GPUDECK · WORKLOAD EVIDENCE

krea2_turbo_fp8_scaled

A paid production workload on the reference fleet, measured on GeForce RTX 4080 and GeForce RTX 5090 on the same days. One workload, its own cells — never averaged with any other workload, because cross-workload aggregation is the substitution error this project exists to expose.

4 ratio cell(s) · window 2026-08-21 → 2026-08-24 · page generated 2026-08-26 11:10 UTC

Dated ratio cells

DAYGOODPUT GeForce RTX 4080 / GeForce RTX 5090ATTEMPTSCOMPLETIONSCARDS
2026-08-210.40×455 / 937416 / 9072 / 6
2026-08-220.30×474 / 1364419 / 13222 / 6
2026-08-230.42×628 / 1131582 / 10992 / 6
2026-08-240.58×60 / 19259 / 1852 / 6

Cost per 1,000 completed, where priced

DEPLOYMENT$/1000 COMPLETEDPRICE $/HRSAMPLEPRICE PROVENANCE
GeForce RTX 5090$1.79$0.651006 attemptsafhubbard/gpu-prices (CC BY 4.0)…
RTX PRO 4500 Blackwell$3.29$0.72336 attemptsafhubbard/gpu-prices (CC BY 4.0)…
Confounds travel with the number. † marks a host confound. Attempts share a card, host and queue - replication is physical cards and card-days, never attempt count.

The measured pair page for these deployments: side-by-side cells. Downloads: ratio cells, value board.