Intelligent Data Centres Issue 64 | Page 41

F E A T U R E

EMBRACING DIRECT-TO-CHIP

COOLING FOR AN ENERGY-EFFICIENT AI ERA

Josh Claman , CEO , Accelsius , argues data centres are the plumbing for all things digital , and the desire for power isn ’ t diminishing any time soon . Consequently , operators should be utilising all emissions to initiate more sustainable solutions .
he AI boom has

T expedited innovation and breakthroughs , bringing change to almost every aspect of people ’ s lives with the guarantee to continue to do so . It has enhanced and streamlined billions of processes , aiding practically all industries , from healthcare to education , and bringing unprecedented change to daily lives .

However , this technology does not come without its downsides . The growing demand for AI-driven power has resulted in an exponential increase in electricity and water usage within data centres , and the current infrastructure cannot withstand the ever-increasing demand without change .
AI chips and workloads already demand more power than past solutions , and data centres are unprepared for the surge in electricity AI will need . Data centres use about 40 % of their power allocation on inefficient air-cooling infrastructures for regular chips and workloads . With AI , data centres will need to draw more energy from the power grid due to the higher computing needs of AI chips . This presents the primary issue – it is no longer a question of ‘ if ’ we will run out of power , but ‘ when ’ we will .
Recent data shows that current data centre racks average about 7kW of power , but when integrating AI hardware , the power consumption can jump to anywhere from 30 – 100kW or more . This 4 – 5x increase in energy use can cause added strain on power grids and create additional heat in data centres .
To put this in perspective , in 2022 , data centres and AI used 460TWh of electricity . The energy usage is expected to nearly triple by 2026 with data centres set to use more than 1,000TWh at the current pace of technological development . The addition of this power consumption is roughly equal to Japan ’ s
www . intelligentdatacentres . com 41