[dahart]

There’s probably a way to get a stable conservative error bound for a parallel sum reduction that’s non-deterministic.

Here’s the PBR book’s chapter on error analysis. It’s way easier to implement error bound tracking than it is to come up with symbolic bounds, and it’s way better than having nothing or even a guess pulled out of thin air. It’s probably worth trying, just to see what happens, it might be surprising. Plus if you implement the error bounds, you may find very interesting and/or enlightening ways to visualize them.

https://pbr-book.org/4ed/Shapes/Managing_Rounding_Error#

[tomsmeding]

Thanks a lot for the insights!