GPU to 72 status...

[flashjh]

Originally Posted by Dubslow

There are ~250 available in the yellow column at the moment. Someone dumped a load of work. (Probably Xyzzy.) Also note that TF to 72 throughput is more than 5 times the P-1 throughput, which while not a problem at lower levels where the P-1/Non ratio is much lower, at the higher expos that ratio is much much greater than 1/5.

If someone has a P-1 server or farm they want me to run, I'll take it With a resoanably fast system it is still taking like 12 hours per P-1, for me. I'm hoping to put another P-1 system together within the next month or so.

Any recommendations on hardware for a really good P-1 system?

[kladner]

Comment deleted by poster. I decided it wasn't worth taking up.

[flashjh]

Originally Posted by kladner

Comment deleted by poster. I decided it wasn't worth taking up.

I read your quote before you removed it... it was fine. (And quite funny).

Perhaps we all need to adjust a few TF workers to P-1 to balance the system? My trouble is P-1 takes so long without a lot of memory and my systems are basically maxed out.

[kladner]

Originally Posted by flashjh

I read your quote before you removed it... it was fine. (And quite funny/true).

OK. I saved it:

Originally Posted by Dubslow

I really wish I was kidding. Take a look at this. (And that's only the tip of the iceberg.)

I looked at your link, and pulled some assignments for P-1 in the 58M range. But I had to settle for assignments which had only been TF'd to 71.

From that, I would ask about the relative priorities of P-1 v TF. Granted, one runs on the CPU and the other on the GPU. But isn't there a chance that TF has at least as much need as P-1? At one point the GPU to 72 assignment page for P-1 stated that exponents factored to less than 72 might be reassigned for more TF.

Perhaps that standard has been modified, since it no longer appears on that assignment page. But still, P-1 work is being done. I can't contribute as much to that effort since the Opteron x2 board apparently died. (Still haven't gotten around to testing the alternate PSU.) But I question if my current allocation of resources needs changing. I have two mfaktc instances running TF and DC factoring--using two PhenomII 1090T cores, with three of the remaining running P-1, and one doing LL and/or DC.

Do you perhaps obsess too much about what isn't happening, at the expense of what is?

(It's hard to resist encouragement.)

[flashjh]

Originally Posted by kladner

OK. I saved it:

Do you perhaps obsess too much about what isn't happening, at the expense of what is?

(It's hard to resist encouragement.)

Now for my original reply... I think we all do that. Especially considering what's involved with G272... afterall manually getting the assignments and working with non-GUI isn't exactly simple.

In the end, the project is moving along as intended. I don't think it was ever expected that we could keep up with the P-1?? Also, hardware issues happen, as you well know. My main P-1 system was down twice in two weeks which put me ~100 P-1s behind. Oh well. We'll do what we can... I think that if exponents get released before P-1 that we 'may' get another chance at them, is that correct?

[kladner]

Originally Posted by flashjh

Now for my original reply... I think we all do that. Especially considering what's involved with G272... afterall manually getting the assignments and working with non-GUI isn't exactly simple.

In the end, the project is moving along as intended. I don't think it was ever expected that we could keep up with the P-1?? Also, hardware issues happen, as you well know. My main P-1 system was down twice in two weeks which put me ~100 P-1s behind. Oh well. We'll do what we can... I think that if exponents get released before P-1 that we 'may' get another chance at them, is that correct?

I am really pretty ignorant of such things, but my impression is that we are sort of bottom feeders (no derogatory implications intended.) We (Spidy, actually) scoop up neglected/discarded material and get it back into the food chain.

As for your original reply:
I guess I am also in a maxed out condition right now, too. The Opteron system did contribute, but fairly slowly. It might have been better off at DC rather than P-1. If I get it to revive, I might go that route and put the one PhenomII core back to P-1.

[Dubslow]

Originally Posted by kladner

I looked at your link, and pulled some assignments for P-1 in the 58M range. But I had to settle for assignments which had only been TF'd to 71.

Philosophical musings and jokes aside (I can't compete there) there were ~250 (and still around ~220) available at 52M TFd up to 72. Why would you go with higher expos and less TF?

