4 ms·

Hmm. It looks like natural gas is a little less than $2.50 / GigaJoule (~29.3kWh) at the moment. One A100 eats about 400W, and apparently OpenAI used about 1023

Hmm. It looks like natural gas is a little less than $2.50 / GigaJoule (~29.3kWh) at the moment. One A100 eats about 400W, and apparently OpenAI used about 1023 of those running for 34 days to train ChatGPT[1]. There's also a bunch of supporting equipment, so maybe they used around 500W * 1023 * (34*24=816)hrs = 417,384kWh? I think that would be around $35k in fossil fuels, assuming none of the datacenter's energy came from renewable sources? [1]: https://www.tooltester.com/en/blog/chatgpt-statistics/ https://www.tooltester.com/en/blog/chatgpt-statistics/

sterlind 4y ago
It's based on GPT-3, right? Is the amount of compute needed to train that factored in?
Retric 4y ago
You just assumed 100% efficient conversation between heat and electricity plus zero transmission losses etc, ignored the training time for earlier versions of ChatGPT or any of the building or developers equipment, and of course Natural Gas prices have only recently fallen so they where vastly higher during the actual training. A more reasonable estimate might be 5 years * 500W * 1,000 systems = 76.7 GWh, ~29.3kWh * 55% = 16.1 kWh. Actual electricity generation isn’t 100% natural gas but this is additional demand on the grid which tends to be supplied by natural gas unless there’s a surplus of green or nuclear power. Also, data centers also use AC etc so it’s really quite arbitrary how you want to estimate things. PS: A similar thing applies to EV’s people tend to use the existing mix of generation to estimate how green they are but it’s not like we are adding hydroelectric dams when people increase electricity consumption. It’s really a question of how marginal increases in demand are supplied.
RandomWorker 4y ago
-420,000 kWh X 30 year average grid emissions data California 0.233kgCO2e/kWh = 98tCO2e which is about the emissions of 7 households in the USA for one 2021 year.

sterlind 4y ago

It's based on GPT-3, right? Is the amount of compute needed to train that factored in?

Retric 4y ago

You just assumed 100% efficient conversation between heat and electricity plus zero transmission losses etc, ignored the training time for earlier versions of ChatGPT or any of the building or developers equipment, and of course Natural Gas prices have only recently fallen so they where vastly higher during the actual training. A more reasonable estimate might be 5 years * 500W * 1,000 systems = 76.7 GWh, ~29.3kWh * 55% = 16.1 kWh. Actual electricity generation isn’t 100% natural gas but this is additional demand on the grid which tends to be supplied by natural gas unless there’s a surplus of green or nuclear power. Also, data centers also use AC etc so it’s really quite arbitrary how you want to estimate things. PS: A similar thing applies to EV’s people tend to use the existing mix of generation to estimate how green they are but it’s not like we are adding hydroelectric dams when people increase electricity consumption. It’s really a question of how marginal increases in demand are supplied.

RandomWorker 4y ago

-420,000 kWh X 30 year average grid emissions data California 0.233kgCO2e/kWh = 98tCO2e which is about the emissions of 7 households in the USA for one 2021 year.