Smoothing the Ramp, Not the Peak: Scheduling-Induced Power Dynamics of LLM Inference and Their Grid-Scale Consequences

What the math gives to ML

The paper identifies chunked-prefill scheduling as a controllable dynamical mechanism for shaping LLM inference power: splitting long-prompt computation into smaller scheduling quanta substantially reduces power ramp rate while leaving peak power nearly unchanged. The effect strengthens monotonically with concurrency and the fraction or size of long-context requests, indicating that scheduling benefits are state-dependent rather than fixed. The transferable neural-network asset is an adaptive inference scheduler that chooses prefill chunk size from measured saturation and explicitly optimizes high-quantile power ramps, latency, and throughput rather than peak power alone. Its success is falsifiable through predicted ramp-rate reductions and reserve-quantile changes obtained by bootstrap resampling of GPU traces.