Numeric precision, and the memory it decides
Numeric precision is the format each weight is stored and computed in. Halving it roughly halves the weights in memory, and it also decides whether a particular card supports the run at all. As of 2026-09-22.
| Effect | Direction |
|---|---|
| Weights in memory | Down, roughly proportionally |
| Card eligibility | Narrower, not wider |
| Output | Changed, by a model-dependent amount |
| Parameter count | Unchanged |
Inclusion rule. Effects of changing the numeric format. An effect that does not occur is listed, because it is the one people expect. Order. From the effect on memory outward to the effect on the count.
1A figure without a format is incomplete
One repository in this register states from 4GB in FP16 and from 5GB in BF16 for two of its releases. The formats are part of the values, and dropping them would make both figures unusable.
Most rows here do not name a format. That is recorded as it stands rather than assumed, because assuming half precision would put a convention into the register that the vendor did not state.
2Lower precision narrows the hardware
An fp8 build needs a card whose hardware supports fp8, so the cheaper configuration is available on fewer machines rather than more. One repository states its card generations and formats in the same sentence for this reason.
A card with plenty of memory and an older architecture satisfies the number and not the requirement, which is the failure mode a bare gigabyte figure invites.
3How often the format is stated here
Four families name a numeric format beside a memory figure, and one of those states two different formats for two of its own releases. The rest publish figures with the format left implicit.
A figure whose format is unstated cannot be adjusted, because halving the bytes per parameter is the largest single lever on the weights footprint. The register records the omission rather than filling it.
A reading note, not an entry: no vendor value appears on this page. Where the register records this term for a particular release, it is on the hardware column and on that family's own page. Nearby terms: Model sharding, GPU architecture.