A Data-Driven Multi-Level Profiling Framework for Power and Utilization Optimization in GPU-centric Data Centers

With increased demand for generative AI and LLM-based digital services across society, data centers are shifting towards GPU-centric servers with higher power consumption and cooling requirements. Operators utilize power profiles at the server’s AC-input boundary for planning and control. When these profiles are interpretable at the component-level, the drivers of peaks, ramps, and sustained plateaus can be attributed to specific components. However, component-level power profiles are rarely available outside controlled testbeds; in operational facilities, these profiles are restricted to authorized teams and fragmented into non-synchronized logs. Existing datasets and simulators rarely provide an end-to-end power profiling approach that links workload scenarios to component activity, component power, and server AC-input power. This study presents a multi-level approach that generates component-activity profiles using a semi-Markov process (Level-1), maps activity to component-level power using lightweight analytical models (Level-2), and generates AC-input power profiles (Level-3). The approach outputs 6-hour profiles at 1-min cadence for component activity, component power, and AC-input power. A guard-and-repair quality-control policy with six physics guards (D1–D6) enforces physical plausibility and accepts 98.79 % (9879 of 10000) profiles. Generated profiles for publicly available workloads (Llama/ResNet variants) are validated against published DGX H100 AC-input traces, and report 3.7–5.95 % mean absolute percentage error and sub-kW RMSE. The performance of the proposed framework is analyzed under varying inlet temperatures, capping policies, and power modes, along with comparative evaluation against existing simulators. The approach is useful for researchers and practitioners designing, evaluating, and benchmarking power-aware planning and control methods at the AC-input boundary.