Originally Posted by flashjh

In the end, the project is moving along as intended. I don't think it was ever expected that we could keep up with the P-1?? Also, hardware issues happen, as you well know. My main P-1 system was down twice in two weeks which put me ~100 P-1s behind. Oh well. We'll do what we can... I think that if exponents get released before P-1 that we 'may' get another chance at them, is that correct?

As far as obsessing over what isn't happening, at least we're not complaining about finding a new prime, or the fact that overall GIMPS work has been slowing down (citation to come). All I know is, in the current state of GPU72, there's significantly more TF than P-1 throughput (duh!). I'm only commenting so that everyone else is aware, and perhaps sway a few others to join the P-1 bandwagon. I myself am only a recent inductee. (As of the new year, when petrw1 announced his work type shift, I had decided to do one LL worker, one DC worker, and one P-1 worker. Before that, I had decided that I would keep their counts even on PrimeNet [which would mean 90% of my work to LL]. And now all three cores are P-1! The graph is pretty awesome ).

Originally Posted by kladner

From that, I would ask about the relative priorities of P-1 v TF. Granted, one runs on the CPU and the other on the GPU. But isn't there a chance that TF has at least as much need as P-1? At one point the GPU to 72 assignment page for P-1 stated that exponents factored to less than 72 might be reassigned for more TF.

Perhaps that standard has been modified, since it no longer appears on that assignment page.

I don't see why it would be. Expos are kept until yellow, and hopefully when they do hit yellow we can P-1 them. @flash: As I pointed out earlier, PrimeNet TF assignments are in the 56M range right now, and though I haven't tested its P-1 assignments, my feeling is that it's pretty similar, if the GPU72 numbers are anything to go by. As soon as we clear everything south of 55-56M, I think we need to seriously reconsider moving back to just PrimeNet, instead of reserving the expos as LLs, as was the original intent and current design of GPU72. At such time, it would be something that we work on only as needed, for those expos that have fallen behind the wave. (By 'as needed', I mean not full time GPU72.)

[nucleon]

Recommended P-1 system - SNB-E cpu with quad chan memory running prime95 v27. With only 2x cores @4.2GHz on p-1 I was doing 15GHzdays/day. The other 4 cores were running GPUs.

I'm a major fan of the SNB-E platform at the moment.

What may help with p-1 workload - p-1 on gpu. But it appears no one is taking that onboard.

The dark horse of course is the new GTX600 series. It may be fast enough on LL not to need CPU p-1 done. Breath isn't held....yet.

-- Craig

[flashjh]

Originally Posted by nucleon

Recommended P-1 system - SNB-E cpu with quad chan memory running prime95 v27. With only 2x cores @4.2GHz on p-1 I was doing 15GHzdays/day. The other 4 cores were running GPUs.

I'm a major fan of the SNB-E platform at the moment.

What may help with p-1 workload - p-1 on gpu. But it appears no one is taking that onboard.

The dark horse of course is the new GTX600 series. It may be fast enough on LL not to need CPU p-1 done. Breath isn't held....yet.

-- Craig

Can v27 do 64 bit yet? How did you manage memory if it can only use 4Gb?

I'm going to take a look at P-1 on GPU once I finish my masters in March (unless someone does it before then)

Jerry

[flashjh]

Originally Posted by NBtarheel_33

Heavens to Betsy. I promise, I really do have an advanced math degree. When looking at "2828", no doubt, I was thinking of the first few *digits* of e: 2.71828, but just forgot about the "71". Oh well, not so exciting then.

As for the pi milestone, I had a screen shot at 314,159, but then lost it when I accidentally refreshed the milestones page. So 314,158 had to be close enough for government work.

In other news, with GPUto72 mowing down the exponents, all of the milestone paces have been picking up as of late. Linear regression based on YTD 2012 data puts the M46 and M47 milestones back within 2017; they had drifted into mid-2018 based on last year's data.

Looks like it's working

[James Heinrich]

Originally Posted by flashjh

My trouble is P-1 takes so long without a lot of memory and my systems are basically maxed out.

More memory typically doesn't help that problem: adding more RAM lets stage2 run more efficiently, so Prime95 responds by increasing bounds (for higher factor probability) at the expense of an even-longer runtime.