Sage Journals: Discover world-class research

Abstract

Background:

In this work, an accurate low-cost power microwatt-meter is designed, built, and tested. This microwatt-meter is found to be capable of measuring power consumption in a wide range of devices and sensors, such as nodes in wireless sensor networks, microcontrollers in Internet of Things devices, and Field Programmable Gate Arrays.

Methods:

The paper examines the usage of the microwatt-meter for the measurement of power consumption in digital devices as they execute program instructions, ranging from a low-level single step (instruction) in an assembly program to a high-level sophisticated program or system.

Results:

The results obtained showed that the microwatt-meter is highly accurate. It can measure the energy consumed by any of the devices/programs/instructions in a digital system with accuracy levels ranging between 0.5% and 2.5% (depending on the application).

Conclusion:

The accuracy of the microwatt-meter is validated using the fundamental conservation of energy law and a comparison with multimeter current measurements is presented.

Keywords

Power aware energy consumption microcontrollers programmable

Introduction

As the Internet of Things (IoT)¹ continues to grow exponentially, the need for improved designs of digital circuits increases.^2,3 Battery-operated designs that consume minimal amount of energy become more important. Such designs include microcontroller units (MCUs), sensory units, Field Programmable Gate Arrays (FPGAs), and so on. Due to limitations of the energy stored in batteries, designers are usually interested in measuring the efficiency of power consumption in various designs and its effects on the operation of the systems. For example, wireless sensor nodes usually have scarce resources especially in terms of energy, and they are expected to operate autonomously in unattended areas for long periods of time.⁴ The depletion of energy on one node may compromise the whole network operation and may degrade the quality of the monitoring process.⁴ Therefore, energy consumption has long been one of the crucial characteristics of wireless sensors to predict the network lifetime and to evaluate the efficiency of different WSN protocols and layouts. For example, the researchers in Reverter⁵ compared the power consumption of resistive and capacitive sensors.

In this work, a new design for a powerful and practical instrument is presented. A microwatt-meter is tailored to measure power consumption in battery-operated programmable devices with an accuracy of 0.5%–2.5% (depending on the application) within few seconds, without requiring significant preparation for the device under investigation. The built instrument is flexible and has a wide range of applications:

It is suitable for usage with any system. All that needs to be done is to power the system under consideration through the circuit. It does not require intervention or embedding the MCU/CPU/FPGA device in a special equipment/board for measurements as in Lee and colleagues,^6,7 for example.

It can be used to evaluate hardware and software for both macro and micro optimizations. This is achieved by evaluating the effect of changes from a single program instruction to a major system component or the whole system.

The designed microwatt-meter relies on measuring the voltage change across a capacitor (as will be discussed in section “Circuit design”). It is more accurate than other instruments that rely on current measurements.^6,8

It is cost efficient compared to other designs.^6,9,10 The built system costs about US$100.

The rest of the paper is organized as follows: The motivation and related work is discussed in section “Related work.” In section “Circuit design,” the circuit design is described in full detail. The power microwatt-meter validation process is presented in section “Power microwatt-meter validation.” Section “Experimental tests” presents the results of the experimental tests. Section “Sources of error” discusses the sources of measurement errors in this design. Finally, section “Conclusion and future work” describes future work and concludes.

Related work

In this section, the methods used to evaluate power consumption in electronic circuits are reviewed. They can be categorized into two main methods: measurement-based or calculation/simulation-based methods.^11–13

Measurement-based methods

In general, power can be measured using one of these two methods.

Using oscilloscopes and multimeters

One approach to measure power consumption is to monitor the current flow from a power source into the system in real time. This can be done using a high-speed multimeter or an oscilloscope, by letting the current pass through a known resistor and measuring the voltage drop across its terminals.

Chang et al.¹⁰ used an oscilloscope, to stream digitized voltages into a PC to calculate the profile of energy consumption. The work presented in Landsiedel et al.¹⁴ predicts the energy consumption of a whole sensor network based on the electric current measurement of a single sensor node. The authors in earlier works^9,8,15,16 measured current using a multimeter to find the consumed power of a CPU. In Kulkarni and Udupi,¹⁷ a method for energy estimation at the instruction level for ARM Cortex M4 processor was presented. The method relies on measuring the drawn current of a java script executed in a loop. Most of the previous work in this area, such as Jevtic and Carreras,^18,19 tried to validate their work by comparing their results with other multimeter/oscilloscope-based methods, or tried to validate their work by checking the repeatability of the measurement, when they should have opted for more fundamental techniques to validate their work.

