I’d like all ai service to publish the energy used in training the model and performing inference.
“Queries uses an average of X kWh of power.
A model training run requires X MWh, and the development of this model over the years required X TWh of power.”
Then we could judge companies by that metric. Off course, rich people would look for the most power-draining model for the sake of it.
development of this model over the years required X TWh of power
This part is kind of hard to measure. When do you start counting? From the first work that informed the research direction eventually leading to this model? From the point where the concept of this final model first came about? Do you split the energy usage between multiple models that came from the same work?
I’d like all ai service to publish the energy used in training the model and performing inference.
“Queries uses an average of X kWh of power. A model training run requires X MWh, and the development of this model over the years required X TWh of power.”
Then we could judge companies by that metric. Off course, rich people would look for the most power-draining model for the sake of it.
This part is kind of hard to measure. When do you start counting? From the first work that informed the research direction eventually leading to this model? From the point where the concept of this final model first came about? Do you split the energy usage between multiple models that came from the same work?