AI-Driven Procurement: Revolutionizing the Industry and Offering Competitive Advantage

Making Generative AI Environmentally Sustainable: Insights and Strategies

In a study from Ajay Kumar from EMLYON Business School (France) and Tom Davenport of

Babson College, the environmental impact of generative AI Large Language Models (LLMs) like ChatGPT, BERT, LaMDA, GPT-3, DALL-E-2, MidJourney, and Stable Diffusion is brought into sharp focus. Despite their remarkable capabilities, these tools come with significant hidden environmental costs, primarily due to their high energy consumption during development and use.

There are considerable energy demands of data centers, which are essential for storing and managing the vast data required by these AI models. These centers contribute to 2–3% of global greenhouse gas emissions, with energy requirements for data storage and management doubling every two years. The energy-intensive nature of GPU chips, which are crucial for running these AI models, is also emphasized.

A crucial point raised is the carbon footprint associated with different stages of AI model development, including training, inference, and the production of necessary computing hardware. For example, training a model like GPT-3 or Google’s PaLM can generate about 300 tons of CO2, equivalent to the annual CO2 output of several average North Americans.

The carbon footprint of generative AI models is influenced by three main factors: the energy used in training the model, the energy consumed during inference (processing new input data), and the energy required to produce the computing hardware and cloud data center capabilities. Larger models with more parameters, like GPT-3, require significantly more energy. For instance, training a model like GPT-4 can use about 300 tons of CO2, much higher than the average individual’s annual carbon footprint. While inference uses less energy per session, it adds up over many uses. Fine-tuning existing models is less energy-intensive than initial training, but widespread use can still lead to high energy consumption. The manufacturing of computers and servers for AI also contributes substantially to the energy footprint.

To address these environmental challenges, the authors propose several strategies to make AI greener:

1. Utilize Existing Models: Instead of creating new large models, companies should leverage existing ones to save energy.
2. Fine-Tune Existing Models: Refining pre-trained models for specific needs is less energy-intensive than training new ones from scratch.
3. Adopt Energy-Conserving Computational Methods: Using methods like TinyML can significantly reduce power consumption.
4. Selective Model Usage: Employ large models only when they offer substantial value, avoiding unnecessary energy consumption.
5. Conscious Application of AI: Prioritize AI use in areas where it offers significant benefits, like healthcare and natural hazard prediction, rather than in less impactful applications.
6. Evaluate Energy Sources: Choose cloud providers or data centers that use environmentally friendly power sources.
7. Reuse and Recycle Models and Resources: Opt for open-source models and recycle tech materials.
8. Monitor Carbon Footprint: Implement carbon monitoring for AI activities to make informed decisions about their environmental impact.

The article concludes by stressing the importance of including ecological considerations in the discourse around generative AI, alongside ethical, legal, and other concerns. Ensuring the sustainability of these technologies is vital for their long-term viability and for maintaining a habitable planet for future debates and developments in AI.

Let’s prioritize sustainability in AI innovations for a greener future! 🌍💡 At NeuralPit, we believe sustainability should be at the forefront of any innovation. By fine-tuning existing Large Language Models to suit specific purposes, we not only improve the performance of the model outputs, helping businesses in maximizing benefits, but also minimize environmental impacts.

Read more about AI applications for business here.