Measurements using a multimeter/oscilloscope are prone to error^6,8,20 because the sampling rate of these devices is significantly lower than the clock speed of the MCU/CPU. Sampling a GHz CPU instruction current in a period of few milli-seconds can hardly be described as accurate.^21,22 The readings of these devices are assumed to be averaged, but in reality they ignore any sudden spike in drawn current during the execution. Another problem with this approach is the millions of collected data points that need to be logged to be processed later.¹⁰ This will complicate the system design and increase its cost as it will require extensive data storage components. Although high-speed oscilloscopes are available commercially, they cost tens of thousands of dollars and depend on a computer to analyze the logged data.

Using capacitors

To measure a household water consumption at the end of the day, it is easier and more accurate to observe the difference in the water level in the main tank, rather than meticulously measuring every single drop of water in real time. This analogy explains the biggest advantage of using capacitors over using multimeters/oscilloscopes.

Contrary to current/voltage monitoring techniques, this method does not require continuous high-speed sampling components. This is because all that need to be done is to record the voltage at the start and at the end of the measurement. The energy consumed can be calculated by measuring the capacitors voltage drop as will be explained in section “Circuit design.”

The authors in previous works^6,23 developed a system that uses fast switching capacitors to measure the energy consumption of a CPU (an ARM7 CPU was used as a test case). In their test, the CPU under study is soldered to the printed circuit board (PCB) and coupled tightly with an on-board FPGA. This makes repeating the measurement for different CPU or MCU lengthy, costly, and inconvenient.

Calculation/simulation-based methods

These methods rely on models that describe the power consumption of CPUs/MCUs,^17,24–27 sensors, and transceivers.^28,29 For example, PowerTossim³⁰ provides a per node estimate of power consumption through estimating the number of CPU cycles executed by each node. Wang²⁴ developed an instruction-level model to estimate the energy consumed by a program. They grouped instructions into categories and calculated the percentage of the instructions in each category.

In practice, CPUs/MCUs are always attached to other components to perform their tasks. Including all possible components in these models is not feasible because of the infinite possibilities. In addition, old models cannot describe newly released hardware. Therefore, power estimation models will always lag state-of-the-art hardware. Models also assume that different CPU/MCU implementations/packages consume the same amount of power regardless of the manufacturer. As will be shown later, the power consumption of the same MCU architecture can differ significantly from one manufacturer to another, or from one implementation to another.

Contribution

In comparison with the previous work, the contribution of the paper is threefold and is summarized as follows:

The paper provides a design for a low cost, accurate, and versatilely applicable power micrometer. The designed microwatt-meter targets battery-powered devices, and it is based on voltage measurements across capacitors, which achieves higher accuracy when compared to typical current-based measurements. Current-based measurements^6,8,9,16,20 rely on high-end multimeters that are capable of logging readings at a high sampling rate (higher rates give more accuracy). Even the fastest multimeter sampling rates are significantly slower than the commonly available CPUs/MCUs and their readings are considered to be averages that may ignore sudden current surges/drops.^21,22 This kind of equipment is usually costly and requires data storage to save the logged readings.¹⁰ The approach used in the microwatt-meter requires neither high sampling rates nor data storage; thus, it is more convenient and much cheaper. There is some similar previous work that used capacitors to measure power consumption,^6,23 but the used approach was not convenient and required a lot of preparation for the device or component to be evaluated. On the contrary, the microwatt-meter can be used directly with the device to be evaluated. All that needs to be done is use the microwatt-meter as a power source for the device or component through any convenient plug or connection. Unlike previous work, there is no need to solder the device under test to a special PCB.¹⁷ We believe the microwatt-meter is unique, and to the best of our knowledge, it is the first to provide such convenience and flexibility at such a low cost.

The microwatt-meter is implemented, and its accuracy is verified and compared to theoretical results. The microwatt-meter showed high accuracy, which validates the implementation.

The versatility of the microwatt-meter is practically demonstrated through a series of evaluation experiments targeting different abstraction levels from single instructions to system-level components.

Circuit design

The microwatt-meter is a controlled and monitored capacitor charge–discharge circuit. A schematic illustrating the functionality of the circuit is shown in Figure 1. As shown in this figure, a large capacitor is charged in a controlled manner using the switch formed by the n-MOSFETs Q1 + Q2. The device under test consumes the power it needs from capacitor C1. The discharge process is controlled using the switch formed by the n-MOSFETs Q3 + Q4. The charge/discharge processes are controlled using an MCU and a control logic that ensures proper operations as will be shown later. If the device under test requires a constant voltage, or an input voltage with a very narrow variation, then the charge stored in C1 is transferred to a smaller capacitor that is charged using the switch formed by the n-MOSFETs Q5 + Q6, and a feedback circuit that monitors C2 voltage. Q5 + Q6 will open when C2 voltage drops below a specific voltage $V_{m i n}$ and will close when a maximum target voltage $V_{m a x}$ is reached. The detailed circuit is shown in Figure 2. The circuit consists of the following major components.

Figure 1.

