Bananas for Scale
ChatGPT Query (Single Response)

Photo from Wikimedia Commons

ChatGPT Query (Single Response)

Large language models have large power bills/Computing

A single ChatGPT query uses roughly 10 times the energy of a Google search, at about 3 watt-hours per response. Running GPT-4 class models at scale requires enormous GPU clusters drawing megawatts of power. As AI usage grows, the energy footprint of language models has become a significant topic in sustainable computing.

Measurements

Energy per query10,800 J
82.4 millionthsGallons of gasoline
5.4 billionthsFireworks shows

About 3 Wh (estimated)

Typical response time5 s
1.4 hundredthsBohemian Rhapsodies
2.4 hundredthsTaylor Swift songs

Varies with length

Typical response size2,000 B
7.6 thousandthsNES game cartridges
426 billionthsEntire Shrek DVDs
2,000Text characters

About 2 KB of text

Single GPU power (H100)700 W
1.8 septillionthsSun luminosities
4.7 tenthsTreadmills
GPT-4 training energy180 trillion J
1.8 nonillionthsSupernovae
43 billionFood calories

Estimated 50 GWh

Browse more in Computing