20 long-form articles

A month of written content in one run, on the biggest open model that fits your card.

The pipeline

Draft 20 articles (~1,400 tokens each), llama-3-3-70b

Fastest card where every stage is a real measurement: NVIDIA B300, 9.7 min for the whole job.

Every GPU, slowest job to fastest

GPU	Total time	Compute	Energy	Basis
NVIDIA H100 NVL	9.7 min	9.7 min	n/a	anchored estimate (0/1 stages measured)
NVIDIA B300	9.7 min	9.7 min	61.26 Wh	all 1 stages measured
NVIDIA B200	10.5 min	10.5 min	67.6 Wh	all 1 stages measured
NVIDIA GH200 Grace Hopper	10.7 min	10.7 min	n/a	anchored estimate (0/1 stages measured)
NVIDIA H200	10.9 min	10.9 min	39.82 Wh	all 1 stages measured
NVIDIA B100	11 min	11 min	n/a	anchored estimate (0/1 stages measured)
NVIDIA H100 80GB HBM3	11.4 min	11.4 min	55.51 Wh	all 1 stages measured
NVIDIA H800 80GB	11.4 min	11.4 min	n/a	anchored estimate (0/1 stages measured)
NVIDIA RTX PRO 6000 Blackwell Workstation Edition	13.4 min	13.4 min	42.96 Wh	all 1 stages measured
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition	14.1 min	14.1 min	n/a	anchored estimate (0/1 stages measured)
NVIDIA RTX PRO 6000 Blackwell Server Edition	14.5 min	14.5 min	60.83 Wh	all 1 stages measured
NVIDIA RTX PRO 5000 Blackwell	17.8 min	17.8 min	46.68 Wh	all 1 stages measured
NVIDIA H100 PCIe	19 min	19 min	n/a	anchored estimate (0/1 stages measured)
NVIDIA A100 80GB SXM4	19.1 min	19.1 min	55.72 Wh	all 1 stages measured
NVIDIA A800 80GB	19.1 min	19.1 min	n/a	anchored estimate (0/1 stages measured)
NVIDIA A100 80GB PCIe	20.4 min	20.4 min	55.11 Wh	all 1 stages measured
NVIDIA RTX 6000 Ada Generation	25.4 min	25.4 min	89.53 Wh	all 1 stages measured
NVIDIA L40S	28.5 min	28.3 min	122.26 Wh	all 1 stages measured
NVIDIA RTX A6000	29.7 min	29.7 min	56.52 Wh	all 1 stages measured
AMD Radeon Pro W7900	33.1 min	33.1 min	n/a	anchored estimate (0/1 stages measured)
NVIDIA RTX 5880 Ada Generation	34.6 min	34.6 min	n/a	anchored estimate (0/1 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 Pro W6800, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon Pro W7800, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 6700, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 6800, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 6800 XT, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 6900 XT, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 6950 XT, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 7600, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 7700 XT, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 7800 XT, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 7900 XT, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 7900 XTX, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 9070, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
AMD Radeon RX 9070 XT, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
GeForce GTX 1080 Ti, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
GeForce RTX 4060, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
GeForce RTX 4080 Super, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
GeForce RTX 5060 Ti, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
GeForce RTX 5070, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
GeForce RTX 5070 Ti, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
GeForce RTX 5080, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
Intel Arc A750, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
Intel Arc A770 Limited Edition, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
Intel Arc B580, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
Intel Arc Pro A60, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA A100 40GB PCIe, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA A100 40GB SXM4, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama-3.3-70B)
NVIDIA A10G, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce GTX 1070 Ti, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce GTX 1080, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce GTX 1660, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce GTX 1660 Super, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce GTX 1660 Ti, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce RTX 2060, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce RTX 2060 Super, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce RTX 2070, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce RTX 2070 SUPER, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce RTX 2080 Founders Edition, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce RTX 2080 Super, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)
NVIDIA GeForce RTX 2080 Ti Founders Edition, blocked at "Draft 20 articles (~1,400 tokens each)" (Llama 3.3 70B)

…and 38 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.