To offer typical numbers, the team that made the analysis followed the requests and the equipment that served them for a period of 24 hours, as well as the inactivity time for this material. This gives them energy by request, which differs according to the model used. For each day, they identify the median prompt and use it to calculate the environmental impact.
Descent
Using these estimates, they note that the impact of an individual text request is quite low. “We believe that the median text prompt of Gemini applications uses 0.24 Watthers of energy, emits 0.03 grams of carbon dioxide equivalent (GCO2E) and consumes 0.26 millilier (or approximately five drops) of water”, they conclude. To put this in context, they believe that energy consumption is similar to around nine seconds of visualization of television.
The bad news is that the volume of requests is undoubtedly very high. The company chose to perform an AI operation with each research request, a request for calculation that simply did not exist a few years ago. Thus, although the individual impact is low, the cumulative cost is probably considerable.
The good news? Barely a year ago, it would have been good, much worse.
Part of this is just circumstances. With the solar energy boom in the United States and elsewhere, it has become easier for Google to organize renewable power. Consequently, carbon emissions per unit of energy consumed have seen a reduction of 1.4x in the past year. But the biggest victories were on the software side, where different approaches have led to a reduction of 33x of the energy consumed by invite.
Most energy consumption to serve AI requests comes from the time spent in personalized accelerator fleas.
Credit: Elsworth, and. al.
The Google team describes a number of optimizations that the company has made that contribute to it. One is an approach called a mixture of experts, which consists in determining how to activate only the part of an AI model necessary to manage specific requests, which can delete the calculation needs of a factor of 10 to 100. They have developed a certain number of compact versions of their main model, which also reduces computer load. The management of the data center also plays a role, because the company can ensure that any active material is fully used, while allowing the rest to stay in a low power state.