Power consumption meter schematic. The dark arrows represent the main current flow. The dotted arrows represent control signals that do not consume any energy from the main energy stored in $C_{1}$ and $C_{2}$ .

Figure 2.

Power consumption meter circuit design.

A large capacitor array

Power consumption measurements are achieved by measuring the voltage across the array at the start and the end of the process. This array is shown as $C_{1}$ in Figure 2. It can either be soldered on the PCB or be connected externally via JC1. The consumed energy is given by

E = \frac{1}{2} C (V_{2} - V_{1}) - E_{l e a k} (t)

(1)

where $E_{l e a k} (t)$ has two components: the lost energy through the components surrounding $C_{1}$ and the capacitors self-discharge when no load is present. The circuit was used using one of two types of capacitors:

Supercapacitors

Several arrays of supercapacitors with capacities ranging from 1 to 80 F and a maximum voltage of 16 V were used. The error in the labeled value is generally very large. It can be as high as 40% in some cases. Therefore, an accurate measurement must be performed to get the actual capacitance. Supercapacitors generally have a significant charge leakage that must be considered after doing a power consumption measurement. The measured charge leakage for one of the used arrays is shown in Figure 3. To minimize the error due to charge leakage, one must use a capacitor array with the lowest possible capacity to minimize the measurement error. Leakage is estimated using predetermined lookup table saved into the microwatt-meter MCU.

2. Aluminum electrolytic arrays

This type of capacitors has a low self-discharge rate, but a significantly lower capacity.

A $0.125 \pm 1 %$ Farad 16 V aluminum electrolytic array, that have an ignorable leakage, is used throughout this paper. This capacity gave excellent results for measurements that lasted few seconds on low power devices.

Figure 3.

52.4 F 16 V supercapacitor charge leakage.

Charging sub-circuit

The switch that enables charging the capacitor is the n-MOSFET Q2 in Figure 2. A Fairchild FDD8447L was chosen for this task because of its 15 A maximum continuous current to enable charging supercapacitors quickly. This transistor has a 8.5 mΩ on-resistance $(R_{O N})$ when switched-on using $V_{G S} = 10 V$ . The low resistance value means that no significant heat is generated by Q2. Moreover, this $R_{O N}$ is negligible compared to the charging-current limiting-resistor $R_{23} = 4$ Ω whose rated power is $100 W$ . The transistor Q2 has a maximum drain to source voltage $V_{D S} = 40 V$ , which enables higher charging currents. These current and voltage limits determine the specifications of the power supply $V_{D D}$ that can be used. Attention must be paid during the PCB design to the traces that carry high currents. The PCB used in the circuit shown in Figure 4 had 70-µm-thick and $90 mils$ wide traces that were as short as possible. The resistance of these traces is 0.3 mΩ which should not generate any significant heat even using the highest charging current.

Figure 4.

A prototype of the power consumption meter circuit in Figure 2.

Since the charge/discharge processes were driven using 5 V logic, the transistors Q2 was switched using another transistor Q1 to step-up its driving voltage in order to minimize Q2 resistance. The logic that drives Q1 will be discussed in section “Charge–discharge logic.”

Discharging sub-circuit

N-MOSFET PSMNR90 from Nexperia was used as a discharging switch, which is shown as transistor Q4 in Figure 2. This n-MOSFET has $R_{O N} = 1.0 m$ Ω, 120 A maximum continuous current, and 30 V maximum drain to source voltage. Q4 is driven using Q3 to step-up its driving voltage in order to minimize its on-resistance. The time it takes to turn this transistor on or off is about 200 ns which is fast enough for its purpose. The states of all transistors Q1 through Q4 is monitored using a set of light emitting diodes (LEDs) not shown in the figures. It should be emphasized here that the PCB traces carrying large charge/discharge currents must be designed carefully to avoid disturbing the MCU ADC readings.

Charge–discharge logic

The states of the charging transistor Q1 and the discharging transistor Q3 are determined using the logical sub-circuit formed by the integrated circuits: 74HC00 NAND gates, 74HC08 AND gates, and U2B AD8642 operational amplifier working as a comparator (see Figure 2). The used 74HCXX chips are TI parts SN74HC00N and SN74HC08N 14-DIP that have 2–6 V supply voltage range, the same LOW/HIGH logic-level voltage range, and the same 17 ns propagation delay. The function of the comparator U2B is to determine whether capacitor $C_{1}$ reached its target voltage during the charging process. This is done by selecting the proper values for the resistors R₁₀, R₁₂, and $R_{24}$ using the following conditions

(\frac{1}{3}) 5 V = \frac{R_{12}}{R_{12} + R_{10} + R_{24 (min)}} V_{c (min)}

(2)

(\frac{1}{3}) 5 V = \frac{R_{12}}{R_{12} + R_{10} + R_{24 (\max)}} V_{c (\max)}

(3)

The values of all resistors must be as high as possible to minimize the leakage from $C_{1}$ . For example, choosing the resistor values shown in Figure 2 will trigger the comparator output at $V_{C 1} = 2 - 12.5 V$ regardless of $V_{D D}$ value by sweeping the trimmer $R_{24}$ from 0 Ω to 1.0 MΩ.

