GPU to 72 status...

[davieddy]

Originally Posted by chalsall

We're "riding the wave"....

... and heading for a wipeout.

[TheMawn]

Originally Posted by chalsall

I am happy to say that the "deep" P-1'ing is very slightly more than keeping up with the LLing.

We're "riding the wave"....

Ah! Glad to hear it. I know it's kind of tempting to stick to the more glamourous first time LL tests since that's where all the money and glory is. Hopefully my old first gen i3 is doing my share of the P-1.

On a different topic, is there any way to see how much trial factoring is being done through a GPU versus through a CPU, now that GPU's are frankly embarassing CPU's?

[chalsall]

Originally Posted by TheMawn

On a different topic, is there any way to see how much trial factoring is being done through a GPU versus through a CPU, now that GPU's are frankly embarassing CPU's?

I don't have access to enough data from Primenet to be able to answer for only GPU vs. CPU. But it is fairly safe to say that this report represents mostly GPU efforts below 67M.

[petrw1]

Originally Posted by TheMawn

Ah! Glad to hear it. I know it's kind of tempting to stick to the more glamourous first time LL tests since that's where all the money and glory is. Hopefully my old first gen i3 is doing my share of the P-1.

On a different topic, is there any way to see how much trial factoring is being done through a GPU versus through a CPU, now that GPU's are frankly embarassing CPU's?

One sign to of this is the (lack of) progress of TF-LMH (Doing TF to 66 from 100M to 999M). I have been casually observing it for a few years. The rate of progress from 65 to 66 is MUCH MUCH less than half the expected rate of progress last round from 64 to 65. I think it is a combination of many of the older PC that were doing this have simply died ... but also many people are getting GPUs and can now do TF in the needed ranges instead of 100M+.

[kladner]

Originally Posted by petrw1

One sign to of this is the (lack of) progress of TF-LMH (Doing TF to 66 from 100M to 999M). I have been casually observing it for a few years. The rate of progress from 65 to 66 is MUCH MUCH less than half the expected rate of progress last round from 64 to 65. I think it is a combination of many of the older PC that were doing this have simply died ... but also many people are getting GPUs and can now do TF in the needed ranges instead of 100M+.

Ah! That reminds me that now I have my faster GPU back, I should throw in an occasional 332M TF job.

[TheMawn]

Originally Posted by petrw1

One sign to of this is the (lack of) progress of TF-LMH (Doing TF to 66 from 100M to 999M). I have been casually observing it for a few years. The rate of progress from 65 to 66 is MUCH MUCH less than half the expected rate of progress last round from 64 to 65. I think it is a combination of many of the older PC that were doing this have simply died ... but also many people are getting GPUs and can now do TF in the needed ranges instead of 100M+.

I suppose if trial factoring ever gets really, really far ahead of the wave, I may send my GPU to play with the big exponents for a while. As a test, I did 65 to 71 for something around 400 million and I got a real kick out of how fast 65 to 66 went.

How does a person go about working on getting those exponents from 65 to 66? I don't see how it can be done through GPU72 because it will only let you factor to a minimum of 71.

[kladner]

It doesn't really matter in GPU72. The lowest factoring level available in that range is 72. Things may be different by other means. Uncwilly is the most likely source of information about such things.

[chalsall]

Originally Posted by TheMawn

How does a person go about working on getting those exponents from 65 to 66? I don't see how it can be done through GPU72 because it will only let you factor to a minimum of 71.

One doesn't. At least currently.

GPU72's agenda is to help with the LL and DC (and now the "100M digits") wave-fronts. Thus, we carefully balance the available firepower with the available candidates, taking into consideration James' analysis as to how deep TFing makes sense.

Currently we're taking everything available in the 60M and above range to 74 "bits". I'm hoping we can start going to 75 bits at 64M or 65M.

[lycorn]

Originally Posted by TheMawn

How does a person go about working on getting those exponents from 65 to 66? I don't see how it can be done through GPU72 because it will only let you factor to a minimum of 71.

That has to be done manually.
Get the exponents from the server pages (you may get there via GPUto72 "Trial Factored Depth" reports) in text form, paste them in an empty worktodo.txt file, and edit it.
Report the findings through the server manual pages.

[TheMawn]

Originally Posted by chalsall

Thus, we carefully balance the available firepower with the available candidates, taking into consideration James' analysis as to how deep TFing makes sense.

Currently we're taking everything available in the 60M and above range to 74 "bits". I'm hoping we can start going to 75 bits at 64M or 65M.

As I recall, that analysis is dependent on a GHz-Days-ish measure of how much work it is to trial factor vs LL-test. The idea is the odds of finding a factor between 2⁷⁰ and 2⁷¹ is roughly 1 in 70 (most of my work has been in the low 70s and, lo and behold, about 1 in 70 turn up a factor) so it makes sense to spend about one seventieth of the time it takes to run an LL test, on trial factoring.

On the other hand, GPUs are ridiculously fast at trial factoring (my GTX 670 is equivalent to 285GHz). If that speed can't be harnessed for LL-tests (I recall reading somewhere that CUDALucas is faster than Prime95, but definitely not THAT much faster) then it might make sense to go deeper with the TF than the analysis suggests. It would make sense on a time basis (factoring an exponent up to 74 would take me about 5 hours, but doing the LL-test would take me nearly 500) but not so much on a GHz-Days basis.

[kracker]

Originally Posted by TheMawn

On the other hand, GPUs are ridiculously fast at trial factoring (my GTX 670 is equivalent to 285GHz). If that speed can't be harnessed for LL-tests (I recall reading somewhere that CUDALucas is faster than Prime95, but definitely not THAT much faster) then it might make sense to go deeper with the TF than the analysis suggests. It would make sense on a time basis (factoring an exponent up to 74 would take me about 5 hours, but doing the LL-test would take me nearly 500) but not so much on a GHz-Days basis.

We don't know of the performance of LL tests on AMD cards yet, but anyways looking on the CuLu benchmarks I don't know why I(only I probably) would run a LL test on a gpu, considering the speed is not a *huge* lot faster compared to a similar level cpu, and I could run TF on my gpu...