1. Introduction
This paper is a part of the research leading to developing a method to increase the efficiency of High Performance Processors (HPP) without interfering with their internal structure. The principle is the use of information not only about the processor thermal state, but also about changing environmental conditions.
The CPU temperature increases rapidly when a large amount of data is processed. Various techniques are used to prevent microprocessor overheating. The most popular solution is a passive heat sink combined with a cooling fan with adjustable rotation speed [
1]. The heat sink receives heat from the processor. The cooling fan generates forced convection conditions, which significantly increases the heat dissipation efficiency from the heat sink to the environment. Peltier modules are also used for cooling very often [
2]. The temperature of one side of the module decreases when the electric current flows through it—the heat is transferred to the opposite side [
3]. The Peltier module requires a cooling fan on the “hot” side of the module to prevent its overheating. Liquid cooling, using prefabricated microchannels integrated with an electronic circuit, is also reported in the literature [
4]. Forced circulation of the cooling liquid supports heat dissipation significantly. This is more effective than air cooling. The other possibility is an active heat sink. It can be implemented if the device is too small to integrate a massive radiator inside it (e.g., laptops) [
5,
6]. Active heat sink is composed of two copper blocks. CPU is mounted on the first one. Second block is placed near the housing edge and its task is to drain the heat outside the laptop. Both blocks are thermally connected by heat pipe channels—a tube filled with liquid. The liquid evaporates (near the CPU) and condenses (near the cooling fan) which is used for heat transfer (
Figure 1).
The above-mentioned methods of cooling processors ensure effective heat dissipation. However, this is not enough in modern systems. High CPU load processing of large amounts of data requires the CPU to be equipped with appropriate algorithms to supervise its operation. One of the basic extensions is the integration of CPU and temperature sensor. It allows implementing advanced thermal load control procedures. The CPU load depends on its temperature. Such solutions (measuring the temperature of the semiconductor structure in real time) are often used [
7,
8]. Using the sensors appropriate algorithms can be implemented to control CPU performance. It is known as Dynamic Thermal Management (DTM). It is a set of techniques that controls the processor temperature and prevents it from overheating; e.g., reducing of the CPU clocking (DFS—Dynamic Frequency Scaling) or supplying voltage (DVS—Dynamic Voltage Scaling) leads to reduce CPU processing speed and to decrease its temperature [
9,
10,
11,
12,
13]. DFS and DVS are based on CPU temperature monitoring using an integrated sensor. In general, the CPU load is determined by the current demand reported by the operating system. However, the efficiency is reduced if there is a risk of overheating. Consequently, the temperature decreases.
In the DTM method the key CPU temperatures are defined [
9]: the allowable temperature limit value (T
max), the first limit temperature value (T
thr1), the second limit temperature value (T
thr2) and the hypothetical limit temperature value (T
coerce). They meet the dependence T
max > T
thr1 > T
thr2 > T
coerce. The CPU works normally when its temperature is lower than T
thr2. When it exceeds this value, the oscillator output frequency or supply voltage is reduced. The processor slows down and its temperature decreases. The algorithm allows for the efficiency increase when the internal temperature of the CPU drops below T
thr1. Of course, in the DTM method it is crucial to determine T
thr1 and T
thr2 temperatures properly in order not to overheat the CPU.
One of the modern approaches is not only the CPU temperature control, but also monitoring of the thermal conditions in CPU surroundings [
14]. These conditions influence the CPU temperature. For example, there is a laptop in a room. Its processor is under heavy load. If the temperature in the room drops (e.g., in consequence of opening a window), the temperature of the heat sink block in contact with the environment drops first (right side in
Figure 1). Then the temperature of the heat sink block in contact with the CPU drops (left side in
Figure 1). Only then the temperature change will be detected by the sensor integrated in the processor. The reaction time will depend on the heat flow speed in the system. The thermal response time of the system (Point Heating Time—PHT) is long due to the high thermal inertia of the processor-radiator system, including a large mass of copper blocks [
10].
