Home › Method
Method & sources
Cost per video = benchmark seconds × number of GPUs × price per GPU-hour ÷ 3600, computed with exact decimals.
- Timings: only from the model makers' own READMEs. Wan publishes its tables as images; we transcribed them and store the image hash. If the image changes, the daily check flags it for re-transcription. Where a README gives a bound ("within 75 seconds"), the cost is shown as ≤.
- Prices: Runpod list prices from its pricing page, both tabs. Vast.ai marketplace median and minimum from its public offers API (1 GPU, on-demand, verified hosts, dph_total). Multi-GPU cost assumes N × the single-GPU hourly rate.
- Matching: an offer must be the benchmarked GPU class with at least the benchmark's peak memory per GPU. Rows labelled A100/A800 or H100/H800 are priced with A100/H100 offers, and the SXM/PCIe variant isn't stated in the benchmarks. H20 has no tracked offer.
- Not included: setup, model download, storage, idle time, failed runs, and your own sampler or step settings.
- Pending: HunyuanVideo (13B), LTX-Video and LTX-2 timings. We found no first-party timing table, and we don't use community numbers.
Sources
- W22: https://github.com/Wan-Video/Wan2.2#computational-efficiency-on-different-gpus (table image sha256 75ee012dcfb0…)
- W21: https://github.com/Wan-Video/Wan2.1#computational-efficiency-on-different-gpus (table image sha256 b0e225caffb4…)
- HY15: https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5
- HY: https://github.com/Tencent-Hunyuan/HunyuanVideo
- Runpod: https://www.runpod.io/pricing
- Vast.ai API: https://console.vast.ai/api/v0/bundles/