Handful more
 

I just checked and these 3 beauties just became available again (assignments expired).

Judging from the bad/good ratio, I'd say these are definitely in the category of things to DC ahead of the curve:
[CODE]exponent Bad Good Unk Sus Solo Mis worktodo
37864429 5 0 2 1 2 1 DoubleCheck=37864429,71,1
40164529 20 3 3 0 1 2 DoubleCheck=40164529,72,1
54881429 32 6 12 14 7 19 DoubleCheck=54881429,73,1[/CODE]

I'll take M37864429 for a DC

[QUOTE=Madpoo;411161]...When I have some time I'll try and pick out a selection of the worst offenders who also did any P-1/ECM/TF work and try to get a feel for how they may have fared with those. I guess things like figuring out how many factors they found out of how many attempts, see if it seems "off".[/QUOTE]

I did a small sampling... 3 machines that met these criteria:[LIST][*]>= 10 bad results[*]0 good results[*]no more outstanding unknown/suspect[*]v5 client only[/LIST]
Of those, 2 of them have only done a small amount of TF work with no factor found.
#1 has 1 factor out of 62 attempts
#2 has 0 factors out of 4 attempts (why is zero plural? It felt weird to say "zero factor found" :smile: )
#3 has 1675 factors out of 109110 attempts

It's the 3rd machine with the large amount of factoring that made me wonder... I guess a factoring 'success rate' of ~1.5% isn't too bad? I don't know.

Could just be that the type of work involved in TF compared to LL isn't as taxing for the system. It had only done 12 total LL tests, all of which were bad.

It had also done a handful of ECM factoring work... 2 factors found out of 296 attempts (all of them were just 3 curves at varying bounds, mostly B1=50000 / B2=5000000)

So even in that worst case looking thing, the rate of factors being found seemed kind of in line with expectations, I think?

[QUOTE=Madpoo;411219]I guess a factoring 'success rate' of ~1.5% isn't too bad?[/QUOTE]
That is exactly one in 65, which fits perfectly "in theory", even with some luck (if he factored to higher bitlevels, then he found more factors than expected!). No action need to be done here. And indeed, GPU TF does not use (lots of) memory, as opposite to GPU LL or GPU P-1, both use (lots of) memory for FFT. One can check the utilization of the memory with GPU-Z, for example, is over 40%-60%, depending on your exponent when LL, and is close to 1%-2% when TF. So, a card can be wonderful at TF and crap at LL, if it has memory problems (just an example).

[QUOTE=Madpoo;411161]Hmm... good question. I'm looking at just the LL results since that's the only thing where we can tell for sure if they were good/bad.

When I have some time I'll try and pick out a selection of the worst offenders who also did any P-1/ECM/TF work and try to get a feel for how they may have fared with those. I guess things like figuring out how many factors they found out of how many attempts, see if it seems "off".[/QUOTE]

[QUOTE=LaurV;411223]That is exactly one in 65, which fits perfectly "in theory", even with some luck (if he factored to higher bitlevels, then he found more factors than expected!). No action need to be done here. And indeed, GPU TF does not use (lots of) memory, as opposite to GPU LL or GPU P-1, both use (lots of) memory for FFT. One can check the utilization of the memory with GPU-Z, for example, is over 40%-60%, depending on your exponent when LL, and is close to 1%-2% when TF. So, a card can be wonderful at TF and crap at LL, if it has memory problems (just an example).[/QUOTE]

Thanks for the sanity check. Yeah, I'd have to say that, on the surface of it, if a machine is lousy at LL, it may still have done just fine at TF.

I know there have been some crazy cases of people doing TF work and missing all kinds of things, and I wonder what special bad-ness was going on to allow that to happen, but it certainly seems atypical.

Unless something comes up later that makes us rethink that concept, I probably won't pursue avenues of trying to find bad factoring machines, just because it's so hard to really tell one way or the other. I've tried a few things now but with inconclusive results. Even when I tried to find machines that did TF work and a factor was later found by P-1 or ECM that *should* have been found by TF, I wasn't really always sure about it, and it wasn't the kind of thing where I could point to a particular set of machines and say "aha!"

