Is AI doomed to be an energy drain?

Gen­er­at­ive AI mod­els, such as OpenAI’s GPT‑4, are presen­ted as all-pur­pose tools. They include a sig­ni­fic­ant num­ber of para­met­ers – now num­ber­ing in the bil­lions – that enable them to per­form any type of task. This plur­al­ity of uses, which leads to com­plex­ity issues, makes these mod­els “in need of optim­isa­tion”, accord­ing to Enzo Tartagli­one, research­er and seni­or lec­turer at Télé­com Par­is (IP Par­is). This com­plex­ity also implies con­sid­er­able energy con­sump­tion.

“Even for an extremely simple query, AI will tend to use all the resources at its dis­pos­al to respond, without exclud­ing those that are not use­ful. This leads to energy waste, and it is really some­thing we need to optim­ise.” This energy con­sump­tion, estim­ated at around 2% of glob­al con­sump­tion in 2024, is driv­ing research towards an altern­at­ive approach: energy efficiency.

OpenAI has made an extremely resource-intens­ive lan­guage mod­el avail­able via serv­ers. This obser­va­tion led research­ers to dis­tin­guish between the resource con­sump­tion of mod­el train­ing and that of its infer­ences, i.e. its use. Although the energy con­sump­tion of train­ing is sig­ni­fic­ant – approx­im­ately 1,287 MWh for GPT‑3, and between 10,000 and 30,000 MWh estim­ated for GPT‑4 – its impact is one-off. The impact of infer­ence, on the oth­er hand, depends on the num­ber of users, which is grow­ing con­stantly. A 2021 study¹ estim­ates that “between 80 and 90% of machine learn­ing work­load at NVIDIA comes from infer­ence. Amazon Web Ser­vices estim­ates that 90% of cloud demand for machine learn­ing is inference.”

Some research­ers believe that a bal­ance needs to be found between a model’s energy con­sump­tion and the task it is required to per­form. If a mod­el is used to dis­cov­er a drug or advance research – which it is cap­able of doing – the car­bon foot­print will be easi­er to accept. How­ever, today, these mod­els can be used for all kinds of tasks, mak­ing mil­lions of infer­ences through the vari­ous requests made of them at the same time. 

At Télé­com Par­is (IP Par­is), the Data Sci­ence and Arti­fi­cial Intel­li­gence for Digit­al­ised Industry and Ser­vices chair focuses on sev­er­al chal­lenges: how to recon­cile the rise of AI and its energy con­straints without sac­ri­fi­cing its poten­tial. “We are explor­ing issues of frugal­ity (editor’s note: seek­ing to “do more with less” and with great­er respect for the envir­on­ment), but also sus­tain­ab­il­ity (editor’s note: meet­ing the needs of present gen­er­a­tions without com­prom­ising those of future gen­er­a­tions),” adds Enzo Tartagli­one. “There is a real ques­tion in the choice of applic­a­tions, because we can­not simply point the fin­ger at AI and say it is a bad thing. Togeth­er with col­leagues, we are start­ing work on gen­er­at­ing mater­i­als for stor­ing hydro­gen. This is also some­thing that AI can offer as a solution.”

Espe­cially since the mod­els we can all use on our mobile phones require com­mu­nic­a­tion with a serv­er. “We need to be aware of the cost of trans­port­ing inform­a­tion, espe­cially when it’s bid­irec­tion­al,” insists the research­er. “There is there­fore a great neces­sity to design mod­els that can be used loc­ally, lim­it­ing the need for com­mu­nic­a­tion with an extern­al serv­er. How­ever, we are talk­ing about mod­els with bil­lions of dif­fer­ent para­met­ers. This requires too much memory for your smart­phone to do without the internet.”

There are there­fore sev­er­al dimen­sions to energy effi­ciency. It is not enough to reduce the num­ber of para­met­ers required for mod­el cal­cu­la­tions, wheth­er dur­ing train­ing or infer­ence, as the Deep­Seek mod­el has done. Action must also be taken on the train­ing data and the data that forms the model’s know­ledge. One solu­tion that stands out is Mis­tral, an open-source French lan­guage mod­el. There is value in divid­ing the main mod­el into a mul­ti­tude of experts that can be activ­ated accord­ing to the task at hand. This is one of the approaches pro­posed for optim­ising these mod­els: dis­tin­guish­ing them by spe­ci­al­ity. “The goal is to take pre-trained mod­els and imple­ment strategies to adapt them with as few para­met­ers as pos­sible to dif­fer­ent, very spe­cif­ic sub­tasks,” explains Enzo Tartagli­one. “This not only avoids the impact of retrain­ing, but also greatly improves energy and performance.”

With more spe­cial­ised mod­els, the amount of know­ledge required to achieve this will also require less data to be acquired. This type of mod­el could there­fore act loc­ally and com­mu­nic­ate with serv­ers in a much more stream­lined way. After all, AI is still a rel­at­ively recent innov­a­tion. And, like most tech­no­lo­gic­al innov­a­tions, it should logic­ally fol­low the same path of optim­isa­tion as its pre­de­cessors. Ulti­mately, frugal AI is more of a move­ment in AI research than a field in its own right. In fact, com­puter sci­ence has always sought to design sys­tems that optim­ise resources and lim­it unne­ces­sary cal­cu­la­tions – frugal­ity is there­fore a nat­ur­al con­tinu­ation of this logic of effi­ciency. Per­haps in the same way that com­puters and phones have become portable?

In any case, the appeal of frugal­ity, in addi­tion to being envir­on­ment­al, is also eco­nom­ic for the vari­ous play­ers devel­op­ing this type of tool. This does mean, how­ever, that there is still a risk of a rebound effect: more wide­spread use due to lower devel­op­ment costs would greatly reduce the envir­on­ment­al bene­fits. How­ever, this approach will undoubtedly not be the only solu­tion to the energy abyss that AI rep­res­ents; address­ing the issue of sus­tain­ab­il­ity will also be essential…

Pablo Andres