Green AI – An Energy-Saving Paradise Lost (So Far)

February 13, 2023

Written by Alex Lawrence

“The potential of Green AI cannot be disregarded: the majority of publications show significant energy savings, up to 115%, at little or no cost in accuracy.”

So says a paper published last month by a team of academics from the Netherlands headed by Roberto Verdecchia of Vrije Universiteit Amsterdam along with TU Delft’s June Sallou and Luis Cruz. The article, “A Systematic Review of Green AI”, examines the state of work to reduce the environmental impact of machine learning and AI.

As networks grow in complexity past human capabilities to manage, professionals developing 5G Advanced and 6G are talking about networks being not just dependent on AI but potentially ‘AI-native’. A larger proportion of network power consumption than ever before will need to be given to AI processes. Meanwhile, both environmental legislation and cost pressures are driving towards energy and carbon reduction, so there is a clear need for ‘Green AI’ approaches to reconcile these otherwise conflicting demands.

Green AI Out Of The Blue

If you haven’t heard of Green AI as a topic area before then you’re in good company. It is perhaps not surprising, however. The ‘Systematic Review’ is an examination of the state of the available research to understand progress so far and where gaps in our understanding lie. Out of the 98 research papers identified in the review, the first was published in 2015 and three quarters were published after 2020, highlighting just how quickly and how recently the issue has risen to significance.

Strikingly, the review addressed questions relating only to the reduction of energy consumption by machine learning/AI algorithms. While this is an important topic, the underlying physical materials can also make a difference both to the energy consumption involved and to the sustainability of the overall system. In many cases there will be a sound argument for implementing AI on standard GPU processors because of their flexibility, and arguably AI’s green characteristics in this regard would evolve in parallel with advances in sustainable chipmaking. However, compared to tailored AI processors these are horribly inefficient, as we discovered in this conversation with Professor Merouane Debbah at TII. It is worth noting that work is ongoing to change some of this.

AI is a complex field involving several stages: an AI algorithm needs to be trained, tested, and then allowed to operate in real life – what the review authors call the ‘inference phase’. The majority of research published to date focuses on the training phase, because that tends to be obviously energy-intensive. Training an AI can soak up a similar amount of energy to a year’s household supply. An inference during the AI’s active life uses only a tiny fraction of that amount. However, as the review’s authors note:

“Given the high execution rate of the inference phase, how the energy consumed by the infrequent execution of the training phase compares to one of the highly-executed inference phase is still an open question. As a call for action, studies should be conducted by considering the energy consumed throughout the whole life cycle of AI models.”

In telecoms terms a fresh focus on the inference phase makes sense. While an AI needs training once the energy consumed to do this is, largely, down to the creator or vendor of the AI. However, the ongoing energy consumption of that AI algorithm throughout a network is a cost to the operator.

From Theory to Practice

A second, perhaps more worrying, point that the review highlights is the academic nature of the research to date, both in terms of who has conducted the research and who the papers have been written for.

Out of the 98 studies, 75 are written exclusively by academics; 20 by a combination of academia and industry stakeholders; and only three exclusively by industry.

What is perhaps more concerning, the review also identified the target readers for the literature. “The vast majority targets academic readers (85 out of 98 papers), while a much smaller portion both academic and industrial readers (8 out of 98)… Few studies are intended also for the general public (5 out of 98).”

Strikingly, there does seem to be real value to be gained from the work so far, despite it being such a young field:

“Overall, more than half of the papers which explicitly report energy saving percentages (27 papers) report a saving of at least 50%”

The review notes a fairly obvious conclusion, which is that it makes little sense for such results to remain in academic circles when the processes and software involved can be implemented today, and indeed already have been mainly in laboratory environments. A move to field trials of various techniques could yield some direct benefits in a very short time. Collaboration with companies involved in, for example, cloud and edge computing; AI algorithm development and management; and network management would open up the potential for rapid wins.

One other element the review highlights is that a minority of studies – 15 out of the 98 papers examined – make tools available to enable Green AI. While it is a shame that more do not offer tools directly designed to help, the fact is that there are fifteen tools out there to help with optimisation of Green AI. According to the review,

“The tools provided are of heterogeneous nature, and range from tools to monitor the resource efficiency of AI algorithms, to tools optimising the energy efficiency for stochastic edge inference, and implementations of convolutional neural networks optimised for energy efficient AI outside the training realm.”

As telecoms networks become increasingly dependent both on software and AI, it is exciting to see that there are already opportunities to reduce the energy costs involved, with many more directions available to follow up. The term ‘Green AI’ might trigger the cynic in some people who take this as greenwashing, but network operators can take heart that there are very real benefits available for the bottom line.

Drop us a line

Interested in contributing to Let's connect to learn more about your ideas.

Drop us a line

Interested in contributing to Let's connect to learn more about your ideas.

Recent Posts

Pin It on Pinterest

Share This