High performance low power CMOS temperature sensor

Abstract

A temperature sensor based on the combination of a temperature variable oscillator and a linear controlled oscillator is proposed, which can realize temperature detection through the characteristic of frequency changing with temperature. The changing frequency is generated by the two oscillators, and by adjusting the frequency linear change, the linearity of the sensor is also increased. Through the frequency digitizer, the digital signal can be output. Compensation through a process compensator improves the accuracy of the sensor after a single point correction. After conducting tests on 15 experimental samples, the accuracy was achieved within $\pm$ 1.2 ${}^{\circ}$ C across the temperature range of 0 to 125 ${}^{\circ}$ C, demonstrating a highly favorable level of precision.

Keywords

Temperature sensor energy conversion rate low power consumption

1. Introduction

With the continuous updating and iteration of advanced electronic equipment, accurate temperature detection and control are becoming more and more important for microprocessor thermal management, dynamic random access memory refresh control and other technologies, and the realization of these technologies requires temperature sensors [1]. It is necessary for the sensor to have as small a volume as possible, as low as possible power consumption, as high as possible accuracy, and as large a measurement range as possible, which are all key requirements for temperature sensors [2].

Precise band-gap temperature sensors based on analog-to-digital converters [3, 4] and temperature sensors based on thermal diffusivity sensing [5] are not suitable because of the high energy consumption and large area of this type of sensor [6]. Therefore, researchers have successively proposed temperature sensors based on delay lines that vary with temperature, temperature sensors with timing comparators, and temperature sensors based on delay-locked loops.

However, in order to achieve satisfactory temperature measurement resolution, traditional temperature sensors require hundreds of inverters, and due to loop capacitors, the area of the sensor is quite large, and the power consumption is also large, so that there is sufficient delay [7, 8]. In addition, when the power supply voltage changes, it does not operate correctly. Generally, the temperature detection accuracy based on the inverter delay module is only 3 ${}^{\circ}$ C [9, 10], mainly due to the nonlinear degree of the delay module (comparison of temperature and frequency), and process variation will cause slope asymmetry and power supply voltage jitter.

Figure 1.

The internal structure of temperature sensor in this paper.

Figure 2.

The relationship of oscillator frequency with temperature.

Therefore, a double oscillator structure is designed in this paper. The two oscillators are composed of temperature variable oscillator and linear control oscillator, which can improve the linearity of the sensor. The innovation is that one of the ring oscillators is inversely temperature-dependent, so that the frequency difference changes linearly with temperature. In addition, compared with the previous temperature detection [7], the error of the temperature sensor is reduced by using the process compensation scheme and power regulation, which is caused by process deviation and power instability. It is done using a 65 nm CMOS process with an area of only 0.01 square millimeters. The maximum error of temperature sensor is $\pm$ 1.2 ${}^{\circ}$ C in the range of 0 $\sim$ 125 ${}^{\circ}$ C by using single point correction method, which reduces the cost of mass production test.

2. Structural design of temperature sensor

The internal structure of the temperature sensor is shown in Fig. 1. This temperature sensor contains two oscillators. One of them is an oscillator that changes with temperature called a “temperature oscillator”, which can generate a frequency difference that changes with temperature. The other is a linearly controlled oscillator, which can improve the linearity of the sensor. However, because the bias circuit is different, each oscillator will have its own distinct characteristics. In a temperature-varying oscillator, its delay time and the oscillator’s series determine its frequency. Similarly, the mobility of a field-effect transistor (MOSFET) and the threshold voltage both decrease with increasing temperature. Since the mobility is significantly reduced, the frequency of the oscillator is negative in relation to temperature, that is, the temperature increases while the frequency decreases, as shown in Fig. 2(a). In the process corner FF TT SS, t stands for typical, s stands for slow (low current), and f stands for fast (high current). Generally, the first letter stands for nmos and the second letter stands for pmos (for example, TT means that nmos and pmos are typical). Generally, the temperature of a ring oscillator is nonlinear to its frequency. This leads to an increase in the error of the temperature sensor. It can be seen from the frequency temperature characteristic curve that if only one oscillator is used, the linear relationship between frequency and temperature cannot be obtained. Therefore, a linearly controlled oscillator is required to generate linearized frequency differences. The linear controlled oscillator can generate an output signal with positive frequency and temperature characteristics, that is, the frequency increases as the temperature increases, as shown in Fig. 2(b). The frequency difference obtained by adjusting the frequency characteristics is more linear. Because the temperature variable oscillator and the linear control oscillator have the same structural basis, the process variation can cancel each other out. The temperature sensor produced by this process not only has improved resolution, but also has higher accuracy, as measured in Fig. 2(c).

