What fuels AI? Follow the lifecycle of a machine-learning model
Each stage of an AI model’s lifecycle involves a huge amount of consumption, especially water and power.
What are the stages in the life of an AI model?
It starts with inception: This involves defining the kinds of problems that the AI model will solve, and setting associated goals.
Then comes design and development: At this stage, creators must build the algorithm, prepare the datasets, build the model in the form of clusters of computing capability. All this needs electricity (vast amounts of it) and water, to keep the energy-intensive data centres from overheating.
Verification: Extensive tests area run on the clusters, to assess whether the AI is behaving as it was meant to. These trial runs also require vast stores of power and water.
Deployment: The trained model is stored and replicated across servers spread out around the world, to distribute the load and make its operations smoother and faster.
Operation and monitoring: Once live, an AI model processes every request via powerful servers.
Re-validation: As more data is gathered and social contexts as well as user needs evolve, the models are periodically re-tested and fine-tuned.
Retirement: This happens when an AI system becomes outdated. Its models and servers are decommissioned, scrapped or replaced. This involves dumping tonnes of chips, circuits and hardware that will degrade and poison soil or water over centuries, unless disposed off securely and responsibly.
Continuous learning: Through its lifecycle, as the AI model adapts by “learning” from new data and feedback, it requires additional computation and storage, all of which comes at an energy cost.
Could AI use something other than water?
In a first-of-its-kind study published in the journal Communications of the ACM in June, researchers from the Universities of California Riverside, Texas and Houston traced how AI uses water, and whether smarter, more sustainable usage is possible.
First, the usage. Data centres draw on power plants that use large cooling towers.
Then there are the servers at the data centres — thousands of servers at each such centre, and thousands of centres — that need to be kept cool as semiconductors overheat during use. These are typically connected to cooling towers too.
When climate conditions are appropriate, data centres may also use ambient air outside the facility to directly dispel heat, the study notes. However, water evaporation would still be needed, especially when the outside air was hot. Water would also be needed to achieve ideal humidity levels for the machines, when the outside air was too dry.
While using ambient air would be less resource-intensive than using water, it can be difficult to practice this in very hot areas or busy districts with multiple data centres. Already, the sound pollution from the drone of large data centres has caused residents to protest, in parts of the US.
The best answer remains smaller models and more considered use of this technology. Click here for more on this.
