As AI models grow in size and complexity, so does their carbon footprint. For example, training GPT-3, a well-known language model, used over 1,200 megawatt-hours of electricity - equivalent to the annual emissions of 276 cars. And that’s just one model. This raises a dilemma: should we continue pushing AI performance at the cost of the planet?

ABB’s researchers explored this dilemma by focusing on the process industries, where AI models are typically smaller than those used in tech giants’ labs. The good news? The research shows that these systems can be optimized to reduce emissions.

To guide engineers and decision-makers, the reserachers developed a framework mapping out the key factors influencing an AI model’s environmental impact across its entire lifecycle—from development to deployment. These include the choice of hardware, the efficiency of the training process, and even the source of electricity used.

One promising strategy is transfer learning, where a pre-trained model is adapted for a new task. This approach can drastically cut down on training time and energy use. ABB also tested various recommendations from academic literature and found that many do work in practice - such as using more efficient algorithms or training during off-peak energy hours.

AI doesn’t have to be a climate villain. With thoughtful design and smarter choices, we can build powerful models that are also environmentally responsible. Companies can take a proactive role in understanding and reducing the carbon footprint of their AI systems - not just for compliance or cost savings, but as a commitment to sustainability, making AI sustainability a reality - helping industries outrun, leaner and cleaner.

Read more in the ABB Review article here