Check "bad" link. It leads to non-overflowed and validated result.
[
it seems good and bad reversed
Your "good" link contains overflow that didn't confirm by other CUDA host.
I start to think it's PROBLEM with app, not driver.
Please, don't use this version on your host.
XP has no driver restart ability. If kernel launch longer than 3 secs it just will be aborted w/o user-noticeable messages.
If I remember right it will fail silently even not reporting error state (and this is CUDA flaw IMHO). So, try to use app with lower number in name.
IF it will run OK then we should mark "8192" as inappropriate for cards of your type (BTW, what GPU do you use? ).
BTW, it's _beta_ app so I'm not sure it's correct to discuss it in non-beta thread. You could report this issue in beta area.