Roundoff Error Penalty

[nevarcds]

Does anyone know how much of a performance hit is caused by the roundoff error checking? I know that sums of the sequence before and after the DWT are taken to insure that the floating-point truncation is not affecting the result.

Do the benchmarks listed on the GIMPS site take this into account?

Thanks!
Nevarcds

[Prime95]

Off the top of my head, on a P4 it is pretty small - maybe 1%. On a P3, it was pretty big - about 3-5%. Athlon - I don't know.

It is pretty easy to check. Use Advanced/Time with and without error checking enabled.

[Danath]

But how reasonable is to use round-off checking? For example, if someone wants to use a PC in GIMPS, but has suspicions about the stability of that computer (I mean its capability to produce good results), will Prime95, by enabling the round-off checking, catch the round-off errors (and turn back to a previous good intermidiate result), thus making the final result to be a good one?

[Prime95]

Yes, any error that round-off checking discovers will be corrected. Unfortunately, round-off checking will not catch every hardware error.

Your best bet is to resolve any hardware problems first, then begin prime searching.

[ewmayer]

Originally Posted by Prime95

Off the top of my head, on a P4 it is pretty small - maybe 1%.

Since 1% is lower than the average LL test error rate, wouldn't it make sense to have error checking always enabled on the P4?

[cheesehead]

Originally Posted by ewmayer

Since 1% is lower than the average LL test error rate, wouldn't it make sense to have error checking always enabled on the P4?

But is the average LL test error rate on P4s the same as the error rate on non-P4s?

Have we seen any analysis of error rate vs. CPU type?