MIT unveils EnergAIzer estimator

Artificial intelligence jobs run on graphics processors that draw different amounts of power depending on the model, input size, and chip. MIT researchers said on April 27 they built a tool called EnergAIzer to estimate that power use in seconds. (news.mit.edu) The project came from MIT and the MIT-IBM Watson AI Lab, with Kyungmi Lee as lead author and Anantha Chandrakasan among the co-authors. The paper is titled “EnergAIzer: Fast and Accurate GPU Power Estimation Framework for AI Workloads.” (news.mit.edu) (arxiv.org) Power models usually need either detailed simulation or direct hardware profiling, both of which can take hours or days before an operator gets an answer. EnergAIzer replaces that bottleneck with a lightweight performance model that predicts the chip-usage signals those power models need. (arxiv.org) (research.ibm.com) In plain terms, the system tries to predict how busy a processor’s parts will be before the job runs, the way a traffic model estimates congestion before cars hit the road. That lets it estimate electricity use without waiting for a full test run on the target machine. (arxiv.org) The MIT team said the method produces reliable estimates in a few seconds instead of hours or days. In the paper, the researchers reported roughly 8% power error on NVIDIA Ampere graphics processors. (news.mit.edu) (arxiv.org) They also said the framework could forecast power on an NVIDIA H100 processor with about 7% error when exploring different frequency settings and architectural configurations. That means the tool is aimed not only at live scheduling, but also at planning future systems. (arxiv.org) The problem it addresses is growing fast. MIT said data-center operators and algorithm developers need quicker ways to measure energy use as artificial intelligence workloads push up electricity demand and sustainability concerns. (news.mit.edu 1) (news.mit.edu 2) MIT said a user can enter details such as the model to run and the number and length of inputs to process, and EnergAIzer returns an energy estimate. The researchers said that could help operators place workloads on the most efficient processors and reduce wasted energy. (news.mit.edu) The paper is listed as accepted to the 2026 IEEE International Symposium on Performance Analysis of Systems and Software. The next test is whether cloud operators and chip designers use fast power estimates as a routine planning number, not just a lab measurement. (kyungmi-lee.github.io) (arxiv.org)