I took a look to DCFootprint to get some ideas of abiotic impact of OpenLLM. The principle was to compare GWP get from our study with its result the adapt through a factor.
1-When typing Jean Zay parameters, I see an automatic correlation between "Total power installed (IT + facility) - MW" and "Total energy (IT + facility), for one year - MWh" which prevents to enter the real numbers, that is Power : 2,5 MW and total consumption : 23 709 kWh/an. May be letting the user let free on this part could be more adapted.
2- So I took a lower total energy to get a result then corrected it by the GPU life duration from our study, estimated to 10 years for H100. We have made some update for a September communication at workshop GREEN-AI ECML-PKDD, through a sensitivity analysis, but for this first step I kept the first publication.
Our study led to a 417,7 t CO2eq. with a life of 10 years, 6 years in Jean Zay then 4 years in US data center.
DCFootprint gave 762,8 t CO2eq. after corrected from GPU lifespan.
I wonder if the gap could come from building lifespan : we consider Jean Zay building as amortised so with 0 impact following Labos 1point5 methodology. Do you?
Best regard
