- Joined
- Jun 20, 2001
- Location
- Vancouver, WA
I hadn't a clue that Whichesters had this kind of a problem (since I don't follows CPUs nearly as much as I should ), but has anybody brought up the possibility of a tweak to the Athlon64's FPU causing the difference? It may very well be stable even with the error, and that Prime95 is just too sensitive (read on if you think I'm loosing my marbles )
As I'm sure you all know, floating point math is inherently error prone. Rounding errors from loss of precision creep in all the time, though are usually kept below notice from additional bits (for example, my calculator will display numbers with 10 digits of precision, but internally does all calculations with 13 digits to prevent errors from being noticible).
The math is still deterministic since the same error will be introduced every time, though it isn't nesscaraly deterministc across kinds of chips. Every manufacturer implements the FPU logic a little different from everyone else, meaning that the way each chip creeps in error will be different. Could it be that Prime95 actually causes the precision error to bleed into it's results, and thus fail when it encounters a different precision error because of different FPU implemetation? It wouldn't really be a sign of instability (since would be the fault of the creators of Prime95 that they don't use precice enough variables), but would still cause it to fail from "rounding errors" and such....
Just random ponderings...
JigPu
As I'm sure you all know, floating point math is inherently error prone. Rounding errors from loss of precision creep in all the time, though are usually kept below notice from additional bits (for example, my calculator will display numbers with 10 digits of precision, but internally does all calculations with 13 digits to prevent errors from being noticible).
The math is still deterministic since the same error will be introduced every time, though it isn't nesscaraly deterministc across kinds of chips. Every manufacturer implements the FPU logic a little different from everyone else, meaning that the way each chip creeps in error will be different. Could it be that Prime95 actually causes the precision error to bleed into it's results, and thus fail when it encounters a different precision error because of different FPU implemetation? It wouldn't really be a sign of instability (since would be the fault of the creators of Prime95 that they don't use precice enough variables), but would still cause it to fail from "rounding errors" and such....
Just random ponderings...
JigPu