Measurements and Analysis of Power Consumption and Execution Time for AI Accelerators

Kurzfassung

The Artificial Intelligence (AI) field is subject to regular update on many of its aspects. In the last years field of AI has gathered interest not only in in- dustry but also electronics consumers. Our prevalent technology in this field is called Artificial Neural Networks (NN). NNs are very versatile algorithm used for many applications and currently in progress to be implemented in embedded systems. This algorithm is a sequential execution of multiple lay- ers described by a defined function. These functions can be expressed in the form of matrix operations, which require a lot of computation power. One of the current solutions is to rely on the computing capacity of a cloud to perform the inference and give the result to the consumer systems (named the edge). This approach introduces communication and privacy concern. Integrating AI systems in the edge requires development of powerful solu- tions under very strict resource constraints. To achieve performance execution of NNs on embedded systems, a char- acterization of time and power consumption of the NN is required. The current solution is to directly deploy the NNs on the target hardware to perform measurements on it. This approach is time consuming because it requires the setup to perform the measurement. Another approach is to use models to predict the behavior of NN on a given hardware. To create an accurate model, it requires a setup able to achieve a good accuracy in the measurement probed. This data is obtained through measurement, charac- terization and profiling. In this context, this document will focus on the measurement of power consumption and the execution time.