The proposed set-up will use an additional temperature difference sensor between the processor and the heat sink. It will monitor thermal changes in the laptop environment. Information about the external conditions’ changes will be delivered to the processor much faster. It allows to take appropriate actions in advance. It means that monitoring of the outside conditions helps to predict the thermal state of CPU in the near future. It permits to react faster to dynamic changes in the environment. It can be applied to optimize the CPU’s efficiency; e.g., the processing speed can be decreased when both—ambient and CPU temperatures increase. An example of the introduced algorithm is TΔT power control method described in the literature [
14]. Two temperature sensors are used in the method. One is integrated inside CPU (standard in modern CPUs). The second one is mounted on the heat sink, near the device housing. The second sensor monitors the ambient temperature and the CPU cooling conditions. The information about the temperature difference between CPU and its surroundings is obtained (TΔT). The results presented in the paper are related to TΔT power control method.
In this work a new temperature difference sensor for the TΔT system was developed to gain information for DTM algorithm. The requirement for the sensor was to monitor the temperature difference between the heat sink area near the processor and the heat sink area in contact with the environment (e.g., with air outside the laptop case—copper blocks, shown in
Figure 1). It was decided to use a thermocouple-based device.
The thermocouple is a simple device based on the Seebeck phenomenon. It consists of two different conductive or semiconductive materials connected by opposite ends, e.g., as shown in
Figure 2. An electromotive force will appear between the junctions when they are held at different temperatures and an electric current will flow in the system [
3]. The value of the resulting electromotive force is directly proportional to the temperature difference between the junctions and to the parameters characterizing thermoelectric materials, Seebeck coefficients (Equation (1)).
where:
α1, α
2—Seebeck coefficients of materials A and B;
THOT,
TCOLD—temperatures of hot and cold thermoelectric junctions, respectively.
The junction placed at a higher temperature is usually called “hot”. The junction placed in the lower one is called “cold”. Seebeck coefficients for conductive materials are in the range from several to a few tens of μV/K, for semiconductors—few hundreds of μV/K.
The wire thermocouples are the most frequently used in metrology. However, it is possible to fabricate thermocouples in the form of a planar structure applied on a flat, thin dielectric substrate, e.g., using screen-printing or magnetron sputtering methods [
15,
16,
17,
18,
19]. Several, several dozen or even several hundred thermocouples can be integrated on one substrate, forming a thermopile. Thermopile is a set of thermocouples connected electrically in series and thermally in parallel. In this way the electrical signal generated by the device is multiplied. If the thermopile consists of
n identical thermocouples, its output voltage will be n times higher. Such devices are used, e.g., to generate “green” energy for microelectronics [
20,
21,
22]. They can be used also to build a temperature difference sensor.
Different dielectric materials can be used as a substrate for planar thermocouples. Rigid materials such as alumina or glass are used, as well as flexible materials such as polyimide or polyester foils. However, if the thermocouples are fabricated in the thick-film technology, using the screen-printing method, it is necessary to use a substrate appropriate to the temperatures used in the technological process (850 °C in standard). The Low Temperature Cofired Ceramic (LTCC) gives great opportunity [
23]. It is a ceramic material supplied in the form of a thin, flexible tape. It can be freely formed in the planar plane—it can be cut, holes and channels can be made [
24,
25]. Single tapes can be folded in multilayers, creating three-dimensional (3D) structures with conductive paths buried inside [
20,
23,
24,
25]. During firing in a suitable time-temperature profile, the organic phase of the LTCC tape is removed. Ceramic grains are sintered. Consequently, a ceramic substrate similar in composition to alumina is obtained. Thanks to this technique it is possible to produce a ceramic substrate with a specific shape and to integrate elements such as conductive paths, via holes, contact pads, etc. into it.
3. Results
Sensor A was fabricated in two variants. In the first (marked A), the screen-printed legs were made from PdAg-based DP6146 ink (approximately 75% Ag). In the second variant (marked A2) from DP6145R ink, based on Ag. The other details were identical. A comparative study between the two was performed to see if the type of material used to screen-print the legs made a difference in sensor performance. The “hot” junctions were assembled near the CPU (CPU_heat_sink in
Figure 3a and
Figure 7b) on the copper block of the active heat sink. The “cold” junctions—on the copper block for heat dissipation (FAN_heat_sink). A thin alumina substrate with four thick-film resistors that simulated four processor cores was used as the CPU thermal model [
29,
30]. The results are presented in
Figure 14.
