Peak memory, and why it is not a minimum

Peak memory is the most a run held at once on the card it was measured on. It establishes that the machine finished with room to spare. It does not establish that a smaller card would have finished at all. As of 2026-09-22.

Inclusion rule. Components of a complete peak-memory statement. A component that is frequently missing is listed, because the gap is what makes a figure unusable. Order. From the part always published to the part most often left out.

1Why vendors publish peaks

A peak falls out of instrumentation. Running a script and reporting the maximum allocated is one line of code; establishing the smallest card that completes the job means testing hardware the vendor may not own.

So the field is full of peaks, and the reader who wants a floor has to notice which kind of figure they are looking at. The register keeps the word peak in the value for that reason.

2A peak can still be planned with

On the same class of card it is a good capacity number: it says how much headroom a run leaves, which decides whether two streams fit on one machine.

It transfers downward badly. A run that peaked at 60GB on an 80GB card may well need a different configuration, or a different offload strategy, to fit anywhere smaller.

3How many cells here hold a peak

Four families publish a peak measurement, four publish a stated minimum, two publish a requirement figure, one publishes the memory a run used, one publishes only a timing and three publish nothing.

The register keeps the vendor's own word inside the value, so a cell that says peak says peak. Normalising the wording would make four rows look like floors they are not.

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: Memory floor, Offloading.