While AI undeniably opens up fascinating possibilities, its enormous energy consumption casts a long shadow.
First, we must acknowledge that accurately quantifying AI energy consumption is an immense challenge. Several factors contribute to this complexity:
1. Lack of transparencyMany AI companies treat their energy consumption as a trade secret, making independent verification difficult.
2. Rapid technological developmentThe AI landscape is changing at a breathtaking pace. Energy consumption data is often outdated before it is even published.
3. System complexityAI systems consist of numerous components – from data centers to end devices – each with different energy profiles.
4. Lack of standardizationThere are no standardized methods for measuring and reporting AI energy consumption, which makes comparisons difficult.
5. Indirect effectsAI can lead to energy savings in many areas that are difficult to quantify.
Despite these challenges, some remarkable insights can be gained that give us an idea of the scale of the problem.
To make the scale of AI energy consumption tangible, we look at two insightful comparisons:
SmartphoneAn average smartphone consumes approximately 2-6 watt-hours (Wh) per day when actively used.
HouseholdIn Germany, the average daily electricity consumption of a household is 8-10 kilowatt hours (kWh). In the USA, this figure is significantly higher at around 30 kWh per day.
AI models compared
- Training the GPT-3 language model required a staggering 1,287 megawatt hours (MWh). This corresponds to the annual consumption of approximately... 420 average German households!
- A large convolutional neural network (CNN) for medical image analysis consumed up to 263,000 kWh for training – enough to 88 German households for one year to supply with electricity
Then the data centers:
- Currently consume data centers that provide the infrastructure for AI and other digital technologies, 4-5% of the world's electricity.
- Forecasts According to this, the AI-related electricity consumption of data centers could increase in the coming years. up to 30% increase.
These figures illustrate the enormous energy intensity of AI systems, especially during the training phase of complex models.
Differences in energy consumption
It is important to understand that the energy requirements of AI systems vary greatly. Several factors influence consumption:
1. Generative vs. discriminative models:
Generative models like ChatGPT or Midjourney, which generate new content, are generally much more energy-intensive. In contrast, discriminatory models, used for analysis and classification, often require less energy.
The larger and more complex an AI model, the higher the energy requirement for training and operation. For example, training GPT-3 consumed 1,287 MWh, which is equivalent to the energy consumption of a nuclear power plant in one hour.
Graphics processing units (GPUs) are energy-intensive. An NVIDIA Tesla V100 GPU consumes approximately 300 watts per hour. Newer generations, such as specialized AI chips like FPGAs, promise significant energy savings.
Fun Fact: AI for language translation or image recognition often requires more training than models for industrial process optimization.
Direct impacts on the environment:
If the figures are to be believed, the entire IT industry, including AI, is responsible for 2-4% of global greenhouse gas emissions – similar to global air traffic. And the training of GPT-3 alone caused CO2 emissions of 552 tons, equivalent to 700 round-trip flights between New York and San Francisco.
A hypothetical switch of all Google searches to AI-based processing would require an annual electricity consumption of 29.3 TWh – as much as the whole of Ireland consumes.
The underestimated water consumption
The often overlooked aspect of water consumption is also worrying:
- 700,000 liters of cooling water were needed for the training of GPT-3.
- Microsoft's global water consumption increased by 34% to 6.4 billion liters in 2022, while Google's increased by 20% to 25 billion liters – presumably due to the increased use of AI.
- A single request to a generative AI system like ChatGPT consumes approximately 500 ml of water.
These figures illustrate that the ecological footprint of AI goes far beyond mere electricity consumption, and we must ask ourselves: how long can we really afford this?
The future will be better, in part.
To reduce the energy consumption of artificial intelligence (AI), several strategies exist. Firstly, work is underway on developing energy-efficient hardware, such as special chips and memristors, which enable more efficient data processing.
Secondly, algorithms and software can be optimized, for example by reducing the accuracy of calculations. Thirdly, AI can be combined with other methods, such as scientific models and human knowledge, to increase its efficiency. Fourthly, it is important that trained AI models run energy-efficiently to enable their real-time application. Fifthly, the energy efficiency of AI models can be measured and improved, for example using the EC-NAS benchmark or by adhering to guidelines for developing energy-efficient AI models.
And what is the government doing? Well, certain governments are promoting research and development of environmentally friendly AI applications. Experts emphasize the importance of energy efficiency in AI development to limit its ecological footprint and enable sustainable use.
Bei Fragen? #fragRoger
Would you like to know more? I’d be happy to visit you—whether at your home, your company, your ERFA group, or your association—and help out with a workshop or a presentation.
Disclaimer: This article was written using AI based on my own research, improved with DeepL Write, and summarized and simplified by Mistral. This article is purely educational and makes no claim to completeness.
PS:
Do you already know my AI Good News?
And there's also the AI Bad News.
Quellen:
