The Hidden Carbon Footprint of AI Models: GPU-Aware Carbon Modeling

The rapid growth of AI technology has sparked transformative innovations but also increased carbon emissions. Recent research found that computer systems' carbon emissions are shifting from operational carbon to embodied carbon, but they did not fully capture the rapidly evolving AI landscape. Most recent research focused on operational carbon, neglecting the long-term environmental impact of embodied carbon. We found two gaps that persist in recent research. First, current carbon modeling focused on Central Processing Units (CPUs), neglecting the carbon modeling of Graphical Processing Units (GPUs). Second, it focused on primary components, neglecting significant contributions from peripheral components to the embodied carbon. These neglects are critical given AI's reliance on GPUs. Thus, modeling GPUs' carbon emissions can provide comprehensive carbon quantification of AI models. To address the aforementioned gaps, we propose a GPU-aware carbon modeling tool that estimates GPU's embodied carbon by considering primary and peripheral components. Using this tool, we quantify the embodied carbon emissions of both GPT-3 and GPT-4. Our findings reveal that the embodied carbon from GPUs constitutes 0.77% of GPT-3's and 2.18% of GPT-4's reported emissions. We predict that hardware production's embodied carbon will continue to rise due to the increasing reliance on GPUs in AI computing. Our research reveals the hidden effects of GPUs' carbon emissions, helping hardware designers and datacenter operators reduce the long-term environmental impact of AI.