The power regulator is used for power regulation, the compensator is used for variation compensation, the controller generates the control signal for all operations, and the output digital signal is converted into the frequency difference by the frequency digitizer. Therefore, the output obtained is not sensitive to process variation as well as power supply variation.

3. Proposed oscillator and bias circuit design

The content structure of the temperature variable oscillator and the linear control oscillator is shown in Fig. 3. The ring oscillator consists of a current-driven delay module that can be used to promote common-mode noise suppression. Polyphase clocks are used to generate high-resolution digital outputs.

Figure 3.

Binary weighted current driven delay unit.

Due to the voltage peak reserve problem, it is difficult to implement the bias circuit of a conventional temperature sensor when the supply voltage is reduced to 1.2 V, as shown in Fig. 4(a). Because it is composed of three MOSFETs stacked, and the threshold voltage of each is less than half of the power supply, its sensitivity is extremely low. To solve this problem, this article uses another structure, as shown in Fig. 4(b). In the picture, the M4 and M7 play the same role as the M1, which ensures that the device operates in a saturated region. The use of this structure can save voltage. If the voltage of the bias circuit exceeds the threshold voltage of the transistor, both M4 and M7 are connected. M7 is connected to the supply voltage level and the source of M8. Diode connection PMOS M8 will turn on. The VBN can charge it via the M8. When the VBN voltage exceeds the threshold voltage, M6 and M9 will also turn on. Finally, the bias voltages VBP and VBN can be obtained.

Figure 4.

The bias circuit of the oscillator.

With the increase of temperature, the $\mu$ (carrier mobility) and VTH (threshold voltage) of the structure field effect transistor MOSFET shown in Fig. 4 both decrease [3, 4], as shown in Eqs (1) and (2):

$\displaystyle V_{\textit{TH}}\left(T\right)=V_{\textit{TH}}\left({T_{0}}\right% )+\alpha_{\textit{TH}}\left({T-T_{0}}\right)$ (1)

$\displaystyle\mu\left(T\right)=\mu\left({T_{0}}\right)\left({\frac{T}{T_{0}}}% \right)^{km}$ (2)

According to $\alpha$ power law, the current through M8 can be expressed as:

$\displaystyle I_{\textit{D.M8}}=\frac{1}{2}\left(T\right)C_{\textit{OX}}\frac{% W}{L}\left({V_{\textit{GS,M8}}-V_{\textit{TH}}\left(T\right)}\right)^{\alpha}$ (3)

The reason is that the M8 operates in a saturated region. The coefficient $\alpha$ represents the speed saturation power depending on the length of the device channel (1 $\leqslant\alpha\leqslant$ 2). By substituting the expressions of $\mu$ ( $T$ ) and VTH(T) into Eq. (3), and simplifying the operation, the following formula can be obtained:

$\displaystyle I_{\textit{D,M8}}=A\left({\frac{T}{T_{0}}}\right)^{km+\alpha}% \left({\frac{B}{T}-\alpha_{\textit{TH}}}\right)^{\alpha}$ (4)

Where, $A$ and $B$ are respectively:

