Hardware and model compatibility

One row per measured run: what was asked of the machine, whether it finished, and how much memory it took to get there.

3 runs · 1 experiment records · 1 model configurations · 0 marked stale

This table is not carrying real measurements yet. 3 of 3 rows come from a placeholder record used to build and test the site. They are labelled Sample and will be replaced by measured runs.

Showing 3 of 3 records

One row per measured run. Records older than 90 days keep their place and are marked stale.
Hardware	Task	Model	Config	Outcome	Peak VRAM	Speed	Runtime	Verified	Grade
RTX 4070 Ti CUDA GPU · 12 GB	Video generation	Wan2.2 I2V A14B Q4_K_M	832x480 · 81 frames · 20 steps · 832x480, tiling off	Success	10.4 GB	268 s	ComfyUI 0.3.60 torch 2.8.0+cu129	Aug 27, 2026	Directly tested Sample
RTX 4070 Ti CUDA GPU · 12 GB	Video generation	Wan2.2 I2V A14B Q4_K_M	832x480 · 81 frames · 20 steps · 1280x720, tiling off	Out of memory	11.9 GB	n/a	ComfyUI 0.3.60 torch 2.8.0+cu129	Aug 27, 2026	Directly tested Sample
RTX 4070 Ti CUDA GPU · 12 GB	Video generation	Wan2.2 I2V A14B Q4_K_M	832x480 · 81 frames · 20 steps · 1280x720, tiling on	Success	9.1 GB	611 s	ComfyUI 0.3.60 torch 2.8.0+cu129	Aug 27, 2026	Directly tested Sample
No records match these filters. Reset the filters, or check whether the hardware you are looking for has been tested yet.

How to read this table

What the outcome column records

A run either finished, ran out of memory, errored, or finished with quality degraded enough to matter. Out-of-memory rows are kept deliberately: for a 12GB card, the ceiling is more useful information than the successes below it.

Grades

Directly tested , Reproduced , Unverified report . Everything currently in the table is directly tested on hardware in this lab. Reproduced and unverified-report grades exist for records that come from elsewhere, and they are labelled rather than mixed in silently.

Stale after 90 days

Drivers, runtimes and model weights move. A record more than 90 days past its last verification is marked stale and stays in the table. Nothing is deleted for being old.

Where the rows come from

Each row is one result inside an experiment record in data/experiments/, published with a fixed seed and the runtime versions it ran against. The records themselves are in the public repository.

Filtering happens in your browser. There is no server behind this page, so nothing about your query leaves the machine. The measurement rules are in the methodology, and mistakes get fixed in public through corrections.