Offloading, and what it trades
Offloading keeps part of a model in system memory and moves it onto the card only when needed. It lowers what the GPU has to hold and lengthens the run, and the size of both effects depends on the machine. As of 2026-09-22.
| Aspect | Transfers to another machine |
|---|---|
| The memory saved | Roughly |
| The time cost | No |
| Whether it is worth it | No |
| That the lever exists | Yes |
Inclusion rule. Aspects of an offload statement. An aspect that does not transfer is listed, because quoting the time cost across machines is the common error. Order. From the aspect that transfers best to the one that transfers worst.
1Both figures are worth publishing
One release in this register states the memory with offload enabled and with the call left out, for the same output. A reader with a small card learns it fits; a reader with a large one learns the run will be faster.
Either figure alone would be accurate and misleading. The friendlier number would mislead anyone budgeting throughput; the larger one would discourage users the release can serve.
2The time cost is not a property of the model
How much slower an offloaded run is depends on host memory bandwidth and on the link between host and card. Two machines with the same GPU can differ substantially.
That is why the register records the memory figures and not a slowdown factor. A slowdown quoted from one vendor's machine would not describe anybody else's.
3Where the trade is visible here
One family publishes both sides of it for the same output, and one discusses the technique without ever giving a figure. The second case still tells a reader that memory was a consideration, and nothing more.
No release in the register publishes a slowdown factor, so what the memory costs in time is not recorded anywhere here. It depends on host bandwidth, which is a property of a machine rather than of a release.
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: Numeric precision, Model sharding.