$\displaystyle A=\frac{1}{2}\mu\left({T_{0}}\right)C_{\textit{OX}}\frac{W}{L}$ (5) $\displaystyle B=V_{\textit{GS,M8}}-V_{\textit{TH}}\left({T_{0}}\right)+\alpha_% {\textit{TH}}T_{0}$ (6)

If we extend the . $\alpha$ power term by using Taylor series, Eq. (4) can be expressed as:

$\displaystyle I_{\textit{D,M8}}=A\left({\frac{T}{T_{0}}}\right)^{km+\alpha}% \left({C_{0}+C_{1}T+C_{2}T^{2}+\ldots}\right)$ (7)

Where $C_{n}(n=1,2,3\ldots)$ The representation depends on $V_{\textit{GS,M8}}$ polynomial coefficients.

The delay of the delay module and the series of the oscillator determine the frequency F of the oscillator. The current of the delay module determines $\tau$ (delay time) as follows:

$\displaystyle\frac{1}{F}\alpha\tau=\frac{C_{\textit{load}}V_{50\%}}{I_{\textit% {sat}}}$ (8)

Among them, Cload is the output point capacitance of each delay module, $V_{50{\%}}$ is the transition voltage during the rise and fall transition, and Isat is the saturation current of MOSFET. The frequency difference between the temperature variable oscillator and the linear controlled oscillator can be expressed by the following Eq. (7):

$\displaystyle F_{T}\left(T\right)-F_{L}\left(T\right)=\left({\frac{T}{T_{0}}}% \right)^{km+\alpha}\{\left({A_{T}C_{0,T}-A_{L}C_{0,L}}\right)+\left({A_{T}C_{1% ,T}-A_{L}C_{1,L}}\right)T$ (9) $\displaystyle\quad+\left({A_{T}C_{1,T}-A_{L}C_{1,L}}\right)T^{2}+\ldots\}$

Among them, $F_{T}\left(T\right)$ and $F_{L}\left(T\right)$ represent the frequency of the temperature variable oscillator and the current control oscillator, respectively. In the above formula, km can be determined by the process (i.e. the impurity concentration) $\alpha$ can be determined by adjusting the device channel length so that there is a linear term in the formula above. Changing the coefficients A and Cn can eliminate other nonlinear terms. For example, if km is $-$ 1.5, we can choose the $\alpha$ of 1.5, so that the relationship between T(temperature) and $F$ (frequency) in the first order terms is linear, and other terms can be simplified by changing the values of VGS and W. If the voltage (VGS – VTH) is set to a small value, the linear controlled oscillator will have an inverse temperature correlation [10].

Delay unit time and saturation current are easily affected by fabrication variation. In this paper, the signal Sel $<$ 0: 3 $>$ of the CMOS temperature sensor to the fabrication process compensator will change the strength of the binary weighted current source of the delay unit, which can compensate for the variation of the fabrication process. This will also keep the oscillator slope unchanged, because km and $\alpha$ do not change with W, so W changes. Does not change the accuracy of the sensor. The two oscillators are supplied with a constant voltage of 1 V, and the regulator is used to reduce the power supply noise (33 dB), which is only $\pm$ 0.1 ${}^{\circ}$ C due to electrical noise. Therefore, the temperature sensor proposed in this paper has strong robustness to process variation and power supply noise.

4. Process compensator

Figure 5 is a block diagram of the process compensator, and Fig. 6 is its timing diagram. Figure 2 shows the relationship between the process conditions and the oscillator frequency. Then the process variation can be reflected by the frequency change of the oscillator, which is the role of the process compensator. The frequency range of the oscillator at the specified temperature in Fig. 5 can be detected by the “switch counter”, whose output is: q0 $\sim$ q2. If process compensation is to be performed, the compensator will reset the stored data (usually this data is typical process Angle). At the same time, the controller computes the output of the oscillator at a specific time (usually the time specified by the reference clock). The output of the counter has been converted by the encoder into a 4-bit select signal Sel $<$ 0:3 $>$ . This model can control the frequency of two oscillators. The compensation process ends after 2S. After that, the selection signal used by the temperature sensor is switched to the signal in test mode. In short, different process angles lead to different frequency curves, and the role of the process compensator is to arrange different frequency curves in the same frequency range within a similar group.

