Skip to main content

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​

Question 1 of 2
Which change typically reduces training energy first?
  • A. Fewer redundant training runs via better experiment tracking and hyperparameter search discipline
  • B. Increasing batch size without limit on every job
  • C. Disabling logging entirely
  • D. Running the same experiment repeatedly without version control
Correct Answer: A

Feedback: Eliminating duplicate runs is often the fastest win before deeper architectural changes.

Why does electricity grid mix matter for “green AI”?
  • A. It changes Python syntax version
  • B. It only affects on-prem laptops
  • C. It determines GPU memory size
  • D. The same compute can produce different emissions depending on how electricity is generated
Correct Answer: D

Feedback: Marginal emissions vary by region and time; location-aware scheduling is a real lever.