In order to be able to start the charging process, three conditions must be satisfied:

The comparator must signal that $V_{C 1 (\max)}$ is not reached yet.

Charging must be enabled by setting pin D5 on the MCU to HIGH.

Start the process by setting pin D11 on the MCU to LOW.

On the contrary, the discharging process can be enabled by setting pin D11 and D6 to HIGH. The status of all of these pins and transistors is monitored visually using LED indicators.

The logic network, formed by 74HC00, 74HC08, and U2B (AD8642), control the source pins of Q1 and Q3. Setting the source of Q1 to LOW will start the charging, while setting the source pin of Q3 to LOW will start the discharging process. The status of the source pins of Q1 and Q3 is monitored using the MCU digital pins D0 and D1, respectively. For the purpose of troubleshooting the circuit using the LED indicators, note that Q2 gate logic is the opposite of Q1 source logic (D0), and Q4 gate logic is the opposite of Q3 source logic (D1).

Using the comparator as a way to control $C_{1}$ maximum voltage, instead of relying on the power supply voltage, enables faster charging by using a high $V_{D D}$ value. Charging is stopped when the output of the comparator goes HIGH. This is essential when using a supercapacitor, as charging it to a voltage equal to $V_{C 1}$ will take a very long time if $V_{D D} = V_{C 1}$ .

MCU

Two MCUs were tested: Arduino Nano 5 V with Atmega328P and NXP MK20DX256 Cortex-M3. The MCU monitors $C_{1}$ voltage using the voltage divider formed by $R_{11}$ and $R_{8}$ . It also monitors the comparator non-inverting input and the stability of the 5 V power line. All digital pins D0 through D6 shown in Figure 2 are controlled/monitored using this MCU. It will also estimate $C_{1}$ leaked/lost energy and calculate the consumed energy during the discharge process. All analog inputs of the MCU are buffered using op-amps and stabilized using 100 pF capacitors since polling those terminals (A0, A1, and A2) by the MCU ADC will cause a voltage drop at those inputs due to the high values of the resistors attached to them. The MCU ADC was configured to run at 15kSPS.

An I2C LCD display is connected to the MCU using jumper $J_L C D$ . When capacitor $C_{1}$ is charged and ready, the MCU will set pin D2 in jumper J_RS to HIGH to signal that the power monitor circuit is ready. The load system, whose power consumption will be monitored, is connected to jumper J_Load. The load system, or an external controller, must set pin D3 in jumper J_RS to HIGH to initiate the measurement process. Setting D3 to LOW will stop this process. The energy consumed, while pin D3 is HIGH, is calculated using equation (1).

Optional constant voltage unit

Most devices have a linear voltage regulator that can handle the variable voltage of capacitor $C_{1}$ as it discharges through the device. However, some sensors do not have a voltage regulator. To be able to handle such sensors/devices that are not tolerant to power supply wide voltage variations, a second smaller capacitor $C_{2}$ is charged using $C_{1}$ via a switch formed by the two transistors Q5 and Q6. The two comparators U6A and U6B, along with the NOR gate shown, ensure that $C_{2}$ voltage remains within a narrow range that can be controlled by carefully setting the trimmers $R_{3}$ and $R_{5}$ . To ensure minimal energy loss during the switching process, IPB180N03S4L n-MOSFET is used. This transistor has 5 ns rise time and $R_{O N} = 1.0 m$ Ω. It can also handle 720 A pulse current. This tolerance to high currents is critical because no resistor is present between $C_{1}$ and $C_{2}$ .

Power microwatt-meter validation

Previous attempts to measure the power consumption of CPUs/MCUs have their own inaccuracies. In order to validate the microwatt-meter, the consumed energy by a known load is compared against what is expected by the fundamental conservation of energy law.

To validate the microwatt-meter design, a known resistor, $R_{L o a d} = 218$ Ω, is connected as a load resistance to the microwatt-meter. The capacitor array, $C = 0.125 F$ , is charged to a known voltage, then discharged through $R_{L o a d}$ for a known period of time using the microwatt-meter which then measures the voltage across the capacitor array. This process is repeated for different time intervals. The lost energy by the capacitor equals the energy dissipated by the resistor. This fact is expressed using the following fundamental conservation of energy formula

\frac{1}{2} C (V_{i}^{2} - V_{f}^{2}) = \int_{0}^{t} \frac{V {(t)}^{2}}{R} d t

(4)

where

V (t) = V_{i} e^{\frac{- t}{R C}}

(5)

is the voltage of a resistor in an RC circuit. The leakage current for the used aluminum electrolytic capacitor array is ignored as the measurement time is typically less than few seconds. Figure 5 shows both the theoretically expected result (the solid line) and the experimental result with error bars (the dots). All experimental data points are within ±2.5% of the theoretically expected result. The average error is 1.2% for all data points.

