Local AI, measured on hardware you can buy
Every number comes with the machine it ran on, the settings that failed, and the date it was checked.
- Measured runs in the database
- 0
- Experiment records
- 0
- Hardware configurations
- 0
- Most recent verification
- none yet
Latest measurements
All benchmark articlesThe first articles are still in draft. The database below already carries the experiment records they are built on.
Two clusters, one set of machines
Local Video
Video generation on a single consumer GPU: ComfyUI, Wan2.2, memory ceilings, and the settings that survive 12GB of VRAM.
Local LLM
Language models on hardware you already own: quantization trade-offs, context length limits, and tokens per second you can reproduce.
Both clusters run on the same two machines, which is what makes the CUDA versus Apple Silicon comparisons worth anything: the model, the quantization and the context length are held fixed across platforms.
How to read a page here
- Directly tested
- Run on the hardware named in the test environment block on that page.
- Official source
- Taken from vendor or project documentation, linked at the point it is used.
- Inferred
- Derived from measurements taken on a different configuration, and marked as such.
- Not tested
- Reported by others and repeated without reproduction here.
- Stale
- Last verified more than 90 days ago. Stale records stay visible rather than being quietly deleted, because knowing a number is old is more useful than not finding it at all.
The rules behind those labels are in the editorial policy, and the way experiments are run and recorded is in the methodology.