Skip to content

Energy Consumption

Overview

This documentation provides an overview of the energy_consumption submodule and its components. The submodule is designed to handle various measurements related to energy consumption and to produce the data reported in the paper.

For the purpose of the study, we consider only the energy consumed by: (i) the circuit structure of the RKM, which can be measured analytically using Kirchhoff's laws, and (ii) the matrix multiplication on a computer, the simplest operation behind the RKM, for which we introduce specific tools in this submodule.

Tools and Limitations

To capture the energy consumed by the computer, we use the following tools:

  • pyRAPL: a Python package that uses Intel's “Running Average Power Limit” (RAPL) technology to estimate CPU power consumption. This technology is available on Intel CPUs since the Sandy Bridge generation.

  • nvidia-smi: the standard NVIDIA System Management Interface to measure GPU energy consumption.

This submodule can measure the energy consumption of both the CPU and the GPU, but it is limited to Intel CPUs and NVIDIA GPUs.

Content of the Submodule

In the src/ directory, you will find several files to analyze the energy consumption of matrix multiplication as a function of the number of hidden nodes Nh of the machine. Running these codes will produce images in out_png and store data in out_csv. We include the images and data used to produce the results reported here.

The specific files for the analysis are:

measure_consumption.py

Measures the energy consumption and power usage of matrix-vector multiplications using PyTorch on a CPU, varying the size of the vectors. It plots and saves the results as a CSV file and a PNG image.

nvidiasmi_gpu.py

Measures the energy consumption and power usage of matrix-vector multiplications using PyTorch on a GPU, varying the size of the vectors. It also measures CPU energy if enabled, and plots and saves the results as a CSV file and a PNG image.

estimate_cpu_scaling.py

Reads energy and time data from a CSV file, fits polynomial models to the data, and extrapolates to predict values for larger sizes (Nh). It plots the original and extrapolated data, saves the plot as a PNG file, and exports the estimated values to a new CSV file.

estimate_gpu_scaling.py

Reads GPU and CPU energy data from a CSV file, fits polynomial models to the time and total energy data, and extrapolates to predict values for larger sizes (Nh). It plots the original and extrapolated data, saves the plot as a PNG file, and exports the estimated values to a new CSV file.

Conclusion

The energy_consumption submodule provides tools to measure and analyze the energy consumption of matrix multiplications on both CPUs and GPUs. By using pyRAPL and nvidia-smi, we can capture detailed energy usage data, which is essential for understanding the efficiency of the RKM's operations. The provided scripts facilitate the analysis and visualization of this data, enabling further research and optimization.