The tested structures differed significantly in internal resistance. This is directly due to the ink resistivity, which is several times higher for PdAg. However, the most important parameter is the voltage signal of the sensor. It was at a comparable level for both tested structures. The sensor output signal of type A is on average ~4% higher than that of type A2. This small difference may be due to the thermal conductivity of the materials, which is significantly higher for the silver-based ink. Consequently, somewhat larger temperature gradient may occur between the thermoelectric junctions of the A structure than in the A2 structure. Since the difference between the A and A2 sensor is small, the A sensor was chosen for further study.
To measure the variants A, N and E of the sensor at the same time, they were placed next to each other on the processor model (directly on the heater)—
Figure 15.
Figure 15a shows the location of the structure with “hot” thermocouple junction on the CPU_heat_sink. Three sensors are visible, from left to right variants N, E and A, respectively. The sensors were pressed down using the test needles to improve thermal contact with the heater surface.
Figure 15b shows the location of the structure with “cold” thermocouple junction on FAN_heat_sink.
After the sensors were placed on the active heat sink, measurements were started to compare all sensors. The processor model was heated up by connecting 12 V and 0.2 A to each of the resistors. Total power delivered to the heater (CPU_heat_sink) was equal to 9.6 W. The temperature of CPU_heat_sink and FAN_heat_sink was monitored by pyrometers. Internal resistance and voltage response of each sensor was measured. The sensors’ responses to changes in cooling conditions of the “cold” block (FAN_heat_sink) were investigated. For this purpose, the measuring set-up was equipped with a 12 V cooling fan, which was cyclically switched on and off. In this way the cooling conditions of the FAN_heat_sink block were changed. Information about changes in thermal conditions should be used by the algorithm controlling the processor. It is important to achieve the information as soon as possible. It should be kept in mind that the typical 4 GHz processor can perform up to 4 × 109 operations per second, so time is essential. All data were cyclically read and recorded by 34970A Data Logger (Keysight Technologies, Santa Rosa, CA, USA).
3.1. Experiment 1
In the first experiment power was delivered to the heater (CPU thermal model). As it was heated up the voltage responses of each sensor and its internal resistance were measured every 10 s. The entire test lasted about 3000 s. All sensors were measured simultaneously under identical conditions, allowing direct comparison. The results are shown in
Figure 16. An efficient active heat sink causes that the temperature difference between the “hot” Cu block (CPU_heat_sink) and “cold” one (FAN_heat_sink) is small. During all experiments, the temperature of both blocks varied from 50 °C to 64 °C and the temperature difference ΔT was established just after a few dozen seconds. The electrical responses of the sensors were stabilized at specific levels:
About 0.4 mV for version A.
About 1.2 mV for version E.
About 0.7 mV for version N.
This means, according to Equation (1), the sensor N measured ΔT of about 5.5 °C, while sensors A and E—about 3 °C. This shows the influence of sensor design on the measurement results. In versions A and E the thermoelectric junctions are located at the soldering points of the thermoelectric wires (points T1 and T2 in
Figure 8). In version N they are distanced from the wires (points T3 and T4 in
Figure 8). Consequently, the sensitivity of the sensor N is almost twice higher.
It can be explained by the fact that thermoelectric wires are located in the air (
Figure 8) near at room temperature. The air cools the wires down. This affects the temperature of the place where the wires are connected to the LTCC substrate—it is locally lower in relation to the rest of the LTCC substrate. If the thermoelectric junction is located there, its temperature is also lower. This is the case with versions A and E of the sensor (
Figure 8,
Figure 10a). In the version N, the thermoelectric junctions are distanced from the soldering point (
Figure 8 and
Figure 10b). Its influence on temperature is much smaller consequently. Note that the wires are relatively thick (250–260 µm) compared to the whole sensor (680 µm) and the screen-printed legs (15 µm). They have also a significantly higher thermal conductivity—12–429 W/m∙K for wires (see
Table 1) compared to about 3 W/m∙K for LTCC.
