models for various Single-Board Computers (SBCs) such as Raspberry Pi, Asus TinkerBoard, etc. This tool stresses the CPU with variable theoretical loads, then collects data on CPU cycles and measures the power consumed by the device for each load. CPU Power Bench automatically applies linear and polynomial regression algorithms on the collected data to generate power models, as we have observed a correlation between CPU utilization and power consumption in SBCs.
Tool CPU Power Bench is an automated benchmark tool designed to accurately generate power estimation
Nowadays, single-board computers (SBCs) are widely used in various embedded systems and Internet of Things (IoT) applications due to their small size, low cost, and low power consumption. This type of device may be powered by batteries or other portable power sources. Power consumption is a critical factor in defining the overall performance and energy eficiency of these devices. Thus, measuring or estimating the power consumption of these devices is a crucial first step toward optimizing their energy usage.
Some state-of-the-art approaches involve the use of hardware meters (such as in [
Other solutions are software-based, they are more accessible, less costly, and more scalable
(such as in [
In this tool demo paper, we introduce CPU Power Bench an automated power consumption model generation tool. In Section 2, we present the energy and metrics benchmarking architecture. In Section 3, we present the relevance of our tool and its use in understanding the energy consumption of each device.
CPU Power Bench is an automated benchmark to accurately generate a power model for singleboard computers (Raspberry Pi, Asus TinkerBoard, etc.). Our tool’s primary objective is to generate power estimation models for any SBC in a simple and automated manner. The tool runs three sets of scripts that stress the CPU with loads ranging from 0 to 100% and collects real-time CPU cycles. Simultaneously, it collects the power consumed by the device from the wattmeter. Then, it uses the theoretical and practical collected data to generate linear and polynomial power estimation models.
CPUPowerBench CPULoad.csv (start_time,end_time,theoratical_U) CPUCycles.csv (time,measured_U) cpu-benchmark component cpu-load generator cpu-cycles collector model-generation component
CPU Load
CPU Cycles Power.csv (time,power)
CPU Wattmeter completeSynchedData.csv
Linear Power Polynonial Power
Estimation Model Estimation Model
Concretely, our architecture, in Figure 1, is composed of three components: a cpu-load generator, a cpu-cycle collector, and a model-generation component. In the following, we detail each of these components.
• cpu-benchmark component:
– cpu-load generator: The benchmark produces a workload by applying various loads
on the processor of the device. Our strategy is to apply a stress load on the CPU,
spanning the entire load spectrum, starting from 0% to 100% with an increase of 5%
in each step. The stress is applied for 60 seconds at each percentage interval, and
we save the workload timestamp, storing all the data in a CSV file (CPULoad.csv).
We initially used the same stress command (or stress-ng) used in the literature [
CPU Power Bench was tested to generate power estimation models for a series of Raspberry
Pis (Zero W, 1B, 1B+, 2B, 3B, 3B+, 4B, and 4B+), in addition to ASUS Tinker Board. It provided
highly accurate power estimation models for each of these devices with error rates as low
as 0.33% and up to 7.81% for linear models, and 0.3% up to 3.83% for polynomial models.
CPU Power Bench has much lower error rates when compared to state-of-the-art approaches:
14.56% for PowerPi [
CPU Power Bench is a tool designed to generate accurate power estimation models in an automated way. It integrates all the required components into a single easy-to-use tool. The tool can be used to simulate a wide range of workloads on the device’s processor, allowing researchers to analyze power consumption under diferent operating conditions. By stress testing the device, users can assess the power consumption of the device under heavy loads and determine the maximum power consumption that the device can handle. One of the key features of CPU Power Bench is its ability to implement power estimation models on a large scale of SBC devices, even if they are not connected to the internet. This feature is particularly useful for researchers and developers to test and analyze the power consumption of a variety of devices. Another advantage of CPU Power Bench is its architecture, which ensures compatibility with numerous devices.
The project leading to this publication has received funding from Excellence Initiative of Université de Pau et des Pays de l’Adour - I-Site E2S UPPA, a French ”Investissements d’Avenir” programme. The source code of the CPU Power Bench is available via CPU Power Benchmark GitHub Project (github.com/joular/cpupowerbench).