A controlled protocol for synchronized advancement cost in GPU shallow-water models: scoped comparisons and implementation decisions with RapidFlow v0.1

Abstract. We present a protocol for comparing synchronized advancement cost in GPU shallow-water models while holding local discrete evolution and absolute selected work fixed. Advancement includes host scheduling, selection, state maintenance and synchronization, and excludes construction and downloads. Its first gate checks sparse execution against the complete discrete operator. A non-flat Thacker candidate demonstrates why this matters: filtered and compacted paths agree with each other while both omit a nonzero momentum update. Changing only selection restores dense agreement through two periods. A default Thacker application passes its declared dense-stage, dry-extension and analytical-accuracy checks, with maximum strict 1 mm wet-set discrepancy of 5.15 % on a fixed common domain. Flat-domain experiments on RTX 4090 and RTX 5090 platforms then quantify implementation choices, with explicitly weaker dense-validation coverage. Across eight stored sizes, the conditional timing comparison finds filtering faster and its relative advantage generally narrows. Restoring omitted work confirms growing savings that the current compaction design does not convert into a net advantage. Preparation fusion helps small cases, but slows the largest flat case on the RTX 4090; its direction there is unresolved on the RTX 5090. The contribution is a reusable rejection and comparison procedure demonstrated by failure, repair and a successful application; flat timing matrices remain conditional implementation comparisons outside the audited representative endpoints. Numerical front failures and same-day timing sessions bound its use for model development and prevent general flood-accuracy or cross-date performance claims.