GPUDECK · OPEN DATA · EVERY PUBLISHED FIGURE, REGENERABLE
Every published figure derives from these files under the numbered rules in METHODOLOGY.md. Each file carries its SHA-256, size, row count and as-of date — and manifest.json holds the same register machine-readable — so a downloaded copy verifies byte-for-byte and a citation names an exact edition. Under-replicated cells are suppressed and disclosed, methodology changes get a dated entry before the data they affect ships, and the GPU-model catalogue is append-only so a published citation stays valid. The generator source is not yet public; until it is, the methodology file is the rebuild recipe and this register is the proof of inputs. The same register is served as a versioned JSON API: the endpoint index lists every route, its shape and its as-of date.
Paid completions per occupied GPU-hour as per-workload ratios, with attempts, completions and replication counts on each side.
CSVCompleted-render latency ratios per workload pairing. This file carries time_ratio and no median_s or absolute rate column of its own; what the record as a whole discloses is set out in METHODOLOGY.md §8.
The value board: cost per 1,000 completed renders per (workload, GPU), with full price provenance columns.
CSVCompletion yield per (fleet, GPU, snapshot): of the attempts a card started, the fraction it finished and was paid for. One row per card per ARCHIVED SNAPSHOT — these are dated rolling windows, not closed calendar days: each row carries its own measurement interval and effective hours, some are shorter than others, and where two overlap the same completions appear in both, so those rows are not independent. The JSON lists every overlapping pair. Cells backed by fewer than two distinct physical cards are suppressed under the same replication floor as the goodput exports, and the suppressed cells are listed rather than dropped. The gap between started and completed is observed non-completion: the record does not classify how much of it is a fault GPUDeck could act on, so it is neither a count of the failure GPUDeck targets nor a measurement of GPUDeck's effect on anything — the software ran throughout and there is no untreated baseline.
CSVThe same yield cells as JSON, plus the suppressed-cell list, the days absent from the archive, and the reason occupied_attempts and occupied_terminated are deliberately not published.
JSONEMPTY AND WITHDRAWN since 2026-08-26. This series came from a third-party cross-provider tracker whose chain of title is unresolved: the licence is real, but the dataset’s own documentation states it is collected by scraping provider pricing surfaces, including providers whose terms prohibit it. The file is kept at zero rows so the URL and its schema stay stable. Current prices are in the per-source files and on /prices; the reasoning is in /corrections.
CSVDirect-source price observations from Lium’s public nodes feed, ingested under an archived recorded basis (methodology §8): one row per rentable node per day, segmented on-demand-single / multi-minimum / spot. Never pooled with any other source.
CSVEvery admitted DIRECT source's published tariff, one row per observation, each naming its own source, unit and currency. Deliberately NOT normalised into a single price column: Hetzner and Leafcloud publish monthly euros while Nebius and Koyeb publish hourly dollars, and converting between them would invent an hours-per-month figure and an exchange rate that no provider stated. Every row carries the archived digest of the page it was read from.
CSVThe longitudinal hardware census: which GPU models are observed on the public fast network, daily since 17 Dec 2025.
CSVThe permanent, append-only catalogue order fixing which GPU is listed first in every paired comparison.
CSVHow every figure is derived, what is deliberately not published and why, versioned with dated changes.
MDThe append-only corrections ledger behind /corrections: every correction to a frozen edition, with severity, affected files and digests. Frozen editions are never edited.
JSONThe third-party dispute register behind /disputes: challenges from outside GPUDeck that a published figure misrepresents someone, with category, status and resolution. Frozen editions are annotated, never edited.
JSONThe operator contribution protocol: submit your own fleet's goodput cells — published as a second fleet under your identifier, never averaged into the first.
MDPublish daily workload, GPU, attempts, completions and occupied GPU-hours under your own fleet identifier. Fleets are never averaged; earnings and worker identity are not accepted. Submitting does not make data part of the record — protocol validation is the gate.