4.4.1. Energy Consumption

In [120], Pinto et al. define energy consumption as an accumulation of power dissipation over time:

$$\text{Energy Consumption} = P \times t \tag{1}$$

Note that Energy Consumption is measured in joules and Power ( *P*) is measured in watts. The relationship between these two quantities can be easily interpreted through an example: if a software program takes 5 s to execute and dissipates 5 watts, it consumes 25 joules of energy. In the case of software energy consumption, attention must be paid not only to the software under execution, but also to the hardware that executes the software, the environmental context of execution, and its duration.

## 4.4.2. CO2 Emissions

In [121], Strubell et al. presented a study that focused on the estimation of the financial and environmental cost of training a variety of recently successful NN models. To estimate CO2 emissions (CO2e), they proposed a simple method based on the multiplication of the energy consumption with the average produced CO2. After measuring the CO2e for several models using different hardware, they concluded that the CO2 required for training one model can range from 12 kg up to 284 t. Note that this CO2e footprint is highly significant when compared with the world average CO2 emissions per capita, whose estimate was 4.56 t in 2016 [122]. Moreover, they evaluated the cost of training these models in the cloud, which raised from USD 41 up to USD 3,201,722, respectively.

#### *4.5. Measuring Edge-AI Performance*

Although this article focuses on Edge-AI sustainability, there are other factors that should be considered during the evaluation of the performance of an Edge-AI system. Specifically, four main metrics are often used for the performance evaluation of AI algorithms [123]: accuracy, memory bandwidth, energy efficiency, and execution time.
