The first version has to repeatedly scan (once per input line) an ever-increasing pattern space. the s/\n/ / takes longer and longer as input lines accumulate.
The second version slurps the whole file first, and then runs a *single* s/\n/ /g command, which goes over the whole file just once.

cd "$(mktemp -d)"
tr -dc '[:print:]' file.txt
time sed ':begin;$!N;s/\n/ /;tbegin' processed.txt

# real 2m29.278s
# user 2m28.486s
# sys 0m0.022s

and this (https://stackoverflow.com/a/1252191) so fast:

time sed ':begin;N;$!bbegin;s/\n/ /g' | processed.txt

# real 0m0.026s
# user 0m0.013s
# sys 0m0.013s

...good explanation :) I understand what you are saying ofc, but I'd used sed for replacements many times before having to use labels whereas N and P had been used before. Thanks for your responses.

Well, labels and branches are (or should be) basic sed knowledge.
":begin" defines a label, and "t" is the conditional branch command, so "tbegin" (or "t begin") branches to label ":begin" if the last replacement operation was successful. Think of it as a kind of conditional "goto".
The less known commands "N", "P" and "D" instead deserve more explanation in my opinion, since here they are integral to newline matching (while labels and branches are only used as part of the program and are not essential to understand how newlines are actually matched).

Hi, that is (kind of) true - it is after all an article called 'HOW TO match newlines in sed'. Simply explaining your examples more fully would have greatly increased the value of this post. Thanks anyway.