CICO \(_{2}\) e: A Compute Carbon Footprint Estimation Tool Based on Time Series Data

In recent years, multiple research papers and tools were published that addressed the interest of estimating the carbon footprint of computational activities. Generally, the presented carbon footprint estimation methods calculate the product of the compute hardware’s energy consumption and a factor expressing the emissions for the corresponding energy produced in the region where the compute hardware is located. However, there are three open issues. First, the methods for determining the energy consumption are inaccurate or lack an evaluation of the accuracy. Second, most tools use as carbon intensity a static long-term average, e.g., over a year, that the tool authors gathered once, but since some regions have a carbon intensity varying each day and throughout the year, the accuracy of using a static carbon intensity is unclear and requires an evaluation. Third, most tools enable estimates for a single computer or a homogeneous set of computers only, excluding the easy carbon footprint estimation for scenarios with a heterogeneous set of compute hardware. In this paper, we make three contributions. First, we analyze the evaluation gap regarding methods for determining the compute hardware’s power consumption. Second, using example cases, we show that estimating a carbon footprint using a static long-term average carbon intensity compared to hourly carbon intensity time series data can lead to large errors; in the worst case among the example, the error was 325.8%. Third, we present a tool assisting with carbon footprint estimates using time series carbon intensity data and also supporting heterogeneous compute hardware.