RIP DCTF.

[chalsall]

Originally Posted by NickOfTime

Well, mine is even poorer with 2/625 at TF 75...
8/375 TF 74

What?!?!?! Something is wrong there (or you're being exceptionally unlucky).

Even after a P-1 run, you should still see something like 1/85 to 1/90 or so from 74 to 75.

[chalsall]

Originally Posted by Gordon

Didn't I read somewhere on here that as bit depth increases, odds of finding a factor decrease? Or is my memory playing up again...

You are correct. A "back of the envelope guestimate" often used around here is ~ 1/ [next bit level]. Probability is slightly lower if a P-1 has already run.

Thanks for running the self-test. Clearly that card is good.

[chalsall]

Just a quick query against the GPU72 database wrt 74 to 75 TF'ing.

6,745 runs, 83 factors found. ~ 1 / 81.3. Most of these were done after a P-1 run.

I was always taught to never ignore things which make you go "Hmmmm... That's strange...". Often leads nowhere; sometimes leads to places important.

[Gordon]

Originally Posted by chalsall

You are correct. A "back of the envelope guestimate" often used around here is ~ 1/ [next bit level]. Probability is slightly lower if a P-1 has already run.

Thanks for running the self-test. Clearly that card is good.

GTX-660 also passed all 20,262 self tests

[TheMawn]

Where is this self-test? I'm 0 for ~230 on 75 bits

[Xyzzy]

We worry (a lot) about the possibility of some sort of error causing our cards to miss factors.

One thing we are monitoring is the GHz-days to find a factor. For each higher bit level it should (?) take twice as many GHz-days, right?

Note in the image below that a factor at 70 bits takes 210.6 GHz-days. Then at 71 bits it takes 361.8 GHz-days. Then at 72 bits it takes 796.3 GHz-days. Then at 73 bits it takes 1,666.9 GHz-days. And finally at 74 bits it takes 1,901.9 GHz-days.

So ~200/~400/~800/~1,600/~1,900 means that we are doing better than expected on the 74 bit work?

(We could be wrong!)

[Xyzzy]

It looks like the doubling of GHz-days applies to DC TF work as well. (Roughly, of course!)

85.8/173.9/259.8/568.3

[kladner]

Originally Posted by TheMawn

Where is this self-test? I'm 0 for ~230 on 75 bits

mfaktc -st
mfaktc -st2

mfaktc -h

[VBCurtis]

Originally Posted by Xyzzy

It looks like the doubling of GHz-days applies to DC TF work as well. (Roughly, of course!)

85.8/173.9/259.8/568.3

Yes. This doubling is rough, but a result of two bits of math: Each bit level is twice as big as the one before it, so it takes twice as long to check the next bit level; second, chance to find a factor is roughly 1/bitdepth per bit. So, each higher bit is slightly less likely to find a factor, while taking twice as long. P-1 tests find some factors that you "would have found", so the actual results are less than 1/75 for 74-75 bits in practice.

Of course, the P-1 effect is roughly the same for 73-74 and 74-75, so the doubling of Ghz-days per factor should still be seen in the data.

[axn]

Originally Posted by VBCurtis

Of course, the P-1 effect is roughly the same for 73-74 and 74-75, so the doubling of Ghz-days per factor should still be seen in the data.

P-1 should become less effective for larger numbers. Just how large a drop it is from 73-74 to 74-75, I don't know. Certainly not enough to significantly affect the "doubling" phenomenon. In fact, the direction of this effect compensates for the drop in TF probability.

[Gordon]

Originally Posted by kladner

mfaktc -st
mfaktc -st2

mfaktc -h

If you have more than one card you still need the -d n option to tell it which card to test