Figure 5.

Comparison between the theoretically expected energy consumption of a 218 Ω resistor, and the experimentally measured consumed energy using the microwatt-meter.

Experimental tests

As mentioned earlier, the microwatt-meter can be used to measure power consumption at different levels of abstraction. Recall that the measurement relies on two control signals, the ready and measure signals as discussed in section “MCU.” Regardless of what is to be measured, the microwatt-meter should be used in the following manner:

Use the microwatt-meter to power the load system under consideration through a suitable connector.

The load system should wait for the ready signal from the microwatt-meter, which indicates that the capacitors are charged to the required voltage and ready to be used. Note that to get accurate measurements, it is important to wait until the capacitors are charged before starting.

The load system starts the measurement by setting the measure signal to HIGH.

Let the load system do the task that needs to be evaluated.

To end the measurement, the load system sets the measure signal to LOW.

It is worth mentioning that the control signals in steps 2, 3, and 5 can be maintained by the system under consideration itself or by an external controller.

In this section, the microwatt-meter ability is demonstrated at three different levels:

System level. As a case study, the energy consumption of an XBee S2B module on four different microcontroller platforms is measured in section “System level energy measurement.” The energy cost of preparing and transmitting a single byte is shown.

Component level. As a demonstration of this ability, the energy cost of only transmitting a single byte using XBee S2B transceiver is calculated in section “Component level energy measurement.” In addition, the energy wasted by the voltage regulator of Arduino UNO R3 board is shown.

Instruction level. The power consumption of adding two 32-bit integers, using C language, on ATSAM3X8E Cortex-M3 microcontroller is presented in section “Instruction level energy measurement.”

A comparison with multimeter current-based measurements is presented at the end of each subsection.

System level energy measurement

As a case study for a system-level measurement, an XBee S2B module is mounted on four different platforms. The platforms use AVR-based boards, three of which use the Atmega328P MCU and the fourth uses Atmega2560 MCU. The three Atmeag328P boards are an original Arduino Uno Rev 3, a generic Arduino compatible, and a Sparkfun RedBoard. The fourth board is an Arduino Mega. The XBee and the MCU board together comprise the load system for which the measurements were done. The task in step 4 is carried out by the MCU and the XBee module together, where the packet gets prepared by the MCU and transmitted by the XBee.

For each board, the XBee is used to transmit payloads of different lengths ranging from 10 to 200 bytes. For each payload length, the number of transmissions was varied from 10 to 500 in steps of 10. The results for the original Arduino Uno R3, the generic Arduino compatible, the Sparkfun RedBoard, and the Arduino Mega are shown in Figure 6. As expected, the consumed energy increases linearly with the number of transmissions. A better estimation of the cost per transmission can be calculated by taking the slope of each of the shown lines. Figure 7 aggregates all of these results together and enables the calculation of the energy cost per byte as shown in Table 1.

Figure 6.

Energy consumed versus transmission repetition count of a ZigBee packet (payload: 10–200 Bytes) using XBee S2B module, mounted on four different MCU boards. XBee module settings: Router, API 2, 115,200 bps, Power Level 4; Arduino IDE 1.8.5 optimization setting: Os; primary left axis: energy (J); and secondary right axis: time (s).

Figure 7.

Energy consumed to prepare and send a single ZigBee packet using an XBee S2B module attached to different microcontroller boards. Each point is the result of fitting the corresponding data set in Figure 6. The error in each point ranges from 0.5% to 2.0% as calculated by Levenberg–Marquardt algorithm. XBee module settings: Router, API 2, 115,200 bps, Power Level 4.

Table 1.

Energy consumed during the transmission of a single byte using XBee S2B module mounted on four different MCU boards.

System: XBee+	Using microwatt-meter		Using multimeter			Energy difference (%)
System: XBee+	Energy/byte	Power (W)	Continuous current draw (mA)	Energy/byte Vi∆t/NBytes (µJ)	Power (W)	Energy difference (%)
Arduino UNO R3	87.96 µJ ± 2%	1.04	123 ± 5	108	1.14	23
Generic	95.99 µJ ± 2%	1.13	120 ± 5	106	1.12	10
RedBoard	55.52 µJ ± 2%	0.69	98 ± 5	86	0.91	56
Arduino Mega	77.53 µJ ± 2%	0.92	100 ± 5	88	0.93	14

MCUs: microcontroller units.

