MIT and MIT‑IBM Lab develop rapid AI power estimator

MIT researchers and the MIT-IBM Watson AI Lab have developed EnergAIzer, a fast estimator that predicts power consumption of AI workloads on GPUs.

• Main announcement: The team (led by Kyungmi Lee with senior author Anantha P. Chandrakasan) unveiled EnergAIzer, a lightweight estimation model that produces reliable GPU power estimates in a few seconds (vs. hours or days for traditional emulation) and achieved about 8 percent error when tested on real GPU workloads; the research is being presented at the IEEE International Symposium on Performance Analysis of Systems and Software.
• Background and details: The method leverages repeatable software optimization patterns to generate fast estimates, includes correction terms derived from real GPU measurements for accuracy, can be applied to future/emerging GPU configurations, and the work was funded in part by the MIT-IBM Watson AI Lab.