6-panel comic page

Plot the page, draw every panel at FLUX quality, then fix the details in edit.

The pipeline

Plot the page and write dialogue, qwen3-32b
Draw 6 panels, flux-1-dev
Clean up each panel, flux-1-kontext-dev

Fastest card where every stage is a real measurement: NVIDIA B300, 2 min for the whole job.

Every GPU, slowest job to fastest

GPU	Total time	Compute	Energy	Basis
NVIDIA B300	2 min	61 s	15.52 Wh	all 3 stages measured
NVIDIA B100	2 min	2 min	n/a	anchored estimate (0/3 stages measured)
NVIDIA GH200 Grace Hopper	2.1 min	2.1 min	n/a	anchored estimate (0/3 stages measured)
NVIDIA H100 NVL	2.2 min	2.2 min	n/a	anchored estimate (0/3 stages measured)
NVIDIA H800 80GB	2.2 min	2.2 min	n/a	anchored estimate (0/3 stages measured)
NVIDIA B200	2.3 min	1.7 min	23.05 Wh	all 3 stages measured
NVIDIA H100 PCIe	2.7 min	2.7 min	n/a	anchored estimate (0/3 stages measured)
NVIDIA H100 80GB HBM3	3 min	2.2 min	23.89 Wh	all 3 stages measured
NVIDIA H200	3.2 min	2.1 min	23.21 Wh	all 3 stages measured
NVIDIA RTX PRO 6000 Blackwell Server Edition	3.6 min	3.1 min	26.83 Wh	all 3 stages measured
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition	3.8 min	3.8 min	n/a	anchored estimate (0/3 stages measured)
NVIDIA A800 80GB	4.7 min	4.7 min	n/a	anchored estimate (0/3 stages measured)
NVIDIA A100 40GB SXM4	4.8 min	4.7 min	1.04 Wh	anchored estimate (1/3 stages measured)
NVIDIA RTX PRO 6000 Blackwell Workstation Edition	5.1 min	3 min	28.55 Wh	all 3 stages measured
NVIDIA A100 40GB PCIe	5.1 min	5.1 min	n/a	anchored estimate (0/3 stages measured)
NVIDIA RTX PRO 5000 Blackwell	5.6 min	5 min	24.38 Wh	all 3 stages measured
NVIDIA A100 80GB PCIe	5.9 min	5 min	24.31 Wh	all 3 stages measured
NVIDIA A100 80GB SXM4	6 min	4.7 min	29.8 Wh	all 3 stages measured
NVIDIA L40S	6.5 min	5.4 min	30.3 Wh	all 3 stages measured
NVIDIA L40	9.1 min	8.2 min	39.9 Wh	all 3 stages measured
NVIDIA RTX 6000 Ada Generation	9.3 min	8.8 min	43.46 Wh	all 3 stages measured
NVIDIA RTX PRO 4500 Blackwell	11.6 min	11.4 min	26.25 Wh	all 3 stages measured
NVIDIA RTX A6000	12.1 min	8.9 min	42.76 Wh	all 3 stages measured
NVIDIA RTX 5000 Ada Generation	12.8 min	12.7 min	35.55 Wh	all 3 stages measured
NVIDIA RTX 5880 Ada Generation	14 min	14 min	n/a	anchored estimate (0/3 stages measured)
AMD Radeon Pro W7900	17 min	17 min	n/a	anchored estimate (0/3 stages measured)
AMD Radeon Pro W7800	27.7 min	27.7 min	n/a	anchored estimate (0/3 stages measured)
AMD Radeon Pro W6800	31.5 min	31.5 min	n/a	anchored estimate (0/3 stages measured)

Cards that can't run this pipeline

Every one of these fails on the same kind of wall, a stage that will not fit in VRAM.

AMD Radeon RX 6700, blocked at "Plot the page and write dialogue" (Qwen3 32B)
AMD Radeon RX 6800, blocked at "Plot the page and write dialogue" (Qwen3 32B)
AMD Radeon RX 6800 XT, blocked at "Plot the page and write dialogue" (Qwen3 32B)
AMD Radeon RX 6900 XT, blocked at "Plot the page and write dialogue" (Qwen3 32B)
AMD Radeon RX 6950 XT, blocked at "Plot the page and write dialogue" (Qwen3 32B)
AMD Radeon RX 7600, blocked at "Plot the page and write dialogue" (Qwen3 32B)
AMD Radeon RX 7700 XT, blocked at "Plot the page and write dialogue" (Qwen3 32B)
AMD Radeon RX 7800 XT, blocked at "Plot the page and write dialogue" (Qwen3 32B)
AMD Radeon RX 7900 XT, blocked at "Draw 6 panels" (FLUX.1 dev)
AMD Radeon RX 7900 XTX, blocked at "Draw 6 panels" (FLUX.1 dev)
AMD Radeon RX 9070, blocked at "Plot the page and write dialogue" (Qwen3 32B)
AMD Radeon RX 9070 XT, blocked at "Plot the page and write dialogue" (Qwen3 32B)
GeForce GTX 1080 Ti, blocked at "Plot the page and write dialogue" (Qwen3 32B)
GeForce RTX 4060, blocked at "Plot the page and write dialogue" (Qwen3 32B)
GeForce RTX 4080 Super, blocked at "Plot the page and write dialogue" (Qwen3 32B)
GeForce RTX 5060 Ti, blocked at "Plot the page and write dialogue" (Qwen3 32B)
GeForce RTX 5070, blocked at "Plot the page and write dialogue" (Qwen3 32B)
GeForce RTX 5070 Ti, blocked at "Plot the page and write dialogue" (Qwen3 32B)
GeForce RTX 5080, blocked at "Plot the page and write dialogue" (Qwen3 32B)
Intel Arc A750, blocked at "Plot the page and write dialogue" (Qwen3 32B)
Intel Arc A770 Limited Edition, blocked at "Plot the page and write dialogue" (Qwen3 32B)
Intel Arc B580, blocked at "Plot the page and write dialogue" (Qwen3 32B)
Intel Arc Pro A60, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA A10G, blocked at "Draw 6 panels" (FLUX.1 dev)
NVIDIA GeForce GTX 1070 Ti, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce GTX 1080, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce GTX 1660, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce GTX 1660 Super, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce GTX 1660 Ti, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 2060, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 2060 Super, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 2070, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 2070 SUPER, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 2080 Founders Edition, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 2080 Super, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 2080 Ti Founders Edition, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 3050, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 3060, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 3060 Ti, blocked at "Plot the page and write dialogue" (Qwen3 32B)
NVIDIA GeForce RTX 3070 Founders Edition, blocked at "Plot the page and write dialogue" (Qwen3 32B)

…and 30 more.

How these numbers are built

Each stage time is the quantity of work divided by that card's measured throughput for that model, from our own bench. The pipeline is assumed to run batched, every image, then every clip, so each model loads once. Model load time is added where we recorded it; our text-generation runs don't carry a load measurement yet, so pipelines with a language-model stage are slightly optimistic. Nothing here is a single timed run of the whole pipeline, and we don't present it as one.