Figure 5.

Block diagram of process compensator.

Figure 6.

Sequence diagram of process compensator.

Figure 7.

Schematic diagram of frequency digital converter.

5. Frequency digitizer

The block diagram of the frequency digitizer is shown in Fig. 7, which includes an 8 bit lift counter, a sampler, and an exact code generator. Figure 8 is a timing diagram of the frequency digitizer. If it is a test mode, the frequency difference is calculated by the control signal and its calculated value is saved. When the count signal is “1”, the output value of the temperature variable oscillator becomes the input clock selected by the multiplexer to the digitizer. The counter then increases the rising edge of the input clock. Similarly, when the lift counter signal is “0”, the multiplexer outputs the input clock to the frequency digitizer using a linear control oscillator. At the same time, the rising edge of the input clock is reduced by the input of the computer. Every time a cycle is completed. The sampled signal will become a “1” and the sampler will store the value in the counter. In this way, the frequency digitizer can calculate and store the frequency difference between the temperature-varying oscillator and the linear-controlled oscillator until the sampler is “1” again. If you need to measure again, input a new signal will reset the counter, which is a complete digital temperature measurement. The conversion time is the same as the computer signal cycle. In order to prevent overflows, the frequency $F_{L}$ of the lifting counter should be lifted in accordance with the following conditions:

$\displaystyle\frac{1}{2F_{L}}<\frac{1}{F_{T\__{\textit{MAX}},L\__{\textit{MAX}% }}}\times\left({2^{8}-1}\right)$ (10)

Figure 8.

Frequency digitizer timing diagram.

Figure 9.

Quantization error of precision code generator.

The frequency of the lift signal is as follows, which can make the counter optimal:

$\displaystyle\frac{1}{2F_{L}}>\frac{1}{F_{\textit{T\_MAX,L\_MAX}}}\times 2^{7}$ (11)

Finally:

$\displaystyle\frac{F_{\textit{T\_MAX,L\_MAX}}}{2\times\left({2^{8}-1}\right)}<% F_{L}<\frac{F_{\textit{T\_MAX,L\_MAX}}}{2\times 2^{7}}$ (12)

In the precision code generator, the control signal can be used to sample the four polyphase clocks (temperature-varying oscillators $<$ 0:3 $>$ ) that are drawn from the temperature-varying oscillator. In frequency digitizers, only the temperature variable oscillator $<$ 0 $>$ is used to calculate the frequency difference. Therefore, when the number of rising edges of the temperature varying oscillator $<$ 0 $>$ is the same, the 8-bit coarse code is equal. However, once the polyphase clock is taken into account, different situations also occur on the same bold code. If N polyphase clocks (with $\pi$ / $N$ phase difference) are used in Fig. 9. The resolution of this temperature sensor is 2N times higher than other temperature sensors [5, 6].

The four polyphase clocks in this article increase the resolution by a factor of eight, as shown in Fig. 7. Compared to using an 11 bit counter, the precision code generator does not increase the resolution, but it can increase the measurement bandwidth by 3 times.

Figure 10.

The relationship between frequency and temperature in an analog temperature-varying oscillator.

Figure 11.

Simulation of the relationship between frequency and temperature in a linear controlled oscillator.

Figure 12.

Simulation results of linearized frequency difference between two oscillators.

Figure 13.

Relationship between temperature and frequency.

