The rapid developme of artificial ielligence has made the supply of electricity needed to be a serious challenge. The latest studies show that the power consumption of artificial ielligence dataceers may soon be outstanding the electricity needed for bitcoin extraction.
According to Tekles, Alex de Ferris-Gao, a doctoral stude at the Institute of Environmeal Studies at the University of Amsterdam, examined this transformation in his research published in Joule. The research, based on the general technical specifications of the devices, analyzes and corporate information, estimates the volume of production and energy consumption of artificial ielligence hardware.
Providing electricity needed for artificial ielligence is a serious challenge for technology companies
Since major technology companies rarely expose their artificial ielligence operations, the researcher has used Triangulation to estimate. He has estimated the amou of electricity needed by these operations by examining the supply chain of advanced chips and the production capacity of key actors such as TSMC.
Accordingly, every Nvidia H100 artificial ielligence chip, one of the main compones of modern data ceers, consumes 2 watts of electricity when implemeing complex models. If we multiply this number by millions of Nvidia H100 chips, the total energy consumption will reach a worrying surface.
The researcher estimates that artificial ielligence hardware may require between 1.5 and 2.5 GW of electricity. This is also higher than the total power consumption of Ireland.
The main wave of increased energy consumption in artificial ielligence is still on the way. TSMC’s Cowos packaging technology allows powerful processors and high -speed memory to become the heart of modern artificial ielligence systems. Du Ferris-Ghao says TSMC has doubled its Cowos production capacity between the ages of 1 and 2, but the demand of chip-produced companies such as Nvidia and AMD is still beyond the release.
TSMC apparely plans to double the production capacity of Cowos. The researcher predicts that if the curre trend coinues, the total electricity required by artificial ielligence systems could reach 2 GW by the end of the year, which is approximately the average electricity consumption of the UK.