The results of the experiment confirm the thesis presented in
Section 2.2.1—the temperature difference between points T1 − T2 was lower than between T3 − T4 ones (
Figure 8). Distancing the thermoelectric junctions from the wires has a significant impact on sensitivity. The internal resistance of the sensors increases slightly as the heater warms up (
Figure 16b) because all used conductors were characterized by the positive temperature coefficient of resistance (TCR). The TCR is about 5900 ppm/K for Ni, 3800 ppm/K for Ag. The TCR for CuNi as well as NiCr depends on the atomic percentage ratio [
31]. It is about 30 ppm/K for used CuNi (Cu56Ni44) and 150 ppm/K for NiCr (Ni80Cr20). However, the temperature during all investigations remained between 23 °C and 62 °C, so the resistance changes were only slight.
To verify the results, numerical simulations were performed in COMSOL Multiphysics 5.3 (Comsol, Burlington, MA, USA). A simplified model was built, which corresponds to the tested sensor and the conditions of the tests. It consisted of two LTCC substrates with dimensions of 5.2 × 6.8 × 0.68 mm
3, with buried Ag and Ni legs (5000 × 300 × 15 µm
3). The substrates were connected by 160 mm long Ag and Ni wires with a diameter of 250 µm. The wires were mounted in LTCC substrates at a length of 2 mm. Typical physical parameters for LTCC, Ni and Ag (shown in
Table 1) were used as input data for the simulation. The ambient temperature was set at 20 °C. The system was heated with a 9.6 W heater. Free convection conditions were selected, the characteristic length was calculated as area/perimeter.
The simulation results are shown in
Figure 17.
Figure 17a shows a general view of the sensor. The LTCC substrate for CPU_heat_sink is on the left side. In
Figure 17b the sensor was cut by the plane passing through the center of the screen-printed legs (in Z axis). The temperature difference between LTCC substrates as well as between the points marked as T1 − T4 in
Figure 8 can be seen. Exact values can be read from the line graphs in
Figure 17c. The blue line indicates the LTCC substrate for CPU_heat_sink, green for FAN_heat_sink. These are the temperature distribution along the printed legs from
Figure 17b—the exact line for CPU_heat_sink (blue) is shown in
Figure 17d. According to the simulation results, the temperatures in T1-T4 points (see
Figure 8) are:
57.5 °C for T1.
55.5 °C for T2.
60.0 °C for T3.
55.8 °C for T4.
According to the simulation results, the temperature difference sensors should indicate:
2 °C—sensors A and E.
4.3 °C—sensor N.
These simulations are for a simplified model. However, results show the same trend as the measurement results and similar values. This confirms that the temperatures at points T1-T4 may be different, as the experimental results show. This allows to conclude that the geometric arrangement used in the N-type sensor is more effective than that used in the A-, A2- and E-type sensors.
3.2. Experiment 2
The purpose of the second experiment was to investigate the response of the sensors to changes in the ambient conditions of the system. Changes in the voltage response of the sensors (signal level) and their response times were monitored. The thermal model of the processor was already heated up. Both CPU_heat_sink and FAN_heat_sink temperatures were stabilized. The fan was cyclically turned on and off, changing the thermal conditions of the FAN_heat_sink block. From a physical point of view, the forced convection coefficient was changed. All sensors were measured simultaneously under identical conditions, allowing direct comparison. Voltage responses of all three sensors were recorded every 1 s. The times of switching on/off the fan are presented in
Table 2.
The whole experiment lasted about 110 s.
Figure 18 shows the responses of the sensors. The black dotted lines indicate the moments when the fan was switched on and off.
The responses of the sensors are almost instantaneous. After on/off the fan, the sensors’ reactions occur in the next measurement cycle. However, the interval of subsequent measurements is quite long—about 1 s. During this time the processor is able to perform many operations. Therefore, it was decided to conduct more detailed investigations (see Experiment 3).
