Energy-saving advantages of ARM processors

Many embedded ARM processor systems are based on battery-powered energy supply, and the power consumption of the processor is critical to the entire SoC chip, so the low power consumption advantage of the ARM processor can fully save energy consumption. In short, the current current diagram of typical power consumption does not depend on standard processes, standard sets or workloads.

EnergyBench provides several tools that can be easily combined with economical and practical hardware to measure typical power consumption using standard methods developed by EMBC. However, in addition to the processor, the specific chip design and peripheral modules integrated into the chip are also important factors affecting the chip power consumption. Although many chip vendors provide power consumption parameters in the datasheet of their products, these parameters are often not comparable. When designers try to compare different processors integrated into SoCs, it becomes very difficult to figure out what the real power consumption of the processor is. This is because vendors often use typical power consumption parameters to describe their processors, but rarely indicate the workload of the processor when making these measurements, which will be a key factor in determining energy and power parameters.

Many systems with embedded ARM processors are battery powered, so ARM is widely considered the "low power leader" in the processor space.

However, the power consumption of the system does not only depend on the processor. In addition, the specific chip design and the peripheral modules integrated into the chip will also affect the on-chip energy consumption.

Although many chip vendors provide power consumption parameters in their product datasheets, these parameters are often not comparable. This is because vendors often use typical power consumption parameters to describe their processors, but rarely indicate the workload of the processor when making these measurements.

In the past, the industry usually focused on the performance of the processor, but with the development of various test benchmarks by organizations such as EMBC (such as benchmarks for automotive, consumer electronics, and network applications), we can get a clearer understanding of the real situation inside the processor. As power consumption issues are gradually becoming the focus of attention in embedded applications, power consumption must be considered as an indicator as important as performance parameters when evaluating processors. The ultimate goal is to help system designers get the best balance between performance and power in portable applications.

EEMBC achieves this goal by developing a benchmark software utility, EnergyBench, which provides real data about energy consumption when the processor is under actual workload. Designers can use EnergyBench and EEMBC performance benchmarks together to compare the energy consumption efficiency of different processors when performing a series of standardized application tasks. When using EnergyBench to view the power consumption of a single device, it is obvious that there is no such thing as "typical power" because the average power varies greatly when running different EEMBC benchmarks. EnergyBench does not reflect the typical power of the processor, but it can be used to get typical power consumption values ​​for certain specific algorithms or applications at a specific performance level.

EnergyBench has been successfully implemented by EEMBC using the LabVIEW platform and data acquisition (DAQ) cards from National Instruments. The DAQ cards provide multiple differential measurement channels, which allow simultaneous power consumption measurements on multiple power inputs (each measurement requires capturing both voltage and current) and a trigger channel. Any ARM processor or vendor using an evaluation board or their own hardware platform only needs to modify their board-level circuitry to enable measurable power input lines and add shunt resistors.

EnergyBench can use the DAQ card to sample the voltage and trigger channels and write all the sampled results to a file. A flexible trigger mechanism enables performance benchmarking to be performed simultaneously with power measurements. This ensures that the measurements represent the power consumption when the benchmark code is running, and do not include the power consumption during the benchmark initialization or record retention phase.

When running benchmarks and power sampling, it is important to ensure that the results are reliable, repeatable, and consistent, which is particularly important for the popularity of the standard. EnergyBench uses a number of methods to ensure these goals:

1. Reliability:
Typically, to obtain accurate statistical results, sampling must be performed at a 2 X N yquist frequency or higher, or at random points.

The EnergyBench sampling module accepts a sampling frequency as input, and then the module must be called multiple times with different sampling frequencies. The sampling points generated by sampling multiple times with unaliased frequencies during the benchmark run avoid any resonance with the benchmark execution. In other words, assuming that the repeated tests of each benchmark occur at roughly periodic intervals, using frequencies that are not aliased with the period ensures that pseudo-random points are sampled in each repetition. This method is easy to implement and guarantees statistical accuracy of the results. Using this flexible method, frequencies that are aliased with the benchmark period can be easily detected, as this will lead to different results in one of the sampling frequencies. If this is detected, a new set of unaliased frequencies is selected and tested until valid results are obtained.

2. Consistency:
Because we can repeat the test as many times as needed and increase the sampling frequency arbitrarily, EnergyBench will perform multiple samplings until the average power consumption can be obtained through accurate statistical data. If the value of each repetition deviates too much, the sampling frequency can be increased to increase accuracy and reduce deviation.

3. Repeatability:
The measurement process is repeated many times to ensure measurement accuracy, and the standard deviation of the final result is calculated. Any detection deviation can also be easily obtained because each run of each benchmark produces the average power consumption of each benchmark repetition. Of course, if it is assumed that the target device provided by the supplier is universal, then the results obtained by testing the device are considered reliable. At the same time, EEMB has always had a very strict certification process to ensure that the selected chips under test are representative.
For the same reason, all semiconductor manufacturers should consider in advance the potential possibility of any small deviations in the production process, and EnergyBench's tests for many potential applications can help manufacturers gain in-depth understanding of how to choose test chips and tests, because these will ultimately affect the power consumption test results.

After a benchmark runs multiple repetitions and captures all the measured sample data, the analysis module calculates the average power consumption of the benchmark test based on this data. Based on the average power consumption value, the EEMBC power analysis module analyzes the samples captured for each test to find the minimum and maximum values ​​in the data samples.

If the variation of a specific sampling frequency is too large, the user can increase the frequency and/or the number of benchmark repetitions until the above sampling data is stable enough so that the confidence interval of the average value falls within the specified tolerance (95%), thereby ensuring the accuracy of the obtained data.

 
The final result of the EnergyBench test is the average power consumption value of the chip when running a certain benchmark test.
The EEMBC certified EnergymarkTM test result is an optional parameter that chip manufacturers can choose to provide to customers together with chip performance parameters as a reference to indicate the power consumption of the processor.

The EnergyBench specification requires a minimum device warm-up time of 30 minutes and an ambient temperature of 70°F +/- 5°F.
These basic conditions are important to ensure consistent results, as evidence suggests that as device temperature increases, power consumption also increases.

The DAQ card allows and the EnergyBench specification requires that all power rails on the processor be measured. Obviously, for low-end ARM based devices with limited pin counts, only a few power rails can be measured. High-end ARM based devices with multiple power inputs must run the benchmark multiple times to capture complete processor power consumption data. The EnergyBench Test Harness includes multiple executables that can measure one, two, or three power rails simultaneously. For processors implemented with multiple power rails (i.e., core power and I/O power), there are two methods to calculate the power consumption of each benchmark repetition. In the first method, EnergyBench uses a DAQ card to measure up to three power rails simultaneously. However, when using this method, because all channels must use the same sampling rate, the sampling rate of the DAQ card may need to be reduced so that the host can keep up with the sampling process (this is due to too many simultaneous input data). Alternatively, the power rails can be measured individually, and the sum of the power consumption values ​​for each power rail is the cumulative power consumption total.
Which method should I choose?

First, some processors have more than three power rails. Even if you measure all three power rails simultaneously, it may still be necessary to measure each power rail individually or consider using a DAQ card with more input channels.

Additionally, the sampling frequency should be proportional to the operating frequency of the processor to ensure that each benchmark samples enough data when tested repeatedly. To accommodate GHz-class processors, a higher sampling rate may be required so that the host only tests one power rail at a time.