Big quote: “The surge in energy usage by GenAI and similar advanced models necessitates more efficient power solutions. AMD’s CEO, Lisa Su, is optimistic about enhancing data center power efficiency by a factor of 100 within the next three years.”
AI services are burgeoning, aiming to upgrade both personal and professional spheres. Google Search integrates its Gemini AI for better search summaries, albeit at a significantly higher energy cost, ten times that of traditional searches, often yielding unsatisfactory outcomes. The rising tide of generative AI demands swift data center and power supply expansions.
An estimate by Goldman Sachs suggests data center energy needs will spike by 160% come 2030, presenting a major challenge for regions like the United States and Europe, where the power grids average 50 and 40 years old, respectively. Data centers are projected to consume 8% of the US’s power by 2030, up from 3% in 2022, as per forecasts. “Achieving this without a significant breakthrough seems unfeasible,” mentions OpenAI co-founder Sam Altman.
At the ITF World 2024 conference, AMD’s CEO, Lisa Su, reflected on past accomplishments and forthcoming initiatives to uplift compute node efficiency. By 2020, AMD aimed to enhance their mobile CPUs’ efficiency by 25% (a campaign termed 25×20) and surpassed this target by reaching 31.7% efficiency.
In light of the growing AI workloads and their power demands in 2021, AMD set a new target, the 30×25 goal, focusing on key areas to reduce power consumption. Enhancements in process node and packaging technology form the cornerstone of this effort, with the adoption of 3nm Gate-All-Around (GAA) transistors signifying a step forward from FinFET 3D transistors, thereby boasting improved power efficiency and performance per watt. Innovations in packaging techniques, such as chiplets and 3D stacking, provide AMD the versatility to incorporate various components into a unified package.
The advancement towards AI-optimized accelerated hardware architectures is another significant focus. Neural Processing Units (NPUs), already prevalent in mobile SoCs like the Snapdragon 8 Gen series, have made their desktop debut with AMD’s release of the Ryzen 8700G, featuring an integrated AI engine. This allows for the offloading of AI-centric tasks, augmenting efficiency and curbing power use.
System-level tuning and software/hardware co-design stand as the other critical aspects of the 30×25 initiative. System-level tuning intensifies the packaging initiative’s efforts to minimize the energy required for in-computer cluster data transmission. Meanwhile, software/hardware co-design strives to refine AI algorithms for optimal efficiency with next-gen NPUs.
Lisa Su remains optimistic about achieving the 30×25 goal and foresees a path towards a staggering 100x improvement by 2027. AMD, along with other industry torchbearers, is actively contributing towards meeting the burgeoning power requirements dictated by our AI-driven existence in this novel era of computing.