The return time of the sensor output signal to the state before the change of cooling conditions is quite long. It can be estimated at over 30 s. However, this is not caused by the sensors themselves, but by the large heat capacity of the entire heat sink block.
The characteristics presented in
Figure 18d shows the influence of cooling condition change on FAN_heat_sink and CPU_heat_sink blocks temperatures. Every switching on the fan causes the temperature drop. Any change of the FAN_heat_sink temperature affects the temperature difference between it and the CPU_heat_sink.
The maximum level of the voltage response of the sensor A was about 1.1 mV, E—1.7 mV, N—0.9 mV. The thermal conditions were equal for each sensor. The sensor version E generates the highest output signal, because it is fabricated using the most thermoelectrically efficient materials (CuNi/NiCr, E-type thermocouple). However, more important parameter is the range over which the output signal changes (ΔU = ET
max − ET
min). The range ΔU can be extracted from
Figure 18 as 0.20–0.25 mV for sensor N, 0.45 mV for sensor E and 0.75 mV for sensor A. Sensor version A seems to have the highest sensitivity to thermal condition changes. It should be noted that the responses of sensors A and N differ from each other despite the fact that they were fabricated using the same thermoelectric wires.
3.3. Experiment 3
The third experiment was similar to the second one, but the measurement interval was reduced. The purpose was to investigate sensors response time with the accuracy better than in experiment 2. The thermal equivalent of the processor was already heated up. Both CPU_heat_sink and FAN_heat_sink temperatures were stabilized. The fan was cyclically turned on and off, changing the thermal conditions of the FAN_heat_sink block. The times of switching on/off the fan are presented in
Table 3. The fan was switched on for 10 s each time. The interval between switch off and switch on was 10, 15 or 20 s.
Both, the sensors voltage responses as well as the temperature of CPU_heat_sink and FAN_heat_sink blocks were investigated. The measurement interval was 0.209 s. This is the limit value for the used Keysight 34970A data logger. Between 650 and 800 measurements were taken for each sensor. Each sensor was measured individually, in a separate measurement cycle (it was necessary to obtain 0.209 s interval). Therefore, the thermal conditions were slightly different for each measurement.
Figure 19 shows the results for sensors A, E and N.
Characteristics presented in
Figure 19 show that the sensors responses are practically immediate. When the fan is turned on as well as it is turned off, the signal starts to change immediately. However, analysis of the subsequent measurement points revealed some differences between the sensors. The sensors increase the level of the generated signal in the first interval—i.e., the first measurement after switching the fan on/off shows a difference in the signal level. This is a slight difference, but it can be detected by the sensors used. For sensor N the change of the signal level by 1% occurs after approximately 0.2–0.4 s (first or second measurement cycle) after switching on/off the fan. On the other hand, the sensor E changes the signal level by 1% after 0.4–0.6 s when switching on and 0.4 s when switching off the fan whereas sensor A—0.4 s after switching on/off the fan.
4. Discussion and Conclusions
The paper presents the concept, fabrication and investigation of the hybrid thermoelectric sensor, which combines wire and thick-film thermocouples. Its task is to monitor the temperature difference at the opposite ends of the active heat sink, typically used in laptops.
Sensors were made by screen-printing of conductive inks on LTCC tapes. Every version of the sensor consists of two LTCC substrates (with printed paths) and thermoelectric wires. The LTCC substrates have direct contact with the copper blocks of the active heat sink, in the paper marked as CPU_heat_sink and FAN_heat_sink (
Figure 3a and
Figure 7b). This topology gives very good surface contact between the sensor and the active heat sink. Moreover, the measurement points (thermoelectric junctions) were distanced from the wires. Thanks to this, the negative influence of heat transport through the wires volume on the results was reduced.
