Z-Image Turbo, 4 images at 1024×1024
- RTX PRO 6000 96 GB, $2.42/h
- 6.5 s · $0.004
- RTX 6000 Ada 48 GB, $1.37/h
- 17.7 s · $0.007
2.7 times faster, and 35% less for 4 images.
The dearer card was cheaper per output.
On 20 October 2026, the RTX PRO 6000 finished three AI jobs 2.7 to 3.5 times sooner than the RTX 6000 Ada, so each output cost 35 to 50% less, though an hour on it cost 1.8 times as much. We ran the jobs on rented cards in Finland.
Faster
2.7–3.5×
The RTX PRO 6000 against the RTX 6000 Ada, job by job.
Measured 20 October 2026
Cheaper per output
35–50%
Less per output on the faster card, though its hour costs more.
Measured 20 October 2026
Price an hour
1.8×
$2.42 against $1.37, each card's on-demand list price on the day.
Read 20 October 2026
Our memory estimate
Within 1%
For LLMs, against the peak memory we measured.
Measured 20 October 2026
Why the dearer card cost less
Cost per output, against the RTX 6000 Ada
Measured 20 October 2026
Three jobs, two cards
Time from submitting the job to its last output. Cost is that time at each card's on-demand list price before VAT, with no start-up, storage or idle time.
Wan 2.2 5B, a 5-second 720p video
- RTX PRO 6000 96 GB, $2.42/h
- 74 s · $0.05
- RTX 6000 Ada 48 GB, $1.37/h
- 200 s · $0.076
2.7 times faster, and 35% less for one video.
A 70B LLM (Q4), 32 answers, 8,000 tokens out
- RTX PRO 6000 96 GB, $2.42/h
- 49 s · $0.033
- RTX 6000 Ada 48 GB, $1.37/h
- 172 s · $0.065
3.5 times faster, and 50% less for 32 answers.
Measured 20 October 2026
Our memory maths against the real peak
GPU memory sampled every 250 ms while the job ran, next to what the site's formula says.
| Job | Card | Our estimate | Measured peak | Our estimate / Measured peak |
|---|---|---|---|---|
| 70B LLM (Q4), 32k context | RTX PRO 6000 | 53.8 GB | 53.6 GB | |
| 70B LLM (Q4), 8k context | RTX 6000 Ada | 45.8 GB | 45.4 GB |
Measured 20 October 2026
Image and video jobs need working memory on top of their weights. Four 1024×1024 images on Z-Image Turbo peaked at 15.2 to 15.4 GB against a largest stage of 12.7 GB; a 720p Wan 2.2 5B video (121 frames) took about 31 GB with every model loaded. Read the checkers' verdicts with that margin in mind.
How we measure
Every number on the site is one of four kinds, and the page says which, with its date.
Measured
We rent the card and run the job. The runs of 20 October 2026, in Finland, used ComfyUI v0.42.1 on PyTorch 2.14.1 with CUDA 13.0 for images and video, and llama.cpp for the LLM, with models from ComfyUI's official templates and Hugging Face. GPU memory was sampled every 250 ms.
Computed
Memory from the model's own config.json and published file sizes, with the formula on the page. For LLMs it landed within 1% of the peaks we measured; image and video jobs need working memory on top, which the pages say.
Read
Prices, plans and limits from each provider's own pages, with the date we read them.
Yours
Anything you type into a tool.
A measurement holds for the card, software and settings it names. We re-run jobs as models, software and prices change, and each page shows the date of its numbers.