Why a pair of figures beats a single number
A single published figure describes one point. Two figures with one variable between them describe a direction, which is what a reader planning a different configuration actually needs, and it is the rarer form of publication. As of 2026-09-12.
| Variable | What the pair reveals |
|---|---|
| Offload state | What memory the technique returns, and at what cost in time |
| Frame size | How much of the figure is activations rather than weights |
| Model size | What the parameter count adds at a fixed output |
| Card count | What sharding actually saves, which is not a division |
Inclusion rule. Variables that can be isolated between two memory figures. Each row is a general measurement shape rather than a vendor value. Order. From the variable a reader controls at run time to the one that needs new hardware.
1The second figure costs one more run
A vendor that has instrumented one configuration can instrument a second for the price of another run. The reason pairs are rare is not cost; it is that publishing a second figure invites questions about the range between them.
That is precisely the range a reader needs. A pair turns a claim into a shape, and a shape can be extrapolated a short distance with some confidence.
2Only a controlled pair is a gradient
Two figures that differ in resolution and model size at once describe two runs. The difference between them cannot be attributed to either variable, and the ratio is a number with no meaning.
Checking what is held constant is therefore the first step with any pair. Where the output is held constant and the model varies, or the card is held constant and the frame varies, the pair is worth something.
3Pairs also bound the extrapolation
A gradient measured between two points says nothing reliable outside them. Doubling a resolution beyond the largest measured configuration is not covered by a ratio taken inside the range.
This is the failure mode a pair invites: a reader who has a direction will follow it further than the measurements support. The register records both points rather than a slope for that reason.
4The sharding pair is the one worth hunting for
Multi-card figures are the least intuitive in this field, because activations and buffers are replicated while only part of the load is split. A single-card figure beside a multi-card one is the only honest answer to what extra cards buy.
Anyone assuming linear scaling will under-budget substantially. It is the one place where the absence of a pair reliably costs money.
5Both sides of a trade belong in public
Where a technique lowers memory and raises runtime, publishing only the lower figure is accurate and misleading. A reader budgeting throughput would plan against a number that assumes the slow path.
Publishing only the higher figure is misleading in the other direction, and discourages users the release can serve. Two figures cost one extra line and remove both errors.
6What to do with a single figure
Treat it as a ceiling on confidence rather than as a specification. Plan at or below the configuration it was measured at, and test before scaling anything.
Where a release publishes one figure and a set of smaller builds instead of a second, the builds are the gradient in a form no column can hold. That is worth reading as an answer even though it is not a value.
7Where the repository figures are kept
The techniques above are general. Which vendors have published what, under which licence and on what hardware, is recorded on the model pages, each figure quoted from the repository it was read from with its date.
- Two figures, one run — who publishes a pair
- Model sharding — what extra cards do not divide
- Offloading — the trade behind the commonest pair
Craft notes on reading releases. No row here is attributed to a vendor, and nothing on this page is a reading of anyone's licence obligations. The sourced material is on the requirements page. Related: Limits of a register, Choosing a build.