GPUDECK · MEASURED PAIR · NO WINNER

GeForce RTX 4080 vs GeForce RTX 5090, one workload at a time

These two cards ran the same paid production workloads on the same days, so their goodput can be compared as a ratio inside each workload cell. The ratio is stated as GeForce RTX 4080 over GeForce RTX 5090 and changes with the workload — which is the finding. There is no single number, no winner, and no ranking.

42 published cells · 4 observation day(s) (2026-08-21 → 2026-08-24) · 42 meet the replication floor · page generated 2026-08-26 11:10 UTC

Every published cell

WORKLOADDAYGOODPUT GeForce RTX 4080 / GeForce RTX 5090ATTEMPTS A/BCARDS A/BCARD-DAYS A/B
chroma-v48-detail-svd_fp82026-08-210.39×42/1552/62/6
chroma-v48-detail-svd_fp82026-08-220.25×25/1872/62/6
chroma-v48-detail-svd_fp82026-08-230.48×55/1572/62/6
chroma-v48-detail-svd_fp82026-08-240.48×40/912/42/4
chroma1-hd_fp8_scaled2026-08-210.34×65/1042/62/6
chroma1-hd_fp8_scaled2026-08-220.28×38/862/62/6
chroma1-hd_fp8_scaled2026-08-230.53×34/542/62/6
chroma1-hd_fp8_scaled2026-08-240.41×63/702/42/4
dark_beast_krea2_fp82026-08-210.37×308/13122/62/6
dark_beast_krea2_fp82026-08-220.33×205/13312/62/6
dark_beast_krea2_fp82026-08-230.38×406/16932/62/6
dark_beast_krea2_fp82026-08-240.59×76/2942/62/6
dark_beast_krea2_identity_edit_v1_22026-08-210.46×120/3902/62/6
dark_beast_krea2_identity_edit_v1_22026-08-220.44×177/5182/62/6
dark_beast_krea2_identity_edit_v1_22026-08-230.43×131/4172/62/6
dark_beast_krea2_identity_edit_v1_22026-08-240.44×144/582/62/6
dark_beast_z_image_turbo_v9_bf162026-08-210.44×80/2902/62/6
dark_beast_z_image_turbo_v9_bf162026-08-220.34×113/4242/62/6
dark_beast_z_image_turbo_v9_bf162026-08-230.33×108/2952/62/6
dark_beast_z_image_turbo_v9_bf162026-08-240.30×95/2522/42/4
krea2_identity_edit_sogni_v0_3_alpha2026-08-210.33×24/572/62/6
krea2_identity_edit_sogni_v0_3_alpha2026-08-220.37×32/852/62/6
krea2_identity_edit_sogni_v0_3_alpha2026-08-230.37×86/1532/62/6
krea2_identity_edit_sogni_v0_3_alpha2026-08-240.51×18/342/52/5
krea2_identity_edit_v1_22026-08-210.33×256/7622/62/6
krea2_identity_edit_v1_22026-08-220.34×244/8132/62/6
krea2_identity_edit_v1_22026-08-230.42×165/7182/62/6
krea2_identity_edit_v1_22026-08-240.56×41/1232/62/6
krea2_turbo_fp8_scaled2026-08-210.40×455/9372/62/6
krea2_turbo_fp8_scaled2026-08-220.30×474/13642/62/6
krea2_turbo_fp8_scaled2026-08-230.42×628/11312/62/6
krea2_turbo_fp8_scaled2026-08-240.58×60/1922/62/6
one_obsession_v22_fp162026-08-210.21×229/1812/62/6
one_obsession_v22_fp162026-08-220.20×244/4532/62/6
one_obsession_v22_fp162026-08-230.34×88/1922/62/6
one_obsession_v22_fp162026-08-240.35×101/1242/42/4
z_image_bf162026-08-210.51×12/152/42/4
z_image_bf162026-08-220.51×12/152/42/4
z_image_bf162026-08-230.45×39/332/62/6
z_image_bf162026-08-240.51×40/222/62/6
z_image_turbo_bf162026-08-230.31×25/722/62/6
z_image_turbo_bf162026-08-240.28×32/572/42/4
What this cannot isolate. A ratio compares two deployments - card, host, virtualisation layer and dispatch together. Rows marked † carry a known host confound and are excluded from headline ranges. Attempts are not independent observations; replication is counted in physical cards and card-days (methodology §7e).

Bring your own hourly quotes to /quote to turn any of these ratios into a break-even price for your workload. Raw cells: gpu-goodput-ratios.csv.