With the recent boom in generative AI, the environmental challenges posed by these technologies have emerged as a key issue to be addressed. Studies have examined the energy consumption, greenhouse gas emissions, and, to a lesser extent, water consumption generated by the infrastructure used by these technologies. Until recently, the environmental costs associated with the manufacture of graphics cards, an essential component in the operation of this infrastructure, had been little studied, and even less so from a multi-criteria perspective.

This study, conducted as part of a collaboration between the ALT IMPACT program and the academic world, aims to fill this gap by providing open data on the environmental impacts of eight graphics cards that can be used in an artificial intelligence context. This study is notable for its exhaustive data collection, which for some cards includes analysis of their elemental composition. Based on these analyses, it also proposes a parametric model that can be used to adapt the assessment to other graphics cards not analyzed as part of this project.

The use phase of these cards, particularly when used for artificial intelligence (training or inference), is by far the most impactful phase. Nevertheless, embedded impacts (raw material extraction and processing, manufacturing, distribution, and end of life) remain significant, particularly in terms of criteria related to raw material extraction (resource depletion, human health). The results show that graphics cards have embedded impacts on key environmental indicators linked to digitalization, including global warming potential (GWP), abiotic resource consumption (ADPe) and fossil resource consumption (ADPf), human resource consumption (PM), and water consumption (WU), which remain consistent with the impacts of other digital equipment. In terms of global warming potential (GWP), the embedded impacts of the cards analyzed range from 60 to 150 kgCO2eq.

The manufacturing phase is by far the main contributor to embedded impacts, particularly for indicators related to abiotic resource consumption, while distribution and end-of-life play a more moderate role. At the component level, the main chip (GPU and VRAM) dominates most impacts, followed by the heat sink (if the card is equipped with one).

Beyond these findings, this study illustrates the crucial importance of primary data from disassembly and detailed composition analyses in improving the reliability of life cycle assessments (LCA). While carbon footprint estimates remain broadly stable for the A100 SXM, which was the subject of a detailed analysis (difference of less than 2%), other impact categories—particularly those related to the use of mineral and metal resources—vary much more significantly (up to +33%).