GPU Power Modelling for LLM Serving With Serving‐Visible Information

ABSTRACT

Large language model (LLM) serving is becoming an increasingly power‐relevant workload in modern data centres, whereas offline serving analysis, trace replay and simulation often lack deployment‐time hardware telemetry. This paper quantifies how much GPU board‐power signal remains recoverable from serving‐visible artefacts under that telemetry constraint. The task is formulated as nonoverlapping 15 s window‐level regression under an analysis‐time feature boundary. Request variables are log‐native, batch and phase variables are scheduler‐derived and coarse execution‐family variables are reconstructed from runtime records or simulator estimates. Device and process telemetry, together with raw platform‐specific execution statistics, define the full‐information reference. Experiments on a Qwen3‐8B vLLM deployment using an A100 PCIe 80GB GPU show that, across the workload regimes represented in the dataset, serving‐visible features retain substantial power‐relevant signal: request and batch information provide most of the visible predictive signal, coarse execution‐family summaries provide the largest visible increment and hidden telemetry defines the remaining upper‐bound gap. The selected execution‐aware serving‐visible predictor is further connected to trace replay through a feature adapter to produce window‐aligned power trajectories for policy‐level performance and power analysis.