Power Aware Dynamic Reallocation For Inference (RAPID)

Disaggregated prefill/decode inference improves utilisation and throughput under fixed hardware capacity, but as model and cluster scales grow, power rather than compute becomes the dominant limiter of performance and cost efficiency.

Joints GPU role assignment with power budgets, adding static and dynamic power reallocation on top of GPU reallocation, so that goodput is sustained under strict power caps rather than only under fixed hardware capacity.

Up to 2x improvement in SLO attainment at peak load compared with a static assignment under the same power caps, 'without an increase in complexity or cost'. CAUTION: the abstract states no baseline cluster size, model, workload or cap value, so this is a directional claim of the authors' own and is weaker evidence than the measurement papers in this slice. The framing sentence - power rather than compute has become the dominant limiter - is the more durable contribution and is consistent with the rack-level power density figures recorded in supply/gb200-nvl72-rack-120kw-liquid-required.

Included because it makes the power-as-binding-constraint argument explicitly in a serving-systems context, which is the datacenter-systems angle rather than the energy/carbon/algorithmic-efficiency line - it treats power as a schedulable resource binding SLO attainment, not as an environmental quantity. It should be read as corroboration of the constraint framing rather than as strong evidence for its own speedup number: a bare 2x with no baseline is exactly the kind of claim SCHEMA.md warns against, and it is recorded here with that caveat attached rather than dropped, because the constraint framing is the valuable part. Distinct from the other power-cap papers in this slice (Joule Point, Where Does the Energy Go, Characterizing Performance-Energy Trade-offs) in that it allocates power across heterogeneous GPU roles rather than tuning a cap on a homogeneous fleet. AFFILIATIONS BACKFILLED 2026-10-04 (paper-fields pass, agent w3p): PDF footnote block, same shape as joule-point-energy-optimal-power-cap and the same two institutions - the two papers share an author group. Filled from the API/PDF rather than from memory, because an affiliation is a fact about a specific paper's author block and inference from the author's employer history is how wrong affiliations get shipped. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string. AFFILIATIONS VERIFIED CORRECT 2026-10-05 (wave 6, affiliation sweep, agent w6-papers). The task brief reported that wave 4 found RAPID credited to a college its authors have no affiliation with, and asked for this to be checked. It no longer is: the field reads 'AMD Research and Advanced Development' and 'University of Wisconsin-Madison', and that is what the paper says. arXiv HTML v1 for 2601.12241 carries the footnote block directly under the byline - Yiwei Jiang (equal contribution, footnotes 1 and 2), Sangeeta Chowdhary (equal contribution, footnote 1), Nathaniel Morris (1), Rutwik Jain (footnotes 1 and 2), Srilatha Manne (1), Sam Bayliss (1), with '1AMD Research and Advanced Development' and '2Department of Computer Sciences, University of Wisconsin-Madison, Madison, Wisconsin, USA', correspondence to sangeeta.chowdhary@amd.com. Five of six authors are AMD-only; two (Jiang, Jain) carry both. RECORDED ANYWAY, because a negative result that is not written down is indistinguishable from a pass. The substantive claim of this record is unchanged and still carries the caveat already stated above it: 'up to a 2x improvement in SLO attainment at peak load when compared to a static assignment' is the authors' own figure with no baseline cluster size or workload in the abstract, so it is directional. arxiv_id 2601.12241 and the 2026-01-18 submission date were re-confirmed against the arXiv API and abs page in this pass.