The error presented is calculated using Levenberg–Marquardt fitting algorithm.

To calculate the consumed power for each system, the time to complete the transmissions stated previously is measured. The results are shown in Table 1. Note that the energy consumption of the three Atmega328P boards is different even though they are using the same MCU with the same claimed specifications. These measurements highlight the weaknesses of relying on simulation and estimation models as mentioned in section “Calculation/simulation-based methods.”

Although the fundamental conservation of energy law is the ultimate reference point in judging energy-metering techniques, a comparison with current-based measurements is presented. All currents throughout this study were performed using BK-388A Test Bench multimeter. Such measurements were done by putting an instruction (sending ZigBee packet n this case) in a loop and recording the drawn current. By measuring the source voltage $V_{s} = 9.3 V$ , the current i, and the time ∆t it takes to perform $N_{t r a n s}$ transmissions, one can calculate the energy per byte using the equation

E n e r g y / b y t e = \frac{V_{s} i Δ t}{N_{t r a n s} N_{B y t e s}}

(6)

where $N_{B y t e s}$ is the number of bytes in a ZigBee packet. The result for all used boards is shown in Table 1. Note that RedBoard has the highest difference between the current-based measurement and the microwatt-meter one. To understand the reasoning behind those numbers remember that the current consumed by a load circuit consists of two parts: the quiescent current or DC part of the total current, and a fluctuating current that changes from clock cycle to another. The multimeter current reading is an averaged one. The larger the DC component compared to the fluctuating component, the lower the error in the averaged reading. The RedBoard has a measured 7.7 mA current when the Atmega328P chip is in power-down mode, while the Arduino UNO R3 board has a measured 41.1 mA current when in power-down mode. This argument highlights another problem with measurement methods that try to monitor current/voltage in real time.

Component level energy measurement

To demonstrate the ability of using the microwatt-meter on a component level, the energy consumption of the XBee S2B module alone is measured. This is accomplished by powering the XBee alone through the microwatt-meter. The Arduino board was powered from a different power supply. In this case, the load system under consideration is the XBee only, and the measurement is controlled by an external controller which is the MCU board. Figure 8 shows the consumed energy for transmitting a 160-byte payload a number of transmissions varying from 50 to 500 using the maximum transmission power level. From the figure, one can calculate that the energy cost per transmission is $7.13 \pm 0.18$ mJ/Tx. In addition, the calculated average power during this process is bounded by 0.6 W, which is consistent with the manufacturer’s specifications. The slight curvature of the line in Figure 8 is due to the non-linear efficiency of the XBee S2B MC33269 voltage regulator. To find the cost of preparing a packet by the MCU, the XBee power consumption is subtracted from the data in Table 1.

Figure 8.

Energy consumed by XBee S2B module only; payload = 160 bytes; XBee settings: Router, API 2, 115,200 bps, power level 4; energy per packet $= 7.13 \pm 0.18 mJ$ .

After putting the XBee module in a transmission loop, the measured current drawn from a 9.3-V voltage source was $64 \pm 2 mA$ . This translates into 9.2 mJ per transmission compared to 7.13 mJ using the microwatt-meter data shown in Figure 8.

Arduino UNO R3 board can be powered either via a 12-V external connector or via the on-board USB connector. Powering the board using the USB connector will bypass the on-board voltage regulator. The on-board voltage regulator will waste energy even if the Atmega328P MCU is powered down or even removed completely from the board.

Arduino UNO R3 board power consumption was examined further under three different scenarios:

The board was powered using the 12-V connector with the Atmega328P in “Power Down” power saving mode.

The board was powered via the on-board USB connector (bypassing the voltage regulator) with the Atmega328P in “Power Down” power saving mode.

The board was powered via the 12-V connector with the Atmega328P completely removed from its socket.

The results are shown in Figure 9. The fitted slopes of the data shown in Figure 9 give the average consumed power for the three different scenarios. The result is summarized in Table 2. It should be noted here that the power characteristics of Atmega328P stated in its data sheet are for the isolated chip only. The results presented here are for the whole board. A comparison between the microwatt-meter measurement and a multimeter measurement is shown in Table 2. Note that in this table, unlike the result in Table 1, the multimeter and the microwatt-meter results are very close. This is due to the fact that the load draws a constant current in this case.

Figure 9.

Energy consumption of Arduino Uno R3 board in three different scenarios. Atmega328P ADCs disabled when powered down. PRR register = 0xFF.

Table 2.

Measured power consumption of an Arduino UNO R3 board under different test conditions.