The LTCC substrates are connected by thermoelectric wires, which allow the sensor to fit any shape of the active heat sink. This significantly simplifies the integration of the sensor with the heat sink. Three versions of the sensor, differing in the applied thermoelectric materials and the location of measuring junctions, were fabricated. They were marked A (or A2), E and N. Different thermoelectric materials determine the output signal level. The described investigations showed that all sensors fulfill the requirements and can be used for the assigned task. Sensor E has the highest electrical response. It uses CuNi and NiCr thermoelectric wires, forming a classical type E thermocouple. However, the signal from sensors A and N is also high enough for practical application.
The investigations shown that different locations of the thermoelectric junctions affect the sensor output signal. Thermoelectric wires have a large volume compared to screen-printed paths. The heat transport by them is therefore much higher. If the wires are close to the printed thermoelectric junctions (versions A, A2 and E of the sensor), they have a relatively strong influence on their temperature in comparison with sensor N, where the junctions are significantly distanced from the wires. This translates into several effects. Firstly, the temperature difference measured by the N sensor is about 80% higher than for sensors A and E (experiment 1,
Figure 16). This is a big advantage. When sensors are fabricated using the same thermoelectric wires, this directly translates into their output voltage level. A comparison of sensors N and A shows that distancing the junctions from the wires increases the output electrical signal from 0.4 mV to 0.7 mV (
Figure 16). The use of thermoelectric materials with better efficiency allows to obtain an even higher signal at the output. A comparison of sensors N and E shows that despite the short distance between the wires and the thermoelectric junctions, sensor E generates a signal of 1.2 mV, i.e., 70% higher than sensor N (
Figure 16). This is due to the fact that Seebeck coefficient of CuNi/NiCr thermocouple is more than 3 times higher than for Ag/Ni combination (68 µV/K and 21 µV/K, respectively). However, in this case other advantages of the N-sensor are lost (described below).
The second effect of distancing the thermoelectric junctions from the wires is increasing the temperature difference between the measuring points (thermoelectric junctions). For sensors versions A and E ΔT
A/E = T
1 − T
2 (see
Figure 8), for version N ΔT
N = T
3 − T
4. Results of the investigation show that ΔT
A/E < ΔT
N (experiment 1,
Figure 16 and
Figure 17). Consequently, the sensitivity of sensor N should be better than for versions A and E. However, the results obtained in experiment 2 do not support the above. The range ΔU over which the output signal changes is much wider for sensors A and E: about 0.25 mV for sensor N, 0.45 mV for sensor E and 0.75 mV for sensor A (
Figure 18). Further extensive comparative studies of the presented sensors should analyze why such results were obtained. One possible explanation is the specific conditions of the experiments. The forced cooling of the FAN_heat_sink block was using a fan. The generated airflow affected not only the temperature of the FAN_heat_sink block, but also the temperature of the thermoelectric wires. Lowering their temperature may have affected the temperature of the “cold” thermoelectric junctions in sensors A and E (in sensor N as well, but with lower impact due to the larger distance from the wires). This improved the performance of sensors A and E relative to N. It can be hypothesized here that the N-type sensor is better suited for monitoring the temperature of the FAN_heat_sink block, while the A and N-type sensors are better suited for monitoring its environment. This should be verified by further research by adding an experiment where the fan is replaced by a water-cooling system, for example.
The third effect is an improvement of the sensor reaction time. The sensor N exhibits a slightly faster change of the output signal than versions A or E. The difference may result from the volume (heat capacity) of the thermoelectric junction—in versions A and E it is a solder joint with a large volume (see
Figure 5). In version N the junction is much smaller (screen-printed legs, 15 µm in thick) and therefore its heat capacity is smaller. This may result in faster sensor response. The difference is noticeable but very small, the results were at the limit of the measuring performance of used instruments. The high-speed measurement system should be set up to better investigate this in future work. It should be noted that the reaction time is a second key parameter in the considered application of the sensor. In this context, the sensor N seems to be the optimal solution. This means that distancing the thermoelectric junctions from the wires is more important than the optimal choice of thermoelectric materials. The voltage response of all versions of the sensor is on a functional level.
The sensors are suitable for measuring the temperature difference in the proposed application, they can be used in further work on TΔT algorithms.