Oh well... maybe down the road we can revisit that notion when I (or someone else) has more bandwidth to worry about it. :smile:

[QUOTE=Madpoo;411219]
#2 has 0 factors out of 4 attempts (why is zero plural? It felt weird to say "zero factor found" :smile: )
[/QUOTE]
Change "factors" to "years", and consider which numbers get plural. 0.3 years? 1.1 years? 1.8 year? 0.03 year?

The singular case is only for, you know, singles. Everything else gets plural, as if the item is a unit in science. If I were giving directions, I'd tell someone a bus stop is 0.6 miles away, though I suppose "six tenths of a mile" contradicts all this...

"no factor found" sounds quite "English" to me... :rolleyes:
(not a native speaker, and we won't say it like that in our native language)

Requesting a Quad check to a prime95 CPU result: [url]http://www.mersenne.org/report_exponent/?exp_lo=39635213&full=1[/url]

[QUOTE=LaurV;411223]..... So, a card can be wonderful at TF and crap at LL, if it has memory problems (just an example).[/QUOTE]

[LEFT][B]With consumer (gamer) cards[/B], totally different tuning is necessary to run LL compared to TF.
[/LEFT]

I have experienced truly bogus TF results, as in PrimeNet rejected Factor Found results. This only happened while I had a power supply whose regulation was "Gone with the Wind," or possibly when I had not realized how bad the power connections had become on my GTX 570.

In general, though I have no way to prove accuracy, I have operated on the principle that MFAKTC tends to crash when the GPU is being pushed too hard for the voltage level. Depending on temperatures, a crash indicates either increasing voltage, or reducing frequency.

On the other hand, very conservative GPU and RAM frequencies, possibly combined with higher voltage, when possible, are in order with CUDALucas.

I have never come within fantasizing distance of Tesla cards, but those, as I understand, operate at more conservative frequencies, combined with ECC RAM. So with the more plebeian lines, the best one can do is emulate the lower frequencies, especially regarding VRAM. If your tweaking software's brand matches your card, i.e. Afterburner with MSI cards, you may be able to experiment with VRAM voltage as well.

At least with CuLu, the multitude of tests (2 levels of residue-based self-tests, memory tests, etc,) give you a much better shot at determining potentially viable parameters, before taking the time to do DCs.

In my case, that time, for a 34.7M DC is about 1D 3-5H on the MSI 580 running at Factory OC 833 MHz, with boosted GPU and VRAM voltage, and the RAM throttled back to 1800 MHz. With a GTX 460, I think I ran at 1700 MHz, though I don't remember the GPU frequency. Around that time, one of the heavy hitters in TF, like Flash or Nucleon, said something about running (probably 580s or 570s) memory at 1600MHz.

I have been doing such tests intermittently, between bouts of cranked up TF work, so it usually takes me 2-3 days to finish a 34.7M DC. The last 3, at least, have not mismatched. I had at least one independently confirmed mismatch early on with the 580 (thanks, Wombatman!) and another which did not match, which I reran with more laid back setting to get a match.

IMPORTANT! When switching between work types like these, be sure to change your frequency and voltage settings BEFORE you switch from TF to LL of any sort. It seems likely that TF settings could introduce LL errors very quickly.

Note that when running TF, it is advantageous to slow your memory way down, as in 1500 MHz, possibly even more. You use less power, and can probably run cooler when you are trying to push the GPU core as hard as is practical.

[QUOTE=chalsall;411054]OK, these are now in GPU72 and being assigned.[/QUOTE]

Just so everyone knows, all but three of these have been appropriately TFed; the last three will be finished by this time tomorrow. I haven't checked, but I don't think any factors were found.

[QUOTE=Madpoo;411214]I just checked and these 3 beauties just became available again (assignments expired).

Judging from the bad/good ratio, I'd say these are definitely in the category of things to DC ahead of the curve:
[CODE]exponent Bad Good Unk Sus Solo Mis worktodo
37864429 5 0 2 1 2 1 DoubleCheck=37864429,71,1
40164529 20 3 3 0 1 2 DoubleCheck=40164529,72,1
54881429 32 6 12 14 7 19 DoubleCheck=54881429,73,1[/CODE][/QUOTE]

40164529 has been assigned to an anonymous user - probably a cat 4 churner. ETA 7 days (supposedly).

The last one is still available.