6. Simulation results

Under different process Angle conditions, the relationship between temperature and frequency of temperature variable oscillator and linear controlled oscillator is shown in Figs 10 and 11. If only the temperature variable oscillator is used for temperature sensing, the output code needs to be uniform under all process conditions using a two-point correction method. In addition, even if the two-point correction method is used, there may be large errors in different process angles due to nonlinear characteristics. This problem can be solved by using the frequency difference between the two oscillators, as shown in Fig. 12. Because the linear controlled oscillator is adjusted to compensate for the nonlinear characteristics of the temperature variable oscillator, the frequency difference is linearized at all process angles. However, the output code of this method still needs to be corrected at two points. In addition, the current of the delay unit also needs to be controlled by the process compensator and the selection signal. It’s the only way to compensate for process variation. Figure 13(a) is obtained after single-point process compensation in the same temperature range. Figure 13(b) shows the output error of the temperature sensor.

Table 1
Performance comparison of temperature sensors in this paper

Reference	Correction method	Accuracy ( ${{}^{\circ}}$ C)	Resolution ( ${{}^{\circ}}$ C)	Conversion rate (sample/sec)	Power loss ( $\mu$ W)	Energy/conversion ( $\mu$ J/ sample)	Supply voltage (V)	Area (mm ${}^{2}$ )	Technology
[5]	More correction method	0.25 ( $-$ 70–130 ${}^{\circ}$ C)	0.025	10	100	10	3.3	4.5	0.7 $\mu$ m
[4]	–	0.2 ( $-$ 55–125 ${{}^{\circ}}$ C)	0.02	–	500	–	1.8	0.2	0.2 $\mu$ m
[3]	A little correction	$-$ 1.8–2.3 (0–100 ${{}^{\circ}}$ C)	0.66	5 k	1200	0.24	1.2	0.12	0.13 $\mu$ m
[7]	A little correction	$-$ 2.8–2.9 ( $-$ 40–110 ${{}^{\circ}}$ C)	0.34	366 k	400	0.0011	1.2	0.0066	65 nm
[10]	A little correction	0.2 ( $-$ 70–130 ${{}^{\circ}}$ C)	0.03	2.2	9.96	4.52	1.2	0.1	65 nm
[6]	A little correction	3.5 ( $-$ 60–160 ${{}^{\circ}}$ C)	–	100	0.3	0.003	–	0.02	0.13 $\mu$ m
This paper	A little correction	1.2 (0–125 ${{}^{\circ}}$ C)	0.2	480 k	500	0.001	1	0.01	65 nm

Figure 14.

Equipment for temperature sensor experiments.

Figure 15.

Measurement error of 15 sensor chips.

7. Actual test results

The temperature sensor in this paper is designed using 65 nm CMOS technology. Figure 14 shows the experimental setup for testing the chip. The thermostatic model number is QRTH-781U, which is used for temperature control, the oscilloscope model is GDS-1062A, and the function signal generator is HP33120A. In the process compensation operation, first set the temperature to the middle of the temperature measurement range (such as 40 ${}^{\circ}$ C). When the temperature of the thermostat is stabilized at 40 ${}^{\circ}$ C, the mode selection signal will activate the process compensation. The compensator automatically detects the process conditions and stores the selection signal, and Sel $<$ 0:3 $>$ controls the delay unit current strength. Compensate as soon as you turn on. At the end of the compensation mode, the selection signal changes to the test mode and the temperature of the thermostat is stabilized at the lowest temperature ( $-$ 40 ${{}^{\circ}}$ C). Measurements are made every 10 ${}^{\circ}$ C in the range of $-$ 40 ${}^{\circ}$ C to 130 ${}^{\circ}$ C. The temperature stability is confirmed for each measurement, and the frequency digitizer is reset to reduce the experimental error. Since the maximum frequency of the temperature variable oscillator is 230 MHz (at $-$ 40 ${{}^{\circ}}$ C), the reference clock is 15 MHz to prevent overflows in any case. Thus, one conversion takes about 2 microseconds and the conversion rate is 480 kS/s. Under these conditions, the sensor resolution of 0.2 ${}^{\circ}$ C/LSB is measured by using 8 bit coarse code and 3 bit precise code. In the range of 0 $\sim$ 125 ${}^{\circ}$ C, the temperature accuracy of the 15 measured data samples is up to $\pm$ 1.2 ${}^{\circ}$ C, as shown in Fig. 15. Fifteen data samples were tested at each temperature. The data are averaged and compared with other temperature sensor test data. The latest microprocessors [1] need to process information from multiple temperature sensors. Therefore, to save power, the temperature sensor must have a small area and a high conversion rate. At a conversion rate of 480,000 samples/SEC and a supply voltage of 1.2 V (1 V for an oscillator-regulated supply), the power loss is 500 $\mu$ W. Thus, the energy per conversion is 0.001 $\mu$ J/sample, which is the minimum compared to recent similar sensors [6, 7, 10]. Table 1 compares the performance of each temperature sensor. If no temperature sensor is required, power consumption can be reduced by turning off both oscillators. Linearization and slope compensation enable single point correction, which can reduce the cost of high-volume production testing.

