New HSM-DC Model Predicts AI Data Center Power Swings with 99.97% Accuracy
AI training jobs cause 10x power swings in seconds, but grid planners can't ignore scheduling...
AI data centers are a new load class with unique power dynamics driven by the bulk-synchronous-parallel (BSP) algorithm used in training jobs. Inside a job, nodes cycle through compute, sync, and checkpoint steps, causing power to swing from full load to near idle within seconds. Across the facility, jobs arrive and leave over hours to days, shifting the number of busy nodes and driving facility-wide swings and peak demand. Existing models that ignore job scheduling smooth out these variations, missing the true peak-to-average ratio.
To address this, researchers Chaudhary, Bera, Newlun, Ben-Idris, and Mitra propose the Hierarchical Semi-Markov Data-Center (HSM-DC) model. It couples two layers: a job-scheduling layer that creates jobs via a non-homogeneous compound-Poisson process with daily, weekly, and seasonal patterns, and a within-job layer that moves each busy node through a five-state semi-Markov chain for BSP steps with Ornstein-Uhlenbeck noise. Configured for a reference facility, the model achieved fit scores of 0.9997 for mean power, 0.92 for spread, and 0.82 for peak-to-average ratio. It also matched queued job share within one point at high load. The paper (arXiv:2607.12222) was submitted to the 2026 CIGRE Grid of the Future Symposium.
- Within-job BSP steps cause power swings from full load to near idle in seconds.
- Job scheduling layer uses non-homogeneous compound-Poisson process with daily/weekly/seasonal patterns.
- Model achieved fit scores of 0.9997 (mean power), 0.92 (spread), 0.82 (peak-to-average ratio) at reference facility.
Why It Matters
Grid planners must account for AI job scheduling to avoid undersized infrastructure and potential blackouts.