Why naming the GPU beats stating a memory figure
Eight of the 15 families here name hardware rather than only memory: a consumer 4090, a single 80G GPU, an H100, an A100 80GB, the RTX 30XX to 50XX series, eight H100s or one 4090, an H100 or H800, and three NVIDIA architectures. A named card can be priced in a minute. As of 2026-09-22.
| Item | What is published | Detail |
|---|---|---|
| MAGI-1 | H100 or H800 x 8, or one RTX 4090 | A count of a named card, stated per model size |
| Mochi 1 | At least one H100 | Recommended |
| Open-Sora | H100 or H800 | The cards the peak memory table was measured on |
| Stable Video Diffusion | An A100 80GB card | Named to carry a timing, not a memory floor |
| Wan | A consumer 4090 | Named for the 5B release |
| HunyuanVideo | A single 80G GPU | What the figures were measured on |
| Cosmos Predict 2.5 | Ampere, Hopper and Blackwell | Architectures, beside a 32.54 GB figure |
| FramePack | RTX 30XX, 40XX, 50XX | Generations, with fp16 and bf16 support required |
| Everyone else | Memory figures only, or nothing | The card is left to the reader |
Inclusion rule. Hardware named in the vendor's own repository. A card mentioned in a community thread is not recorded. Order. Families naming a specific card first, then those naming a generation or architecture, then families naming neither.
1Memory alone leaves the expensive decision to the reader
Sixty gigabytes could be one card or two, at very different prices and very different throughput. Naming the card collapses that ambiguity and lets somebody get a per-hour figure before committing anything.
It also makes the vendor's figure reproducible: a reader with the same card can check the claim, which is not true of a bare memory number.
2Measured on and recommended are different claims
Some families say their numbers came off a particular card, others say what they advise, and one states a count of cards the model needs. Measured on is a statement about an experiment, recommended is a statement about an expectation, and a count is a statement about the shape of the machine. Only the first can be repeated.
The register records the wording rather than flattening all three into a requirement column, because the three support different amounts of weight.
3Neither a card nor a figure is a cost
A card name plus a memory figure still leaves out how long a generation takes and how many attempts a usable shot needs. Those two multiply everything, and no repository in this register publishes the second at all.
So the honest summary is that naming a card moves a reader from cannot estimate to can estimate the floor, which is real progress and is not the same as knowing the cost.
- HardwareTested on a single 80G GPUwhat the figures were measured on
- HardwareApproximately 60GB VRAM on a single GPU, with at least one H100 recommendedthe vendor's own figure
- HardwareH100/H800 x 8 for the 24B model, RTX 4090 x 1 for the 4.5Bstated as a count of a named card
- HardwareThis model requires 32.54 GB of GPU VRAM, on Ampere, Hopper or Blackwell hardwarea figure and the supported architectures
- HardwareNvidia GPU in RTX 30XX, 40XX, 50XX series that supports fp16 and bf16; the GTX 10XX/20XX are not testedcard generations and numeric formats
- HardwareSVD-XT takes about 180s on an A100 80GB carda timing on a named card with no memory floor stated
- Hardware60.3GB peak GPU memory at 768px on one GPU, and 44.3GB across four or moremeasured on H100 or H800 cards
- HardwareThe repository states TI2V-5B runs on consumer-grade graphics cards such as a 4090, generating a 5-second 720P video in under 9 minutes on a single consumer GPUvendor's own figure
4Running the model without the hardware
Instead of local hardware for “Naming the card”, SceneMixer runs hosted models; its guides describe the result:
Measured across every job that finished successfully on scenemixer.com in the thirty days to 20 September 2026: analysis of a whole book runs about 2 minutes 49 seconds at the median, one episode's storyboard 1 minute 54 seconds, one video segment 3 minutes 20 seconds, and compositing an episode 3 minutes 52 seconds.
SceneMixer, Turn a Novel into a Video Series with AI — “How long does it take to produce one episode?” (read 2026-10-13)
Wan 3.0 is Alibaba Cloud's video generation model on the Model Studio platform, called through an asynchronous submit-then-poll API, in two tiers: the standard wan3.0-video and the faster wan3.0-video-prime.
SceneMixer, Wan 3.0 for AI Short Drama (2026): Parameters, Reference Limits, Moderation Rules and the 30-Second Take (read 2026-10-13)
5Sources
Repository statements and hosted rates are linked from the model pages and the method page, checked 2026-09-22. Hardware figures are set out on requirements. Nearby: How long a clip, Nobody sells these, Licence vs weights.