor at the charge/discharge node, Vref defines the sawtooth waveform magnitude, and p is the ratio of the current mirror. Considering the extra delay in the feedback loop, which is independent of the charging current, such as the delay caused by the comparators, the logic gates, and the discharging switch, M5, an offset component denoted as τos should be added to the clock cycle period, : In this work, the nominal value of the charging time, , and it follows that ⋅ ⋅ ≪ 1. To make the following derivation simple, Equation (14) is re-arranged and approximated as: , and then Equation (15) is simplified to: Ideally, the oscillator output clock cycle period, τ, is calculated as: where C 0 is the timing capacitor at the charge/discharge node, V ref defines the sawtooth waveform magnitude, and p is the ratio of the current mirror. Considering the extra delay in the feedback loop, which is independent of the charging current, such as the delay caused by the comparators, the logic gates, and the discharging switch, M5, an offset component denoted as τ os should be added to the clock cycle period, τ: In this work, the nominal value of the charging time, p·I 0 , is about 0.2 ms. Simulation shows that the extra delay contributed by the comparator, the logic gate, and the discharging switch is about 3.5 µs, 30 ns, and 3.5 ns, respectively. In general, τ os is much smaller than C 0 ·V re f p·I 0 , and it follows that τ os C 0 ·V re f p·I 0 1.
To make the following derivation simple, Equation (14) is re-arranged and approximated as: Let I OS = p · I 0 · τ os C 0 ·V re f p·I 0 , and then Equation (15) is simplified to: In this design, the oscillator output clock is divided by m times, and then the number of oscillation cycles within a given measurement time, τ cnt , is counted by the frequency-to-digital converter (counter) in Figure 1. The length of τ cnt is controlled by a reference clock. The nominal value of τ cnt is 1000 ms, and m = 4. The counter output number, D out , is given as: Combining Equations (16) and (17) yields: Substituting Equation (12) into Equation (18) gives: It is not necessary to know the value of each individual parameters in Equation (19). Let: Additionally, Equation (19) can be simplified as: T is defined as: Then, Equation (22) can be written as: It is anticipated that the values of k and b in Equation (24) vary randomly chip by chip with the random process variation, and other random effects, such as the reference clock period error. On the other hand, it can be seen from Section 3 that the value of the coefficient, K TC , in Equation (23) can be viewed as deterministic for all the chips fabricated using a given process.
To find the actual temperature value, T, from the counter reading, D out , the following calibration procedure is applied to compensate for both the random linear errors in k and b, and the deterministic nonlinear error as analyzed in Section 3. A two-point calibration is used to calibrate the linear errors. That is to say, each temperature sensor chip needs to be measured at two exactly-known temperature points, T 1 and T 2 . Step 1. For the specific chip, under the first given temperature settling (20 • C in this work), use a high performance sensor to measure the temperature, T 1 , and record the counter reading, D out1 , within a given measurement time; Step 2. Under the second given temperature setting (−30 • C in this work), use a high performance sensor to measure the temperature, T 2 , and record the counter reading, D out2 ; Step 3. Use Equation (23) to calculate T 1 = T 1 1+K TC (T 1 −T 0 ) and T 2 = T 2 1+K TC (T 2 −T 0 ) . Note that K TC = 7.10 × 10 −5 with T 0 = 300 K; Step 4. Use the linear Equation (24) to calculate the values of k and b with (T 1 , D out1 ) and (T 2 , D out2 ) from Step 1 to 3; Step 5. For each output value, D out , given by this sensor, use Equation (22) to calculate the actual temperature, T, with the obtained values of k and b from Step 4, and the values of K TC and T 0 from Section 3.
The proposed temperature sensor has been designed as a part of a wireless sensing SoC, and the calculation in Steps 3 to 5 can be implemented by programming the embedded microcontroller (MCU) in the SoC. For those application scenarios without the MCU, the calibration function can be easily implemented using a small digital logic circuit.

