A Question on The Carbon Footprint Estimation.

[yaoyiran]

Hi,

The carbon emission is estimated in the DALL-E Mega Model Card and DALL-E Mini Model Card respectively. However, I have a concern about if the correct number of TPU-3 chips is fed to the Machine Learning Emissions Calculator.

DALL-E Mini: in the model card, the reported emission is 7.54 kg CO2-eq. It is trained on a single TPU v3-8 for 72 hours. However, I can find from Google cloud's website that a TPU v3-8 contains 8 tensorcores and 4 chips. So the input to the Machine Learning Emissions Calculator should be 4 X 72 hours (rather than 72 hour which derives 7.54 kg CO2). So the emission should be 4 X 7.54 kg CO2-eq, in my opinion.

DALL-E Mega: in the model card, the reported emission is 4503 kg CO2-eq. It is trained on a TPU v3-256 for 1344 hours. However, I can find from Google cloud's website that a TPU v3-256 contains 256 tensorcores and 128 chips. So the input to the Machine Learning Emissions Calculator should be 128 X 1344 hours (rather than 32 X 1344 hours which derives 4503 kg CO2). So the emission should be 4 X 4503 kg CO2-eq, in my opinion.

Could you double-check this a bit? Thank you so much!

[borisdayma]

Many thanks for noticing it.
I think that you are correct.
Do you want to do a pull request on the model cards?

[yaoyiran]

Thanks for your double-check! I just pulled requests for DALL-E Mega and DALL-E Mini respectively on their Hugginface model card pages.