GPUDECK · AN OPEN RECORD, NOT A LEADERBOARD
Renting a GPU shows you one number — dollars per hour — and hides the one that matters: what an hour actually finishes. GPUDeck is an independent, open-data record of both, measured on a real fleet, dated, and downloadable to the last digit. Built to be cited, so it never publishes a figure it cannot regenerate from the open files. The 60-second version of why hourly price lies →
Paid completions per occupied GPU-hour, per workload, from one production render fleet — published as ratios inside exact workload cells, with the confounds named and no cross-workload aggregate invented. The same pairing can swing double digits between two dated windows; that finding is the argument for measuring this way.
View the observatoryWhich GPUs supply-side operators actually run — and abandon. A dated public census of a live render network, archived daily, with class lifelines showing cards entering and leaving the fleet population over time.
View the censusEvery published figure is regenerable from the files: goodput and throughput ratios, break-even price multiples, census history, class catalogue, methodology and citation file. Under-replicated cells are suppressed and disclosed, never padded or averaged away.
Read the methodologyOne operator, disclosed. The goodput cells come from a single production fleet. That is a limitation and it is printed on every figure — physical-card support and known host confounds travel with each result.
Ratios, not rankings. Nothing here says which GPU is best. It says what one workload finished on one pairing in one dated window — the honest unit of comparison.
Reversible by design. Methodology changes get a dated entry before the data they affect ships, and the class catalogue is append-only so a published citation stays valid.
CC-licensed, with a citation file. If you publish from it, the archive keeps your citation reproducible.