Arduino UNO R3 board description	Using microwatt-meter		Using multimeter			Difference (%)
Arduino UNO R3 board description	Supply voltage range (V)	Power (W)	Supply voltage (V)	Current (mA)	Power (W)	Difference (%)
Regulated power Atmega328P PowerDown	8.0–8.8	0.394	9.3	41.1	0.382	3
Unregulated power Atmega328P PowerDown	4.1–4.7	0.150	4.9	32.3	0.159	6
Regulated power chip removed from board	8.0–8.8	0.360	9.3	41.1	0.382	6

Atmega328 P ADCs disabled when powered down. PRR register = 0xFF.

Instruction level energy measurement

The addition of two 32-bit numbers is chosen to demonstrate the ability to calculate the energy of a single instruction. The increasing energy cost of repeating this instruction is shown in Figure 10. From this figure, it is found that the energy of a single 32-bit addition instruction is $79.3 nJ \pm 2 %$ . A current-based measurement draws a current of 76.1 mA using 9.3 V voltage source. Using the data in Figure 10, it can be shown that the average energy per instruction is 72 nJ. That is a $9 %$ lower than the microwatt-meter measurement.

Figure 10.

Energy consumption of adding two 32-bit integers on Arduino Due board using C code. Compiler optimizations are disabled. The slope gives the energy per instruction: $79.3 \pm 0.6 nJ$ /instruction. The dotted straight line is the experimental data best line fit.

Sources of error

There are several sources of error in the data shown. Most of the error is due to the following factors:

Capacitors self-discharge and leakage through the surrounding components. This factor is important when using supercapacitors. However, it is ignorable when dealing with aluminum electrolytic capacitors. The error from this source was minimized by measuring the capacitor self-discharge rate, and then programming the microwatt-meter MCU to take this into account when calculating the consumed power.

Analog to digital conversion: Any voltage read by the MCU is reduced first by a voltage divider before sending it to the ADC. This is done to protect the MCU analog inputs from high voltages. The used MCU has a 12-bit resolution. This accuracy is dropped by a factor of three because of the used voltage divider. A separate external ADC with a higher resolution should enhance the results significantly.

Voltage regulators efficiency and (non)linearity. XBee modules, for example, have 3.3 V regulator that can handle up to 20 V power sources. The efficiency of this regulator drops with higher input voltages. This source of error depends on the load system design. This error can be minimized by working in the linear region of the load system voltage regulator.

The timing induced by the exchanged control and handshake signals may result in a negligible error. For example, the discharge transistor Q4 has a 200-ns transition time. The power dissipated during this time is negligible when compared to the few seconds of measurement period. However, it should be taken into account if the measurement time is less than 1 s.

Main capacitors $C_{1}$ and $C_{2}$ capacity measurement error. The stated capacitance of commercially available capacitors is usually very large (10%–20% or even more). To increase the energy measurement accuracy, the capacitors’ capacity was measured by connecting them in simple RC circuit with a known resistor. Since the energy stored in a capacitor depends linearly on its capacitance, the calculated stored energy will have the same error percentage.

Conclusion and future work

In this paper, a simple, low cost, yet efficient, design of a power consumption meter for programmable and/or battery-operated devices and sensors is presented. The device ability to measure energy on system level, component level, and instruction level is demonstrated. The cost of sending a singly byte using XBee S2B transceivers is calculated. The device accuracy is demonstrated by calculating the cost of adding two 32-bit integers using Cortex-M3 microcontroller. The results obtained were consistent with the manufacturers specifications. But more importantly, the accuracy of the device was verified using the fundamental conservation of energy law. Its accuracy is found to be as high as 0.5%.

The initial results for the power consumption of assembly instructions and code written in higher level languages are very promising. Different instructions and architectures of various MCUs/CPUs/FPGAs and wireless sensors will be tested. The device design will be upgraded with a better ADC and MCU units.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Mohammad Y Al-Shorman

References

Atzori

Iera

Morabito

The internet of things: a survey. Comp Networks 2010; 54(15): 2787–2805.

Yan

Sun

Qian

Energy-aware sensor node design with its application in wireless sensor networks. IEEE Trans Instrument Measure 2013; 62(5): 1183–1191.

Weddell

Merrett

Harris

, et al. Energy harvesting and management for wireless autonomous sensors. Measure Contr 2008; 41(4): 104–108.

Akyildiz

Sankarasubramaniam

, et al. Wireless sensor networks: a survey. Comp Networks 2002; 38(4): 393–422.

Reverter

Power consumption in direct interface circuits. IEEE Trans Instrument Measure 2013; 62(2): 503–509.

Chang

Kim

Lee

HG.

Cycle-accurate energy measurement and characterization with a case study of the arm7tdmi [microprocessors]. IEEE T VLSI Syst 2002; 10(2): 146–154.

Lee

Choi

, et al. Cycle-accurate energy measurement and characterization of FPGAs. Analog Integr Circ S 2005; 42(3): 239–251.

Johnson

MR.