Measurement Results
The proposed temperature sensor chip was designed and fabricated as a function part of a SoC in a 0.13 µm CMOS process, and the die microphotograph of the SoC is shown in Figure 5a. Note that the temperature to current converter and the bias/reference generation circuit are shared by this temperature sensor and some of the other function parts in the SoC, and they are located apart from the current controlled oscillator and the frequency-to-digital converter, as shown in the sensor circuit layout given in Figure 5b. The total active area of the reported temperature sensor, including all the function blocks shown in Figure 1, is 0.0014 mm 2 . The proposed temperature sensor has been designed as a part of a wireless sensing SoC, and the calculation in Steps 3 to 5 can be implemented by programming the embedded microcontroller (MCU) in the SoC. For those application scenarios without the MCU, the calibration function can be easily implemented using a small digital logic circuit.

Measurement Results
The proposed temperature sensor chip was designed and fabricated as a function part of a SoC in a 0.13 μm CMOS process, and the die microphotograph of the SoC is shown in Figure 5a. Note that the temperature to current converter and the bias/reference generation circuit are shared by this temperature sensor and some of the other function parts in the SoC, and they are located apart from the current controlled oscillator and the frequency-to-digital converter, as shown in the sensor circuit layout given in Figure 5b. The total active area of the reported temperature sensor, including all the function blocks shown in Figure 1, is 0.0014 mm 2 . For the test purpose, the SoC was packaged in a 64-pin quad flat package (QFP-64). A photo of the decapped QFP-64 package with the SoC die in it is given in Figure 6. For the test purpose, the SoC was packaged in a 64-pin quad flat package (QFP-64). A photo of the decapped QFP-64 package with the SoC die in it is given in Figure 6. For the test purpose, the SoC was packaged in a 64-pin quad flat package (QFP-64). A photo of the decapped QFP-64 package with the SoC die in it is given in Figure 6.  The measured typical power consumption of the presented sensor is 0.15 µW (including all the function blocks shown in Figure 1), with a 1.1 V power supply at room temperature. The simulated power consumption breakdown is shown in Figure 7. The measured typical power consumption of the presented sensor is 0.15 μW (including all the function blocks shown in Figure 1), with a 1.1 V power supply at room temperature. The simulated power consumption breakdown is shown in Figure 7. Four chips randomly selected from the same lot were measured to validate the presented nonlinearity characterization and calibration. The sampling rate in the measurement is 1 sample per second, with an external reference clock of 1 Hz generated by a crystal oscillator with less than 25 ppm temperature drift over the measured range of −60 °C to 40 °C. Actually, the reference clock frequency does not need to be quite this precise, since this frequency error will be cancelled out during the calibration.
The measurement was carried out by placing the temperature sensor printed circuit board (PCB) board inside a temperature and humidity chamber (model ASR-0220 manufactured by ESPEC). A T2000 handheld thermometer manufactured by Xiatech Electronics with ±0.1 °C inaccuracy was used to measure and set the chamber temperature. Figure 8 shows the PCB used to test the sensor, and the measurement environment (the chamber). Four chips randomly selected from the same lot were measured to validate the presented nonlinearity characterization and calibration. The sampling rate in the measurement is 1 sample per second, with an external reference clock of 1 Hz generated by a crystal oscillator with less than 25 ppm temperature drift over the measured range of −60 • C to 40 • C. Actually, the reference clock frequency does not need to be quite this precise, since this frequency error will be cancelled out during the calibration.
The measurement was carried out by placing the temperature sensor printed circuit board (PCB) board inside a temperature and humidity chamber (model ASR-0220 manufactured by ESPEC). A T2000 handheld thermometer manufactured by Xiatech Electronics with ±0.1 • C inaccuracy was used to measure and set the chamber temperature. Figure 8 shows the PCB used to test the sensor, and the measurement environment (the chamber).
during the calibration.
