This is how much water Gemini consumes every time you ask him a question
Google published a study to measure the impact of artificial intelligence on the environment. The company decided to make the process transparent by providing an estimate of the energy consumption and emissions generated by a text message. According to Google, a simple Gemini query requires only five drops of water, a number that's hard to believe.
According to a post on the Google Cloud blog , the tech giant conducted a study to measure the impact of its AI systems. The study analyzed Gemini's energy consumption, carbon emissions, and water usage during inference, the part where the model generates answers.
According to the data, any Gemini text question consumes 0.24 watt-hours of energy, emits 0.03 grams of carbon dioxide, and consumes 0.26 ml of water, which is equivalent to about five drops.
Google notes that these figures are "significantly lower" than many public estimates, and that using its AI produces an effect comparable to watching television for nine seconds. The company attributes this progress to the efficiency of its AI systems, which have reduced energy consumption by up to 33 times over the past year.
To arrive at this result, Google uses a comprehensive methodology that encompasses multiple aspects. The company measures not only the power consumption of active chips during inference, but also inactive chips to gain an idea of the system's total dynamic power. Google also considers the power consumption of the CPU and RAM, as well as cooling systems, power distribution, and other supporting components.
- Does Gemini really only consume five drops of water?
The most surprising value in this study is water consumption. According to Google, a text-based question directed to Gemini requires only five drops of water. This is possible thanks to the efficiency of AI systems, which require less water for cooling.
The company added: "Our comprehensive systematic estimates (0.24 watt-hours of energy, 0.03 CO2 equivalent, and 0.26 ml of water) take into account all the key elements of a global AI service. We believe this is the most comprehensive view of the overall AI footprint."
While Google is confident that its study provides a complete picture, many researchers question these figures. Shaoli Ren, a professor at the University of California, Riverside and author of a study measuring AI water consumption, accused Google of withholding crucial information.
Ren points out that Google's study doesn't take into account indirect factors, such as the water used by power plants that power data centers. These plants use cooling towers that convert water into steam that is released into the atmosphere. The researcher added that in the United States alone, Google's data centers consumed 12.7 billion liters of fresh water to cool their servers in 2021.