So … have to ask … What difference does this make to me?
Moderately sure most folk here are less interested in the exact internals, and more interested in the end-to-end effect. Or to put different, we do care about internals, but only when we know how that ties into the end result.
As aside … tried an LLM (or three) to look up a vaguely recalled article from long ago, and the answers were wildly wrong. Not sure a few more bits precision would help. 
Seems LLMs are very bad at finding information not directly linked from the web.
Read an article in the early 1980s (published by ACM or IEEE) that drew on experience with early networks. The gist is that end-to-end check is the only one that matters, and intermediate checks are only ever (hoped) performance optimizations.
(The elaborate ISO OSI network was then accepted standard, but over-complicated.)
The graphs are cool, but we all want to know what difference this makes in end.
Also have a speculation: In scientific computing exact fidelity in result is critical (as you do not want to fly into a planet). In compute for LLM, a higher error rate might be accepted as trade for performance, as - frankly - no one could tell. Do not know if that trade is made on recent hardware. Might explain your results.
Nearest reference:
https://web.mit.edu/Saltzer/www/publications/endtoend/endtoend.pdf