Anyone using GPUs to do DC, LL or P-1 work?

[Xyzzy]

For P-1 testing we think even 512MiB per core is fine. (Correct us if we are wrong!)

http://www.mersenneforum.org/showpos...5&postcount=10

[garo]

Originally Posted by Xyzzy

For P-1 testing we think even 512MiB per core is fine. (Correct us if we are wrong!)

http://www.mersenneforum.org/showpos...5&postcount=10

Absolutely right. I run my instances with 2GB of RAM but that's only because I have that much to spare. I wouldn't think twice about running with with 512MB. It's totally fine.

[tigreroars]

Originally Posted by TheMawn

Have you got 8 GB total or 8 GB per core? If you have 8 GB total, you can leave one doing P-1 with as much memory as you feel comfortable with and put the rest to DC or LL. If it's 32 GB total, then I really don't know. You get diminishing returns as you commit more memory to P-1 so you could do a mix of LL/DC and P-1 as you desire.

Both work types are useful, though we do currently have a surplus of P-1...

I have 1 P-1 unit (at a time) going. 1 Primary and 1 Helper core on it with 8 gb dedicated for that thread.

I've been semi-following the threads and missed the release of the GPU P-1. However, based on what I understand from following the threads, the problem with the GPU P-1 programs is the lack of memory that you can devote to it. So, since I'm not using the computer, I decided to throw a chunk of memory at it and 8gb sounded like a nice round (power of 2?) number.

So, this may or may not be the appropriate thread for this, but I'm trying to figure out the best use of CPU power. LL/DC LL seems like it's CPU only. P-1 the advantage might be the sheer amount of memory I can throw at it. TF is right out. Can you guys advise?

Edit for Paragraphs.

[chalsall]

Originally Posted by tigreroars

So, this may or may not be the appropriate thread for this, but I'm trying to figure out the best use of CPU power. LL/DC LL seems like it's CPU only. P-1 the advantage might be the sheer amount of memory I can throw at it. TF is right out. Can you guys advise?

LL/DC can be done on a GPU as well, but right now GPUs should be devoted to TFing since that's were they are most efficient, and we're still not TFing as deeply as we should optimally be doing (we should really be going to 75 "bits" for most cards, but only have the fire-power to sustain 74).

We currently have an over-supply of CPU P-1ing fire-power, ironically largely because of Oliver, the creator of mfaktc (the GPU TFing program) who also has access to some serious CPU resources!

But, if you enjoy P-1'ing (and you certainly have the memory for it!), keep doing it. If you're not committed to that work-type, using your CPU for either LL'ing or DC'ing would be helpful.

Keep in mind an important tenet of GIMPS: people should do what they enjoy the most with their own time, hardware and money!

Hope that helps.

[Mini-Geek]

Originally Posted by chalsall

LL/DC can be done on a GPU as well, but right now GPUs should be devoted to TFing since that's were they are most efficient, and we're still not TFing as deeply as we should optimally be doing (we should really be going to 75 "bits" for most cards, but only have the fire-power to sustain 74).

Does that mean that if the GPU/CPU firepower levels changed so that GPUs could go to 76 before the CPUs LL them, we'd stop at 75 and start LLing on GPUs part-time? I think that sounds right, just trying to be sure I (and anyone else reading this) am clear on that.

[chalsall]

Originally Posted by Mini-Geek

Does that mean that if the GPU/CPU firepower levels changed so that GPUs could go to 76 before the CPUs LL them, we'd stop at 75 and start LLing on GPUs part-time?

No. The "transition point" is based on James' in-depth analysis of where a GPU can LL (or DC) more candidates than can be eliminated by TF'ing.

I can't find the link at the moment, but basically we should be optimally TF'ing to 75 for most of the LL candidates, but right now we are only just staying ahead of the LL and DC (mostly CPU) workers.

[TheMawn]

Originally Posted by Mini-Geek

Does that mean that if the GPU/CPU firepower levels changed so that GPUs could go to 76 before the CPUs LL them, we'd stop at 75 and start LLing on GPUs part-time?

Originally Posted by chalsall

No. The "transition point" is based on James' in-depth analysis of where a GPU can LL (or DC) more candidates than can be eliminated by TF'ing.

I can't find the link at the moment, but basically we should be optimally TF'ing to 75 for most of the LL candidates, but right now we are only just staying ahead of the LL and DC (mostly CPU) workers.

I think Chris misread you. I believe you are correct. If the GPU TF firepower became so great that we COULD TF to 76, we would probably just build up our buffer for a while and then suggest some people migrate to GPU LL.

[chalsall]

Originally Posted by TheMawn

I think Chris misread you. I believe you are correct. If the GPU TF firepower became so great that we COULD TF to 76, we would probably just build up our buffer for a while and then suggest some people migrate to GPU LL.

You are correct; I misread.

Yes, if we had the fire-power / buffer to TF to 76, then we would only go to 75, and some GPUs cycles would optimally be devoted to LL'ing / DC'ing.

We're a long way from being there at the moment, though....

Edit: Oh, and James' analysis is at http://www.mersenne.ca/cudalucas.php?model=12 The graph isn't working for me at the moment, but it might be my browser.

[garo]

Originally Posted by chalsall

Edit: Oh, and James' analysis is at http://www.mersenne.ca/cudalucas.php?model=12 The graph isn't working for me at the moment, but it might be my browser.

The graph hasn't worked for me for at least a couple of months.

[LaurV]

The graph does not work since few months. I had a long PM discussion with James (which is momentarily in holiday, see one of his recent posts) about those graphs, he uses Google Charts, and the final conclusion was that Google Charts are broken (see the GC forums), and he is waiting for Google to fix them, as they broke their own code.

What I later found (and most probably James doesn't know yet, due to his holiday) is the fact that the functions he used are not broken, but obsolete. They had a "grace period" last summer, which expired in November or so.

Most probably, James will fix them after he comes back.

[dragonbud20]

I am new here and I don't really know much about GPU work for GIMPS but if you wanna tell me how to do what your trying to do I am more than happy to contribute the power of my GTX 590