UNPICKING THE NUMBERS WITH GREEN ALGORITHMS
It’s hard to estimate exactly how much energy is chewed through by various AI, driverless systems and so on – not least because of hidden factors such as the type of hardware, how long it’s used, and the energy source. Loïc Lannelongue discovered this the hard way while trying to uncover the carbon emissions of his own work in computational biology.
In 2020, he saw the paper describing the carbon emissions of an AI model as being on par with manufacturing and lifetime use of five cars. Though he uses machine learning in his work on molecular simulations, most computations are run by what he calls “huge algorithms” and he wondered about their carbon footprint. “I thought it would be a two-week break from my PhD, but three years later, here I am, so that didn’t quite work out,” Lannelongue said.
There were simply no available calculators – so Lannelongue and his colleagues built one. The aim was to make the Green Algorithms tool work across research areas or industry, regardless of how much information was available. More data helps improve the accuracy, but all that’s required is how long the model runs – “usually something you know because you’re waiting for the results” – and how much resources are used, be it one GPU or ten. Again, scientists tend to know that because they generally need to request access.
The other key piece of information is where you’re located, as the local energy mix has a massive impact. The carbon intensity of energy in France, which is to say how much carbon is produced per kilowatt hour, is below 50g, versus more than 800g in coal-loving Australia. “You can fine-tune it with more information to make it very accurate,” Lannelongue added, such as how much memory is available and what hardware you’re running on. “You don’t need all these tiny details, as the default values will give you the right order of magnitudes, which is the goal here.”
Papers published via his lab now include an environmental impact statement including the carbon footprint, and he’s seen more and more people using his Green Algorithms tool to do the same, with hundreds of calculations run weekly. “It’s nice to see people picking up on the issue and being mindful of that,” he said. You can find out more yourself at greenalgorithms.org.