AMD says AI energy efficiency rises fourfold by 2026
Wed, 19th Aug 2026 (Today)
AMD says it has achieved an estimated fourfold increase in AI energy efficiency by mid-2026, putting it ahead of its interim target on the way to a 2030 goal.
The improvement exceeded the projected threefold increase for this stage and was more than double the historical trendline at the same point. AMD measures progress against a 2024 baseline as it works towards a 20-fold increase in rack-scale energy efficiency for AI training and inference by 2030.
The update comes as data centre operators face rising electricity demand from AI systems and growing pressure to improve the economics of large-scale model training and inference. Energy use, cooling limits and physical infrastructure have become central constraints as businesses add more compute to support AI workloads.
AMD says its estimate reflects a system-level approach spanning compute silicon, memory, interconnects, software and rack design. It argues that rack-level efficiency gains depend on how processors, accelerators, memory, networking and software work together, rather than on advances in chips alone.
That focus mirrors a broader shift across the semiconductor industry, where suppliers increasingly present AI infrastructure as a tightly integrated stack. As model sizes grow and data movement rises, memory bandwidth and interconnect efficiency have become as important to overall power use as raw processing performance.
AMD also set out a 2030 scenario based on a representative AI training workload. Under that projection, about two AMD racks would deliver the same compute as 570 racks in 2024, reducing use-phase electricity by 20 times and carbon intensity by 28 times.
The same efficiency gains could also produce 20 times more compute, measured in floating point operations per second per watt, using the same amount of energy. Those calculations rely on projected product performance and power characteristics, as well as external energy and workload assumptions.
System design
According to AMD, the main drivers of AI system performance beyond software are compute capability, memory bandwidth and interconnect bandwidth. It says process technology, architecture, memory integration and high-speed interconnects are central to improving performance per watt across AI systems.
Software also remains part of the equation. AMD cited its ROCm software stack, open standards and work with customers as factors that can help improve deployment efficiency for AI and high-performance computing workloads.
Sam Naffziger, Senior Vice President and Corporate Fellow at AMD, linked the latest figure to the company's broader engineering approach.
"The next wave of AI efficiency will depend on tighter co-optimisation across compute silicon, memory, interconnects, software and rack-scale system design. Our estimated 4x improvement through 2026 reflects the strength of our approach and the progress AMD is making across the full system, putting us ahead of our projected pace toward the 2030 goal," said Sam Naffziger, Senior Vice President and Corporate Fellow at AMD.
Wider report
Alongside the AI efficiency update, AMD published its 2025-26 Corporate Responsibility Report covering its operations, supply chain and community programmes. It said it reduced operational greenhouse gas emissions by 30% between 2020 and 2025 and sourced 58% of its global electricity from renewable sources.
AMD also said it audited 99% of its manufacturing supplier factories for responsible business practices. Those figures form part of a broader effort by semiconductor companies to show progress on environmental and supply-chain issues as customers and investors scrutinise the resource demands tied to AI growth.
The methodology for the 2026 result combines measured product data with modelled estimates where final performance data was not available. Progress towards the 2030 target is assessed by comparing annual rack configurations with the 2024 baseline using a performance-per-watt measure.
For data centre operators weighing how to expand AI systems, the significance of such targets lies in whether higher output can be delivered without matching increases in power draw. AMD says customers need more useful compute from the energy available as power and cooling constraints become limiting factors in scaling AI and high-performance computing.