From Servers to Sites: Compositional Power Trace Generation of LLM Inference for Infrastructure Planning

A team from Stanford University has published a research paper, 'From Servers to Sites: Compositional Power Trace Generation of LLM Inference for Infrastructure Planning,' introducing a novel framework to model the unique and volatile electricity consumption of large language models. Traditional datacenter power models fail to capture the specific workload patterns of LLM inference, where GPUs rapidly cycle between high-power compute states (prefill), lower-power generation states (decode), and idle. This new model breaks down power usage into two compositional components: workload-driven state transitions and configuration-specific power distributions within those states.

By learning from measured power traces, the framework can synthesize accurate load profiles for new traffic conditions and server configurations, scaling from individual GPU servers up to entire facility-level demand. The researchers validated the model across multiple LLMs, tensor-parallel settings, and GPU generations, achieving a median absolute energy error below 5% for most setups while preserving critical temporal patterns. This enables downstream analyses previously impossible with flat 'nameplate' power assumptions, such as evaluating safe levels of electrical oversubscription, planning for power modulation, and characterizing utility-facing load profiles for grid interconnection studies.