GPU to 72 status...

[petrw1]

Originally Posted by chalsall

The latter.

Is it just me that finds it curious that short of begging for last 2 years George has been unable to get enough P-1'ers to keep ahead of the the LL wavefront; but in this sub-project the P-1'ers are so plentiful they are running out of assignments???

[kladner]

Originally Posted by Dubslow

Edit: Now that I think about it, I'm getting 95% load without an OC, and with OC, the load only decreases. So I think that's where your extra juice is coming from.

The 95% refers to GPU usage? I get that kind of level from two instances. With one it's only around 70%. You're running one instance, right? But you get to put another i5 core to other uses. I fully expect i5 cores to outperform Phenom II cores, so your overall throughput for one machine might be similar. But that brings up the fact that I can give two cores to mfaktc and still have four for other stuff: P-1 at the moment. That, and I'm running a second machine with two Opteron cores doing P-1.

It would be interesting to stack up the output from i5 + GTX460 vs 1090T + GTX460.

[petrw1]

Originally Posted by chalsall

Thanks for all your work petrw1, and for being the first big P-1 contributor!!!

Looks like monst, delta_t et al are going to be able to help fill the void you leave behind.

You're welcome; it was fun. I've never been on top of any report before on GIMPS or any of its related subprojects.

I was the first to 100, 200, 300, 400, 500 and 600 but now there are a dozen over 100.
And it won't be long before monst, delta_t and other surpass my 600.

P.S. Less than 3 days from my first DCs.

[Dubslow]

Originally Posted by kladner

The 95% refers to GPU usage? I get that kind of level from two instances. With one it's only around 70%. You're running one instance, right? But you get to put another i5 core to other uses. I fully expect i5 cores to outperform Phenom II cores, so your overall throughput for one machine might be similar. But that brings up the fact that I can give two cores to mfaktc and still have four for other stuff: P-1 at the moment. That, and I'm running a second machine with two Opteron cores doing P-1.

It would be interesting to stack up the output from i5 + GTX460 vs 1090T + GTX460.

I have an i7 slightly OC'ed to 3.9 GHz (when Turbo Boost feels like it). I had a 1055T before I got this, and I'm quite sure that even at stock speeds, my four cores got more throughput than the 1055T. OTOH, you're higher than 2.8 GHz, but OTOOH, I am OCd to 3.9 now. With the 460 at 751 MHz core clock, pinned to one core, I get 95-99% load. I've determined in the last day or two that OCing the 460 to 850 MHz doesn't do anything except decrease the (relative) load. (Avg. rate and time per class do not improve.)

[kladner]

Well, single 1090T core at 3.5 GHz can't come close to saturating the 460. I just stopped one instance and got this:

58M exponent, 70-71 factoring
60% utilization
time -dropped from ~4.8-4.9s to 4.25-4.35s
SievePrimes eventually dropped from 12,000 to 5000
Wait went from ~3.8 to 2.5

I misremembered the utilization as being around 80% on a solo run.

It does argue that the i5 does a much better job of feeding mfaktc.

[garo]

Originally Posted by petrw1

Is it just me that finds it curious that short of begging for last 2 years George has been unable to get enough P-1'ers to keep ahead of the the LL wavefront; but in this sub-project the P-1'ers are so plentiful they are running out of assignments???

Indeed. Could I urge chalsall to reserve small P-1 candidates a bit more aggressively? Having said that, I managed to reserve a cluth under 50M today so maybe it was just luck!

[davieddy]

I'm not disagreeing with anyone, but I do think that if any strategy is to be advocated for prime searching (and Jeez have I come in for some stick for that!) we should recognize that a CPU can do LL tests and preceding P-1 (if necessary), but it can't sensibly "GPU to 72".
This blindingly simple observation consigns the "P-1 before the last bit(s) of TF" idea to oblivion.

So help me Mod.

David

[bcp19]

Originally Posted by garo

Indeed. Could I urge chalsall to reserve small P-1 candidates a bit more aggressively? Having said that, I managed to reserve a cluth under 50M today so maybe it was just luck!

Um, what exactly is a cluth?

[chalsall]

Originally Posted by garo

Indeed. Could I urge chalsall to reserve small P-1 candidates a bit more aggressively? Having said that, I managed to reserve a cluth under 50M today so maybe it was just luck!

"Spidy" can only do what Spidy is doing...

The availability of P-1 work is a function of how fast the GPUers are TFing (plus those candidates which expire which have not had P-1 work done already).

But... I have changed the P-1 assignment strategy slightly such that if there are no candidates TFed to 72 which need P-1 work done, those TFed to 71 will be assigned.

[Dubslow]

Originally Posted by davieddy

I'm not disagreeing with anyone, but I do think that if any strategy is to be advocated for prime searching (and Jeez have I come in for some stick for that!) we should recognize that a CPU can do LL tests and preceding P-1 (if necessary), but it can't sensibly "GPU to 72".
This blindingly simple observation consigns the "P-1 before the last bit(s) of TF" idea to oblivion.

So help me Mod.

David

Since when is any of this new? chalsall currently assigns P-1 after 72 bits is complete, as he just said above me.

[davieddy]

Originally Posted by bcp19

Um, what exactly is a cluth?

I was going to ask the same question, but all my suggestions were unprintable.