Source
Mohan et al. (2020), LIT-050 — ARXIV-2007.06775.
What this practice would mean
Decoding and augmenting into half precision on the host, so that what crosses the bus is half the bytes and no cast is needed on the far side. Where the step already runs in mixed precision (SOTA-016) the conversion has to happen somewhere, and doing it before the transfer rather than after is free.
The source does not carry this one
LIT-050’s contribution is the measurement of data stalls and three mitigations in CoorDL — caching, coordinated prefetch across jobs, and avoiding redundant preprocessing. Half-precision decode is not among them, and the note’s own summary of the paper does not mention it.
The recommendation may well be right; what it lacks is this paper as evidence for it. Recorded as a citation to check rather than a body making a case, because the alternative is a body that reasons from a source that does not say what it is credited with — which is the failure the whole worklist in #107 exists to find.
Two things would settle it: a source that actually measures the transfer saving, or the observation that on a pipeline whose stall is decode-bound rather than bandwidth-bound the practice buys nothing, in which case the condition belongs in the title.