Environmental impact of ML/AI

Instruction and application

Complete

ML delivers value, but training and serving models consume electricity, cooling and hardware. As adoption grows, carbon and water footprint become first-class engineering concerns—not optional polish.

The carbon footprint of ML

Large training runs can be energy intensive; reported estimates vary by hardware, region and methodology. Treat public numbers as order-of-magnitude guides, then measure your own workloads with cloud tooling and experiment logs. Classic reference: Strubell et al., 2019.

Drivers of energy use

✨

Accelerators:

GPUs/TPUs and interconnects draw sustained high power; cooling adds overhead.

✨

Grid mix:

the same FLOPs have different emissions depending on region and time of day.

✨

Inefficient workflows:

redundant sweeps, oversized models and repeated preprocessing multiply cost.

Hidden inefficiencies

✨

Hyperparameter brute force

without search strategy or early stopping.

✨

Over-processing data

when a smaller, curated dataset would generalise as well.

✨

Oversized architectures

when a smaller model or distilled variant meets the SLA.

Efficiency levers

✨

Model-side:

pruning, quantisation, distillation, efficient architectures (MobileNet, EfficientNet), transfer learning instead of training from scratch.

✨

Systems-side:

distributed frameworks (Ray), caching/dedup, autoscaling, carbon-aware scheduling where available (example discussion).

✨

Software-side:

green engineering habits from the Green Software Foundation—avoid wasted cycles, measure what you optimise.

Example: radiology pipeline optimisation

Researchers reported large energy savings by tightening architecture and dataset handling while preserving diagnostic accuracy—see related work on efficient medical imaging ML.

Pause and think

What is one training habit you could change next sprint to cut wasted GPU hours without lowering the acceptance criteria that matter?

Action item: Quiz

A machine learning engineer at a startup is developing a new image recognition system. They are considering using a very large, pre-trained model to achieve high accuracy. However, they are also concerned about the environmental impact of their work, as their company has committed to reducing its carbon footprint. Which factor should they prioritise?

A) Deploying the model as quickly as possible to gain a competitive advantage.B) Selecting the largest, most complex model architecture available.C) Evaluating the energy consumption of training the model and exploring options for more efficient architectures or transfer learning.D) Using the maximum amount of computational resources, regardless of energy consumption, to ensure the highest possible accuracy.

A product manager is planning the development of a new AI-powered customer service chatbot. They have the option of deploying the chatbot on a cloud computing platform that primarily uses renewable energy or one that relies on fossil fuels. Which approach aligns with the principles of sustainable AI development?

A) Choosing the cloud provider that offers the lowest upfront cost, regardless of its energy sources.B) Selecting the cloud provider with the most comprehensive suite of AI tools, even if it uses fossil fuels.C) Prioritising the cloud provider that uses renewable energy sources, even if it involves slightly higher costs.D) Avoiding the use of cloud-based solutions altogether and developing a less scalable, on-premises solution.

A compliance officer is tasked with developing guidelines for their organisation's use of AI. They need to include a metric for evaluating the sustainability of AI deployments. Which metric would be most relevant to include?

A) The model's accuracy and F1-score.B) The number of AI models deployed across the organisation.C) The amount of electricity consumed during the training and operation of AI models.D) The diversity of the AI development team.

In progress