Method & measurement
How tests are run, what the speed metrics mean, how evidence is classified, and how captures are collected.
Protocol and measurement
Each report records the device, model file, quantisation, framework version, runtime parameters and test date. Commands, raw logs, metrics and captures share a run ID. Missing conditions are labelled rather than inferred.
Current reports use llama-bench prompt-processing and token-generation metrics, with -p 512 -n 128. The benchmark excludes tokenisation and sampling time, so tokens/s is not end-to-end response speed in a chat interface.
Reports distinguish failure to load from slow or unstable behaviour after loading. A single run does not describe every time of day, driver version or GPU with the same marketing name. Conclusions apply only to the hardware and context range actually recorded.
Evidence levels
- Tested here：Linked to a run ID, configuration and raw records from this site.
- External test：Measured by another source; cited where used and kept separate from our own results.
- Estimate：Derived from a stated formula or model, not a measurement.
- Not tested：No verifiable result is available; we do not interpolate and label it a test.
Current test hardware and capture environment
The current primary test device is NVIDIA GeForce RTX 3060 Laptop GPU (6 GB VRAM) with 64 GB system memory. Each experiment records CPU, driver, framework and parameters separately. Do not treat this laptop’s results as results for a desktop RTX 3060 or another device.
Desktop captures and metrics in the current report come from the same run. Capture used an Xvfb virtual display under WSL; it was not a physical desktop session. Screenshots document the process, while raw logs and structured data are the basis for checking numbers.
Corrections and reproduction
If you spot an error, identify the page, metric and run ID. Report updates keep a revision record. For reproduction, compare the model file, software version, power mode, context and offload parameters.