GPUDECK · THE 60-SECOND VERSION
Every rental marketplace shows you one number: dollars per hour. But you don't buy hours because you want hours — you buy them to finish work. And two GPUs holding the same hourly price can finish very different amounts of it.
GPUDeck measures it nightly on a real production render fleet. One real, dated cell — two cards, one workload, same day, both meeting the replication floor with no host confound. Regenerate it from gpu-goodput-ratios.csv:
| WORKLOAD | GPU | COMPLETED / ATTEMPTS | PAID-HOUR GOODPUT |
|---|---|---|---|
| dark_beast_krea2_identity_edit_v1_2 | RTX 4080 | 86 / 88 | RTX 4080 finished 0.51× the work per paid hour of RTX 5090 |
| RTX 5090 | 38 / 38 |
2 and 5 physical cards over 2 and 5 card-days, one operator. Replication is physical cards and card-days, never attempt count.
Why this is goodput and not cost per 1,000. Turning it into cost needs an admitted price for both cards from one source. GPUDeck has one for RTX 4080 and none for RTX 5090, so the cost pair is not published rather than estimated — the gap is the honest state of the record, not an omission. The value board shows what a price does exist for.
The same hardware pair can also flip on a different workload — which is why GPUDeck never averages workloads together and never publishes a single “best GPU”. A ranking would be a lie with better typography. What you get instead: the per-workload record, dated, with sample sizes, and every number downloadable so you can check the arithmetic.