THE SUSTAINABILITY INTERVIEW
software, power, cooling, security and serviceability in a co-designed rack-scale platform.
The significance is that it addresses efficiency at the architecture stage rather than adding it after the system is built.
“ Helios reflects an important shift in how we’ re thinking about AI efficiency,” Justin says.“ It’ s not something that’ s addressed after a system is built.”
Power and cooling systems affect the efficiency of the entire rack. A component that performs well in isolation may create additional energy demand if it requires disproportionate cooling or complicates system operation. Helios allows those elements to be designed and optimised together. AMD describes the platform as incorporating liquid-cooled, serviceable modules and modular fabrics. Justin says the sustainability implications could be substantial if the 20x30 goal is achieved.
“ Achieving our 20x goal could reduce the infrastructure needed to train a typical AI model from more than 550 server racks in 2025 to two,” he says. This could cut related electricity use and carbon emissions by 95 % or more, according to AMD’ s estimate. Alternatively, a data centre could deliver 20 times more computing using the same amount of energy.
Efficiency meets customer economics Energy efficiency is increasingly becoming a commercial requirement for AMD’ s customers.
“ Customers need substantially more useful compute, but power, cooling and space are finite and expensive,” Justin says.
sustainabilitymag. com 29