Methodology

Posts here contain two kinds of figure, and they are not treated the same way.

Cited figurescome from someone else — a company’s earnings release, a regulator’s filing, an agency dataset, an analyst estimate. These are attributed inline to whoever published them. Where the source is a company reporting on itself, or an analyst estimating rather than observing, the post says so.

Measured figures were produced on hardware in Kathmandu for the post they appear in. These are labelled as measured, and the conditions are published with them.

Any post containing a measured figure states, at minimum:

• The exact hardware, including memory, since it is usually the binding constraint.
• The software and model versions, and quantisation level where relevant.
• The number of runs behind a reported figure, and whether it is a mean or a single run.
• The network path and time of day, for anything latency-related, since both move the result.

The intent is that a reader with comparable hardware can rerun the test and get a comparable answer. Where a run script exists, it is published with the post.

Measurements taken on a single consumer machine on a single consumer connection are not laboratory results. Thermal throttling, background load, ISP routing, and time of day all affect them. Single-machine numbers are reported as what one machine did on one day, and should not be read as a general benchmark of the hardware or the service.

Where a post compares a measured figure against a cited one — local hardware against a published GPU benchmark, say — the two were not produced under the same conditions, and the comparison is indicative rather than controlled. Posts doing this say so on the page.

If a number here is wrong, it gets fixed and the fix is disclosed. See the corrections policy.