GPU to 72 status...

[LaurV]

Hey Chris, I have some issues with you!

First, there is no 37M available from GPU72 for DC, as we discussed before. I had to take a couple of them directly from PrimeNet to give some work to do to my card. Remember, some FFT with power of 2 which works for 37M expos is about 30% more efficient on GCN cards (which we discussed in the clLucas' thread). Now I switched that card to DCTF so you don't need to take any action, jut let you know. The 37M expos reserved for DC magically disappeared from gpu72 assignemnt page.

Second, last of the exponents from the batch was a mismatch, 37000291, so I did a TC to his majesty and turned out my DC was good and original residue was bad (unless there is a bug in clLucas, because both DC and TC were done with the same software, on the same card, and they matched - edit: I am still keeping the residues till problem solved). So, if someone else is reserving this expo for GPU work, they will work in vain. Therfore I am thinking that the exponent is a very good candidate for your R7 systems, if any of them is free...

Before someone else grabs it...

[flashjh]

If it's ok, I grabbed it to test a newer system I have...

[kracker]

If it please your majesty... a few 38M exponents as well?

[chalsall]

Originally Posted by kracker

If it please your majesty... a few 38M exponents as well?

Please don't even make that joke (and I understand and appreciate that it was a joke).

If anything, I am your servant. You are not a subject...

Let me see what I can do....

[flashjh]

Originally Posted by LaurV

Second, last of the exponents from the batch was a mismatch, 37000291...

Done

[LaurV]

We switched few of our cores to assignments directly from PrimeNet until GPU72 decides to give us the assignment types we requested, i.e. DC, and not P-1, see a former discussion about a rebel core which didn't want to get the right type of assignments from gpu72, now the plague spreaded to other cores too, therefore we killed the proxy and hung its groins on the gate... for everybody to see that we killed it. We are going to finish the (wrong type of) assignments we got, but we will not request other through the Gpu72 proxy for a while, from those cores.

Just to let you know.

OTOH, we are doing now TF to 71 for all (30-33M) exponents we got for DC. We can do 53 of those things in the same time we do one DC, so if we find a factor, this DC is cleared (and deleted from our list) and we save/gain some time. We already found a factor for 31388729, (nice one, starts with 1234...., unreported yet, it will be reported automatically when its time come by the batch, in aout one hour). This credit also does not go to Gpu72 (i.e. will not be seen on those nice graphics, but it can be seen on Gimps), as we didn't get the TF assignments from there. Sorry. We can't do better for now, till the problem is solved.

[garo]

Originally Posted by LaurV

OTOH, we are doing now TF to 71 for all (30-33M) exponents we got for DC. We can do 53 of those things in the same time we do one DC, so if we find a factor, this DC is cleared (and deleted from our list) and we save/gain some time.

This does not make sense if all those 30-33M are already TFed to 70 bits. Assuming you are using the same hardware, by doing TF to 71 you clear 53/71 exponents in the same time you do a DC. So how exactly are you saving/gaining time?

[chalsall]

Originally Posted by garo

This does not make sense if all those 30-33M are already TFed to 70 bits.

Every DC candidate below 33M is already TFed to at least 70 "bits". Every available DC candidate in the 33M range is TFed to at least 71 bits.

[LaurV]

Originally Posted by garo

This does not make sense if all those 30-33M are already TFed to 70 bits. Assuming you are using the same hardware, by doing TF to 71 you clear 53/71 exponents in the same time you do a DC. So how exactly are you saving/gaining time?

Well, it depends on your luck :D. In fact your chances are lower, considering that some P-1 was also done on the range (like 53 in about 90, and not 71). But I am a lucky guy, see below.

You do not "clear" 53 exponents in this time. You try to clear them, that is why is called trial factoring. You may be lucky and find factors faster then it would take you to do a complete DC, therefore "clearing" those exponents for which you found factors. You also might be unlucky, and find no factor, in this case the DC LL tests still will have to run for all of them, so you wasted the time spent to do TF.

You may also stop immediately after you found a factor, therefore the rest of the time is "saved".

Now put into the equation the number of TF you can run in the same time you could do a DC (yes, on the same GPU hardware, don't count the CPU), and the probability to find a factor for exponents you got, and you will have a system which you can solve and see how much TF you need to do, to "gain". Factor your luck inside

Concrete: I found 2 factors (the one mentioned before, and 31420847 factored few minutes after) in 6 hours, going through the 27 exponents I had assigned last night (with a single gtx580, it takes 23 minutes and 19 seconds to do a 31M from 70 to 71). To do two full DC tests on the same card it would take 50 hours. So, I "gained/saved" 44 hours of work to clean those 2 exponents.

Of course, it is like trading Forex, you are not always so lucky, but probabilistically I am on plus side, that is why I am still doing it.

What was your question, again?

[garo]

Originally Posted by LaurV

Well, it depends on your luck :D. In fact your chances are lower, considering that some P-1 was also done on the range (like 53 in about 90, and not 71). But I am a lucky guy, see below.
<snip>

Of course, it is like trading Forex, you are not always so lucky, but probabilistically I am on plus side, that is why I am still doing it.

What was your question, again?

Right! Let us work against probability on a maths project. Nice one.

[c10ck3r]

Originally Posted by LaurV

You also might be unlucky, and find no factor, in this case the DC LL tests still will have to run for all of them, so you wasted the time spent to do TF.

Wasted? Not necessarily. You've just doubled the depth the exponents have been TF'd to, which (if there is ever a reason or desire to prove one or more factors for these exponents) gives a higher ground to search from. Plus, it allows the GPU to do what it is best at (TF) instead of work that is better fitted for CPUs.