The Environmental Impact of Generative AI: A Double-Edged Sword
Written on
Chapter 1: The Rise of Generative AI
The year 2023 marked a significant surge in the use of generative AI, bringing both excitement and concern. With the rapid expansion of this technology, critical questions arise: What is the energy consumption of your preferred AI model? What amount of carbon dioxide does it generate?
Some time ago, we explored how machine learning (ML) can play a crucial role in combating climate change by enhancing the efficiency of processes that produce carbon emissions. Numerous companies and researchers are actively employing AI models for this very purpose.
Furthermore, AI's adaptability offers solutions to various challenges tied to climate change, such as combating deforestation and assisting communities in need. However, we must also acknowledge the darker side: training an AI model requires substantial electrical energy, contributing to carbon dioxide emissions and, consequently, global warming.
Understanding the energy consumption and environmental impact of the AI industry is vital for developing strategies to mitigate its effects. The question is more pressing than it appears, especially considering the widespread adoption of both text and image-generating AI in 2023. The energy costs associated with AI extend beyond training to include inference as well.
As the IT sector's energy consumption escalates—particularly in data and cloud services—recent studies indicate that data center usage has doubled from 2017 to 2021, with annual growth rates between 20% and 40%. By 2022, these emissions accounted for approximately 1% of global energy-related greenhouse gas emissions.
Indeed, this data can also be harnessed for artificial intelligence. Storing this information requires electricity, and generating greenhouse gases is an inherent part of training models.
Most research has concentrated on the energy costs of training models, while inference often gets overlooked. It is true that training consumes significantly more energy than inference, as there are no parameter updates during inference. However, the scale at which models are deployed today is considerable, given the millions of daily users for platforms like Google Translate, ChatGPT, and Midjourney. According to Amazon, 80% of ML cloud computing demand is attributed to inference.
Section 1.1: Measuring AI Inference Consumption
As demand for AI services surges in 2023, how much energy does inference consume for each model? This is a challenging question due to varying architectures and tasks. Recent research sheds light on this matter.
The study began by analyzing 10 prevalent ML tasks across five modalities:
- Text-to-category (text classification, token classification, extractive question answering)
- Text-to-text (masked language modeling, text generation, summarization)
- Image-to-category (image classification, object detection)
- Image-to-text (image captioning)
- Text-to-image (image generation)
After selecting 88 models trained for these tasks, researchers conducted multiple inference tests to ensure statistical significance, utilizing the Code Carbon package to quantify greenhouse gas emissions.
The analysis revealed that classification tasks, whether involving images or text, require significantly less energy compared to text-generating tasks. Notably, multimodal tasks, such as image captioning and generation, are the most energy-intensive. Overall, the findings indicate that text-related tasks are generally less costly than those involving images.
For context, charging an average smartphone demands around 0.012 kWh of energy. The most efficient text generation model consumes energy equivalent to 16% of a full smartphone charge for 1,000 inferences. In contrast, the least efficient image generation model can use up to 950 smartphone charges (11.49 kWh), or nearly one charge per generated image.
While these figures represent averages, actual consumption varies with model choice and text length. Furthermore, larger models—those with more parameters—tend to consume more energy. The study found a direct correlation between the number of parameters in a model and its energy consumption.
The research also examined multipurpose models like BLOOM. By analyzing different models across several datasets for two discriminative tasks (sentiment analysis and question answering), it was found that multipurpose models consume significantly more than single-task models.
The authors attribute this to the model structure: a task-specific model can efficiently predict a class using a binary classification, while a multipurpose model requires more energy for predictions across its entire vocabulary.
Section 1.2: The Inference vs. Training Trade-off
An essential question for AI practitioners and policymakers is determining when the energy costs of model inference become comparable to those of model training and fine-tuning. Tracking the lifecycle of a model post-training is complex, and comprehensive data on training energy consumption is often lacking.
The authors presented data showing the number of inference queries needed to equal training energy consumption. While these numbers may seem vast, they must be contextualized. For instance, ChatGPT reportedly has up to 10 million daily users, with its webpage receiving 1.7 billion visits in October 2023. Given the frequency of user queries, inference energy consumption can surpass training energy expenditure within days or even weeks.
Moreover, many models are readily downloadable, leading to underestimations of their energy impact. Additionally, many models undergo fine-tuning before deployment, which incurs further emissions.
The first video titled Artificial Intelligence for Climate Change Mitigation delves into the implications of generative AI on environmental sustainability and how it can be harnessed to combat climate issues.
Chapter 2: The Cost of Generative AI
The analysis of 88 models for various tasks highlighted that the energy costs associated with generative AI far exceed those of other services. A noteworthy trend is the industry's shift toward multipurpose models. Major companies like Google and Microsoft now primarily utilize foundation models, such as PaLM and GPT-4, following the success of GPT-3 and advancements in few-shot learning.
While there are advantages to deploying generative zero-shot models for their versatility, there is little justification for their use in scenarios where tasks are clearly defined, such as web searching, given their substantial energy demands.
In other words, for certain tasks, smaller models can perform just as effectively as their larger counterparts. Techniques like knowledge distillation may prove beneficial in such cases, allowing smaller models to replace larger ones in most situations.
The authors acknowledge that their study is not exhaustive of all deployment scenarios and limitations, but they aim to provide foundational data points for future comparisons and assessments of various models.
As intriguing as these findings are, they represent just the beginning of understanding the environmental impact of AI models. Unfortunately, many models currently in use lack detailed energy consumption data.
What are your thoughts on this issue? Should there be regulations addressing AI emissions? Share your opinions in the comments.
If you found this discussion enlightening, consider exploring my other articles or connecting with me on LinkedIn. You can also access my repository for weekly updates on ML and AI news. I welcome collaborations and projects, so feel free to reach out.
Here is the link to my GitHub repository, where I compile code and resources related to machine learning, artificial intelligence, and more.
The second video titled AI and the Energy Required to Power It Fuel New Climate Concerns discusses the implications of AI's energy demands in the context of climate change, raising important questions about sustainability and future practices.