The study, conducted by a team of researchers led by Maximilian Dauner at Hochschule München University of Applied Sciences in Germany, was published on June 19 in the journal Frontiers in Communication. The research highlights the environmental trade-offs associated with the quest for more accurate AI models.
The study focused on the carbon emissions produced by different large language models (LLMs) when answering a series of questions across various topics, including algebra and philosophy. The researchers used the Perun framework to analyze the performance and energy consumption of 14 LLMs, ranging from seven to 72 billion parameters, on an NVIDIA A100 GPU. They then converted the energy usage into CO? emissions, assuming each kilowatt-hour of energy produces 480 grams of CO?.
It found that reasoning models, which use a "chain-of-thought" approach to break down complex problems into smaller, logical steps, were significantly more energy-intensive than their more concise counterparts. On average, reasoning models generated 543.5 tokens per question, compared to just 37.7 tokens for concise models. This disparity in token generation led to a substantial increase in CO? emissions....<<<Read More>>>...
