A faster way to estimate AI power consumption

Researchers from MIT and the MIT-IBM Watson AI Lab have developed a new tool called EnergAIzer that can predict the energy consumption of AI workloads in seconds. As artificial intelligence continues to grow, data centers are projected to consume up to 12 percent of total U.S. electricity by 2028. This rapid estimation method offers a practical solution for data center operators and algorithm developers to improve the sustainability of AI by identifying energy-efficient configurations before deployment.

Overcoming traditional modeling limitations

Traditional methods for predicting energy consumption often involve breaking down workloads into individual steps and emulating how every module within a graphics processing unit (GPU) is utilized. Because AI training and data preprocessing tasks are massive, these simulations can take hours or even days to complete. This slow turnaround makes it impractical for operators to compare different algorithms or hardware configurations to find the most efficient path forward.
To address this, the research team focused on the repeatable patterns inherent in AI workloads. By leveraging the structured optimizations that software developers use to distribute work across parallel processing cores, the team created a lightweight model. EnergAIzer captures these power usage patterns to generate reliable estimates significantly faster than conventional emulation techniques.

Achieving accuracy through real-world data

While the initial estimation model was fast, the researchers recognized that it needed to account for additional energy costs, such as the fixed power required for program setup and the inefficiencies caused by hardware fluctuations or bandwidth limitations. To ensure accuracy, the team gathered real measurements from GPUs to generate correction terms, which they applied to the EnergAIzer framework.
When tested against real-world AI workloads, EnergAIzer produced results with only about 8 percent error. This level of accuracy is comparable to traditional methods that require much longer processing times. The tool allows users to input specific details, such as the AI model being used and the volume of inputs, to receive an immediate energy consumption estimate. Users can also adjust GPU configurations or operating speeds to see how different design choices impact power usage.

Future applications for sustainable computing

The versatility of EnergAIzer allows it to be applied to a wide range of hardware, including emerging designs that have not yet been deployed. Because the tool provides direct feedback, it encourages developers and operators to prioritize energy efficiency throughout the development process.
The research team, which includes lead author and MIT postdoc Kyungmi Lee and senior author and MIT provost Anantha P. Chandrakasan, presented their findings at the IEEE International Symposium on Performance Analysis of Systems and Software. Future work will focus on testing the tool on the latest GPU configurations and scaling the model to manage multiple GPUs collaborating on a single workload. The researchers aim to provide a comprehensive solution that helps hardware designers, operators, and developers across the entire stack reduce the environmental impact of AI.