Google LLC has released technical data about the environment footprint of its Gemini artificial intelligence apps.

The company shared the information in a paper published today. According to Google, the median text prompt to a Gemini app uses 0.24Wh, or 0.24 watt-hours, of energy.

That’s enough to power a highly efficient lightbulb for more than a minute. Prompts that ask Gemini to generate multimodal content or analyze a large amount of data use considerably more energy.

Gemini is powered by Google’s TPU series of custom artificial intelligence processors. The company detailed that those chips account for 58% of the 0.24Wh used by a median Gemini text query. It’s unclear which specific TPU models were used in the evaluation.

Google’s newest TPU chip, which debuted earlier this year, offers more than 30 times better energy efficiency than the first processor in the series. It’s deployed in water-cooled server clusters that can host up to 9,216 TPUs. Google says that each such cluster consumes nearly 10 megawatts of electricity, which is enough to power a few thousand homes.

While measuring Gemini’s power usage, the company accounted for fluctuations in the energy consumption of TPUs. “This includes not just the energy and water used by the primary AI model during active computation, but also the actual achieved chip utilization at production scale, which can be much lower than theoretical maximums,” Google executives Amin Vahdat and Jeff Dean wrote in a blog post. 

A quarter of the power used by the median Gemini prompt goes to the servers in which TPUs are deployed. In particular, that power is consumed by the machines’ central processing units and memory chips.

Some of the TPU servers in Google’s data centers are idle. Those servers are not offline, but rather operate in a low-power mode that allows the search giant to activate them quickly if an AI cluster requires additional capacity. Google says the idle machines account for 10% of median Gemini prompts’ power consumption.

Auxiliary data center infrastructure accounts for the remaining energy usage. Google says that this equipment includes, among others, cooling and power distribution systems..

The company’s newly published paper also contains other data points about Gemini’s environmental footprint. Google researchers estimate that the median text prompt causes its equipment to emit 0.03 grams of carbon dioxide. In the process, each such request consumes 0.26 milliliters of water, which is equivalent to about five drops.

“Over a 12-month period, while delivering higher-quality responses, the median energy consumption and carbon footprint per Gemini Apps text prompt decreased by factors of 33x and 44x,” Ben Gomes, chief technologist of Google’s Learning & Sustainability group, wrote in a blog post.

The research suggests the environmental footprint of Gemini apps is similar to that of ChatGPT. In June, OpenAI Chief Executive Officer Sam Altman disclosed that the average ChatGPT query consumes 0.34 watt-hours and 0.32 millimeters of water.

Image: Google