Justin Murrill CSO AMD
“ We achieved 38x,” Justin says.“ Our current goal is the third in a track record that goes back over a decade.” The shift from processors and nodes to racks reflects the changing structure of AI infrastructure.
“ Performance and efficiency are no longer determined by an individual processor,” Justin explains.
Data movement can create bottlenecks and consume energy, so AMD is pursuing system-level co-design.
“ By optimising compute architecture, memory bandwidth, data movement, interconnects, software and system design together, we can address bottlenecks,” he says.
The approach also responds to customer expectations. Operators increasingly want complete computing platforms they can deploy and manage as integrated infrastructure.“ Customers want to buy platform compute solutions, not just components,” Justin says,“ and advance their AI capabilities without sacrificing energy and sustainability performance.”
How Helios makes efficiency architectural AMD Helios is the company’ s open, rackscale AI architecture. It brings together AMD EPYC central processing units, AMD Instinct graphics processing units and networking within one system. The platform is designed around open standards, including specifications developed through the Open Compute Project. AMD says Helios combines compute, memory, networking,
sustainabilitymag. com 27