8. Conclusion

This paper presents a CMOS temperature sensor with two oscillators. The sensor uses an adaptive ring oscillator and an additional process compensator to improve linearity and process immunity. After using the one-point correction method, the maximum error of the temperature sensor is $\pm$ 1.2 ${}^{\circ}$ C in the range of 0 $\sim$ 125 ${}^{\circ}$ C, which reduces the cost of mass production testing. The entire model has an area of 0.01 mm ${}^{2}$ . When the conversion rate is 480 kS/s and the supply voltage is 1.2 V (the regulated supply for the oscillator is 1 V), the power loss is 500 $\mu$ W. The energy converted per time is 0.001 $\mu$ J/sample, which is the minimum value compared to other sensors. Adjusting the power supply of the oscillator can further reduce the error of the temperature sensor.

Footnotes

Acknowledgments

This project was funded by Chongqing Education Science planning project “Three Education Reform and Practice Research Based on the Construction of Two Platforms and one Workshop integrating Production and Education with gold Class” project No. 2021-GX-046.

References

Byun

. Categorization and characterization of time domain CMOS temperature sensors. Sensors. 2020; 20(22): 6700.

Lin

Ning

. A CMOS-thyristor based temperature sensor with+ 037. C/-0.32 C inaccuracy. Micromachines. 2020; 11(2): 124.

Cai

Wang

Zhang

Wang

. A pipeline-TDC-based CMOS temperature sensor with a 48 fJ⋅K2 resolution FoM. Electronics. 2021; 10(13): 1542.

Guo

Sun

Qiao

. A CMOS temperature sensor based on duty-cycle modulation with calibration. Analog Integr Circ S. 2020; 102(1): 79-89.

Sugiura

Miura

Nakano

. High-temperature operational piezoresistive pressure sensor on standard CMOS process. IEEE T Circuits-II. 2022; 70(2): 726-730.

Masui

. A CMOS Temperature Sensor with On-chip Photovoltaic Cells. IEEJ T Electr. 2023; 18(3): 401-407.

Xie

Prouza

Theuwissen

. A CMOS-imager-pixel-based temperature sensor for dark current compensation. IEEE T Circuits-II. 2019; 67(2): 255-259.

Someya

Mahfuzul Islam

AKM

Sakurai

Takamiya

. An 11-nW CMOS temperature-to-digital converter utilizing sub-threshold current at sub-thermal drain voltage. IEEE J Solid-ST Circ. 2019; 54(3): 613-622.

Goyret

Cassani

Carbonetto

Garcia-Inza

. Low-power temperature-independent CMOS measurement circuits for resistive gas sensors. Int J Circ Theor App. 2023; 51(6): 2562-2577.

10.

Xie