Steffen ChristensenApr 22, 2024

AI is so small in energy use compared to cryptocurrency.

Training GPT-3 used 1.3 GWh, BLOOM 0.4 MWh. The massive GPT-4 used between 29 and 62 GWh. Using the higher number for GPT-4, we could estimate training Google's frontier model Gemini Ultra at 150 GWh.

In 2023, crypto used 120,000 GWh of energy.

Show threadSteffen ChristensenApr 22, 2024
Source for crypto: Cambridge Bitcoin Electricity Consumption Index (CBECI) estimates.
https://www.eia.gov/todayinenergy/detail.php?id=61364#:~:text=The%20CBECI%20estimates%20that%20global,to%20have%20been%2027%2C400%20TWh
Tracking electricity consumption from U.S. cryptocurrency mining operations - U.S. Energy Information Administration (EIA)

Energy Information Administration - EIA - Official Energy Statistics from the U.S. Government

Show threadSteffen Christensen
Sources for AI:
Computations used in training from Our World In Data, https://ourworldindata.org/grapher/artificial-intelligence-training-computation?time=2016-11-17..latest
Computation used to train notable artificial intelligence systems

Computation is measured in total petaFLOP, which is 10¹⁵ floating-point operations estimated from AI literature, albeit with some uncertainty. Estimates are expected to be accurate within a factor of 2, or a factor of 5 for recent undisclosed models like GPT-4.

Our World in Data
Apr 22, 2024 at 5:07pmWeb
Show threadSteffen ChristensenApr 22, 2024
Energy used for GPT-3 and BLOOM: "Estimating the carbon footprint of BLOOM, a 176B parameter language model," Luccioni et al., 2022.
Show threadSteffen ChristensenApr 22, 2024

Energy used for GPT-4.

Source 1: ScaleDown, "The carbon impact of large language models: AI's growing environmental cost," https://tinyml.substack.com/p/the-carbon-impact-of-large-language

Source 2: Towards Data Science, "The carbon footprint of GPT-4," https://towardsdatascience.com/the-carbon-footprint-of-gpt-4-d6c676eb21ae

The Carbon Impact of Large Language Models: AI's Growing Environmental Cost

A guide to the Energy Demands and CO2 Emissions of Leading LLMs in a Sustainability-Conscious Era

ScaleDown
Show threadSteffen ChristensenApr 22, 2024
Energy used for Gemini Ultra is from scaling GPT-4 energy use to Gemini Ultra's estimated training effort of 50 billion petaFLOPs.
Show threadMarsh RayApr 22, 2024
@Wikisteff Wow, that comic depicting electric power generation as radioactive cooling towers emitting “smoke” is fundamentally dishonest.
Show threadSteffen ChristensenApr 23, 2024

@marshray Honestly I think was a straight-up mistake, but it's hard to tell without context.

Certainly unhelpful to say "energy is use is bad because CO2" and then use a zero-marginal emission power plant.