Fast, accurate power measurement and optimization for microprocessor platforms. PhD Thesis, University of Illinois at Urbana-Champaign, Champaign, IL, 2015.

Kavvadias

Neofotistos

Nikolaidis

, et al. Measurements analysis of the software-related power consumption in microprocessors. IEEE Trans Instrument Measure 2004; 53(4): 1106–1112.

10.

Chang

Nagel

Muftic

Assessment of energy consumption in wireless sensor networks: a case study for security algorithms. In: IEEE international conference on mobile adhoc and sensor systems (MASS 2007), Pisa, 8–11 October 2007, pp. 1–6. New York: IEEE.

11.

Ahmad

Gani

Ab Hamid

, et al. A survey on energy estimation and power modeling schemes for smartphone applications. Int J Comm Syst 2017; 30(11): e3234.

12.

Moschitta

Neri

Power consumption assessment in wireless sensor networks. In: Fagas

(ed.) ICT-energy-concepts towards zero-power information and communication technology, pp. 203–224. London: Intech, 2014.

13.

Adamo

Attivissimo

Di Nisio

, et al. Comparison of current sensors for power consumption assessment of wireless sensors network nodes. In: 2017 IEEE international workshop on measurement and networking (M&N), Naples, 27–29 September, pp. 1–4. New York: IEEE.

14.

Landsiedel

Wehrle

Gotz

Accurate prediction of power consumption in sensor networks. In: The second IEEE workshop on embedded networked sensors (EMNETS-II), Sydney, NSW, Australia, 30–31 May 2005, pp. 37–44. New York: IEEE.

15.

Oliver

Boemo

Power estimations vs. power measurements in Cyclone III devices. In: 2011 VII southern conference on Programmable Logic (SPL), Córdoba, Argentina, 13–15 April 2011, pp. 87–90. New York: IEEE.

16.

Oliver

Acle

Boemo

Power estimations vs. power measurements in Spartan-6 devices. In: 2014 IX southern conference on Programmable Logic (SPL), Buenos Aires, Argentina, 5–7 November 2014, pp. 1–5. New York: IEEE.

17.

Kulkarni

Udupi

GR.

A simplified method for instruction level energy estimation for embedded system. Eur J Eng Res Sci 2017; 2(5): 56–59.S

18.

Jevtic

Carreras

Power measurement methodology for FPGA devices. IEEE Trans Instrument Measure 2011; 60(1): 237–247.

19.

Jevtic

Carreras

A complete dynamic power estimation model for data-paths in FPGA DSP designs. Integration 2012; 45(2): 172–185.

20.

Milenkovic

Jovanov

, et al. An environment for runtime power monitoring of wireless sensor network platforms. In: SSST’05 proceedings of the thirty-seventh southeastern symposium on system theory, Tuskegee, AL, 20–22 July 2005, pp. 406–410. New York: IEEE.

21.

Di Nisio

Di Noia

Carducci

CGC

, et al. High dynamic range power consumption measurement in microcontroller-based applications. IEEE Trans Instrument Measure 2016; 65(9): 1968–1976.

22.

Janković SP and Drndarević VR. Microcontroller power consumption measurement based on PSOC. In: 23rd telecommunications forum (TELFOR), Belgrade, 24–25 November 2015, pp. 673–676. New York: IEEE.

23.

Shin

Shim

Joo

, et al. Energy-monitoring tool for low-power embedded programs. IEEE Design Test Comp 2002; 19(4): 7–17.

24.

Wang

An improved instruction-level power and energy model for RISC microprocessors. PhD Thesis, University of Southampton, Southampton, 2017.

25.

Krunic

Popovic

Krunic

, et al. Energy consumption estimation for embedded applications. Elektronika ir Elektrotechnika 2016; 22: 44–49.

26.

Dâmaso

Freitas

Rosa

, et al. Evaluating the power consumption of wireless sensor network applications using models. Sensors 2013; 3: 3473–3500.

27.

Bazzaz

Salehi

Ejlali

An accurate instruction-level energy estimation model and tool for embedded systems. IEEE Trans Instrument Measure 2013; 62(7): 1927–1934.

28.

Ekstrom

Bergblomma

Linden

, et al. A bluetooth radio energy consumption model for low-duty-cycle applications. IEEE Trans Instrument Measure 2012; 61(3): 609–617.

29.

Macii

Petri

An effective power consumption measurement procedure for bluetooth wireless modules. IEEE Trans Instrument Measure 2007; 56(4): 1355–1364.

30.

Shnayder

Hempstead

Chen

, et al. Simulating the power consumption of large-scale sensor network applications. In: Proceedings of the 2nd international conference on embedded networked sensor systems (SENSYS ’04), Baltimore, MD, 3–5 November 2004, pp. 188–200. New York: ACM.

A practical microwatt-meter for electrical energy measurement in programmable devices