It really isn't noticeable until about 5k UPDATEs on a single row.
Don't know why, and never dealt with a scenario where this would even come up, but that this doesn't perform well inside a transaction isn't surprising to me. Kinda surprised it works well at all actually. I'd probably try and rework the processing algorithm to create an unlogged temporary table with data from the row's initial state, manipulate until my heart's content, then take the final result and update the single live row with the final state.