The measurement was carried out by placing the temperature sensor printed circuit board (PCB) board inside a temperature and humidity chamber (model ASR-0220 manufactured by ESPEC). A T2000 handheld thermometer manufactured by Xiatech Electronics with ±0.1 °C inaccuracy was used to measure and set the chamber temperature. Figure 8 shows the PCB used to test the sensor, and the measurement environment (the chamber).  The measured temperature sensor outputs for the temperature range of −60 to 40 • C before the cycle-number-to-temperature conversion and calibration are shown in Figure 9. For the 100 • C range, the sensor digital output, D out , has a maximum difference of about 200, which indicates a measurement resolution of 0.5 • C. Clearly shown in Figure 9, there exists random linear errors (the slope error and the offset error). Though not quite visible, further calculation shows that there also exists hyperbolic nonlinear error as predicted in Section 3. The measured temperature sensor outputs for the temperature range of −60 to 40 °C before the cycle-number-to-temperature conversion and calibration are shown in Figure 9. For the 100 °C range, the sensor digital output, , has a maximum difference of about 200, which indicates a measurement resolution of 0.5 °C. Clearly shown in Figure 9, there exists random linear errors (the slope error and the offset error). Though not quite visible, further calculation shows that there also exists hyperbolic nonlinear error as predicted in Section 3. All the measured chips were then calibrated with the data acquired at the temperature points of T1 = 20 °C and T2 = −30 °C. After obtaining the calibration parameters (k and b), the temperature range of −60 °C to 40 °C with a step of 10 °C was measured for each chip. Figure 10 shows the temperature measurement error after the conventional linear error calibration. Since the nonlinear error is not taken care of, the maximum measurement error reaches 5.3 °C (the full range error is −1.8 °C to +5.3 °C). All the measured chips were then calibrated with the data acquired at the temperature points of T 1 = 20 • C and T 2 = −30 • C. After obtaining the calibration parameters (k and b), the temperature range of −60 • C to 40 • C with a step of 10 • C was measured for each chip. Figure 10 shows the temperature measurement error after the conventional linear error calibration. Since the nonlinear error is not taken care of, the maximum measurement error reaches 5.3 • C (the full range error is −1.8 • C to +5.3 • C). All the measured chips were then calibrated with the data acquired at the temperature points of T1 = 20 °C and T2 = −30 °C. After obtaining the calibration parameters (k and b), the temperature range of −60 °C to 40 °C with a step of 10 °C was measured for each chip. Figure 10 shows the temperature measurement error after the conventional linear error calibration. Since the nonlinear error is not taken care of, the maximum measurement error reaches 5.3 °C (the full range error is −1.8 °C to +5.3 °C).  Figure 11 shows the temperature measurement error after the proposed nonlinear error calibration using the steps given in Section 4. The maximum measurement error decreases to only 1.1 • C (the full range error is −0.9 • C to +1.1 • C).  Figure 11 shows the temperature measurement error after the proposed nonlinear error calibration using the steps given in Section 4. The maximum measurement error decreases to only 1.1 °C (the full range error is −0.9 °C to +1.1 °C). Note that the sensor design was optimized for the 1.1 V power supply. The measurement results shows that when the power supply deviates from the nominal 1.1 V by ±0.1 V, the oscillator output frequency may vary by ±0.4%, which corresponds to a temperature measurement error of about 1.2 °C. The presented sensor accuracy is sensitive to the power supply variation, and it needs to be used with the fixed power supply of 1.1 V for the best performance.  Note that the sensor design was optimized for the 1.1 V power supply. The measurement results shows that when the power supply deviates from the nominal 1.1 V by ±0.1 V, the oscillator output frequency may vary by ±0.4%, which corresponds to a temperature measurement error of about 1.2 • C. The presented sensor accuracy is sensitive to the power supply variation, and it needs to be used with the fixed power supply of 1.1 V for the best performance.
The performance of the presented temperature sensor chip is summarized and compared to other state-of-the-art designs in Table 1. Compared to other designs, the presented chip shows the lowest measurement temperature with the smallest chip area and the lowest power consumption, by using the proposed nonlinear error calibration. However, the power consumption reduction is actually at the cost of a narrowed measurement range at the high temperature end, which is clearly shown in Table 1. Note that the charging PTAT current was set quite small (~10 nA) to reduce the chip power consumption. On the other hand, the leakage current of the discharging switch transistor M5 approaches the charging current at high temperature. Consequently, the presented sensor design will fail at high temperatures. * Including the power consumption of all the function blocks shown in Figure 1, but not including that of the digital calibration function.

Conclusions
In this paper, a CMOS temperature sensor was presented for temperature monitoring down to −60 • C in cold chain applications. The nonlinear error in the conventional PTAT current based sensor circuit was characterized for the first time, and a two-point calibration method was proposed to compensate for the nonlinear error in addition to the traditional linear error calibration. With the proposed nonlinear calibration, the measurement error decreased to −0.9 to +1.1 • C for the temperature range of −60 to +40 • C. The temperature sensor chip occupied a die area of 0.0014 mm 2 , and the typical power consumption was only 0.15 µW from a 1.1 V power supply, which outperforms similar designs in the literature.