AMD Improves AI Compute Efficiency, Reduces Energy Consumption by 97% Compared to Five-Year-Old Systems

AMD has embarked on a mission to enhance the energy efficiency of rack-scale AI clusters, with a bold ambition of achieving a 20-fold efficiency improvement by the year 2030. This initiative aims to make AI computations more scalable and environmentally sustainable as the demand for computational resources continues to grow.

AMD’s Commitment to Energy Efficiency in AI

[Press Release]: For AMD, energy efficiency is a fundamental principle that has influenced our design philosophy and product roadmap for years. Over the past decade, we have set and met ambitious, publicly announced targets to boost the energy performance of our offerings. Today, we proudly announce that we have exceeded our 30×25 goal, while also setting our sights on an exhilarating new target for the coming years.

During the recent Advancing AI conference, we revealed that AMD not only met but surpassed the 30×25 goal established in 2021, which aimed to enhance the energy efficiency of AI training and high-performance computing nodes by 30 times between 2020 and 2025. Although achieving this milestone is a significant accomplishment, our journey does not end here.

As artificial intelligence continues to expand and evolve, the necessity for complete end-to-end AI system designs becomes increasingly apparent. To maintain our leadership in energy-efficient designs, we are setting an audacious new target: a 20-fold increase in rack-scale energy efficiency for AI training and inference, beginning from a 2024 baseline and aiming for completion by 2030.

Defining a New Standard for AI Efficiency

With the growth of AI workloads and the incessant rise in demand, it is clear that improvements limited to node-level efficiency will not suffice. The greatest advancements in efficiency are now achievable at the system level, which lies at the core of our 2030 objective.

We are confident in our ability to attain this ambitious 20-fold increase in rack-scale energy efficiency in AI training and inference by 2030, which stands to exceed the projected industry improvements from 2018 to 2025 by nearly three times. This target encompasses performance-per-watt enhancements spanning the entire rack, including CPUs, GPUs, memory, networking, storage, and the synergistic design of hardware and software—an evolution made possible through our comprehensive end-to-end AI strategy aimed at scalable and sustainable data center operations.

Real-World Impact of Improved Efficiency

Achieving a 20-fold enhancement in rack-scale efficiency, at an acceleration rate nearly three times the previous industry average, will have profound consequences. Using the training of a representative AI model projected for 2025 as a reference, the anticipated benefits include:

Consolidation of racks from over 275 to less than one fully utilized rack.
A remarkable reduction in operational electricity consumption by more than 95%.
A decrease in carbon emissions from approximately 3, 000 metric tons to just 100 metric tons of CO2 equivalent during model training.

At AMD, we are thrilled to pursue these opportunities to not only elevate performance levels but also redefine the possibilities when energy efficiency takes precedence. As we progress towards our goal, we will keep our stakeholders informed about our advancements and the positive impacts these improvements will have across the ecosystem.