GPU to 66?
 

Hello fellow GIMPers!
I've been looking at the GPU272 group's posts and progress, and started to question the idea. I (think I) understand the desire to push the leading and DC waves to higher bit levels in order to potentially reduce the amount of LL needed. I'm sure the effort is worth it to the LLers and DCers, but is it the best use of the GPUs?
So, I would like to tenatively propose a sister project, GPU to 66, with the intention of pushing exponents between 80,000,000 and 1,000,000,000 to 66 bits, leaving the first 80M to the GPU272 crew. This will consist of taking 2,095,083 exponents from 64 to 65 bits, and 20,415,407 exponents from 65-66 bits, minus any no longer needed due to factors in 64-65.
Anybody with me?
Johannes Schuck

I do not have a GPU to take my comments for what they are worth.

1. The regular TF-LMH will complete the remaining 64-65 bit within a month. And then they will start again at 100,000,000 and go to 66 bits.

2. This will be about a 2-year assignment.....a long time; sort of but not really because it will still be LONG before the LL wave-front will ever get there.

3. I seem to recall reading elsewhere that GPUs preform best at higher bit levels.

mfakt* design means they're more efficient at longer runs, which in general means lower exponents. But higher exponents also need more TF, so they could be long runs too.

As for your idea, the GPU's are pretty much solely responsible for all TF work between 0 and ~60M, and right now we're not caught up. I think this will happen in the next 3-6 months, i.e. within then I think we'll be ahead of the LL wavefront, at which point your idea is worth more consideration IMO. Also go take a look at the LMH forum; there are some people who use GPU's for TF-LMH, although higher than 100M. There is even a special mfaktc version that is more efficient for the short runtimes.

I think it makes more sense to do a depth-first search (i.e. TF as deep as you can for the LL candidates that will be handed out in the next 0-6 month, while largely ignoring the rest), especially given the relatively new, and therefore likely volatile, nature of the GPU/CPU balance and speeds.

In 5 years our GTX5900Ti Extreme "Limited Edition" liquid-nitrogen-cooled GPUs will trial factor all of those in a matter of a few days.

:soapbox:

[QUOTE=Mini-Geek;284007]I think it makes more sense to do a depth-first search (i.e. TF as deep as you can for the LL candidates that will be handed out in the next 0-6 month, while largely ignoring the rest), especially given the relatively new, and therefore likely volatile, nature of the GPU/CPU balance and speeds.[/QUOTE]

Have I told you lately that I love you?
:smile:

David
x

[QUOTE=Mini-Geek;284007]I think it makes more sense to do a depth-first search (i.e. TF as deep as you can for the LL candidates that will be handed out in the next 0-6 month, while largely ignoring the rest), especially given the relatively new, and therefore likely volatile, nature of the GPU/CPU balance and speeds.[/QUOTE]

Isn't that what George started a few months ago when he added 2 or 3 more bit levels in the 50M range?

Or are you proposing more bits yet?

[QUOTE=petrw1;284060]Isn't that what George started a few months ago when he added 2 or 3 more bit levels in the 50M range?

Or are you proposing more bits yet?[/QUOTE]
Yep.
Unfortunately the ~26,000 primenet TF assignments have ground to a halt at 71 bits. The optimal bit level is still governed by the GPU firepower available ATM, but the times they are a changin'.

David

@OP
Bad idea. If you factor all these exponents to higher levels, slower CPUs will not be able to contribute meaningfully to GIMPS. It will be a long long time before even 100M exponents are handed out for LL tests so we need to concentrate on the immediate job first. Make sure every exponent under 60M due to be handed out for LL is factored to 71/72/73 bits.

Sounds like Garo half agrees with me (albeit loathe to admit it).
The brutal fact is that CPUs have been rendered redundant for [B]any[/B] TF.
If the fast guys would let up a bit on DCs (do the man-sized LL work instead), slower CPUs can do them. If the DC wavefront were smaller (say 1/3) than the LL wavefront, the chance of finding a prime per GHzDay would become comparable to that of first time LLs.

David

What do you mean by 'smaller'? I personally put one core to P-1, one to LL, one to DC, and one to mfaktc. That way by numbers I get the most P-1, second most DC, and least LL, while maintaining the same 'GHz-Days' throughput. Unless the DC throughput rate matches the LL throughput rate, I will continue to keep at least one core on DC.

Also, most of us agree in general with garo, and by proxy with you. You just keep spouting off anyways. (We're giving it all she's got, captain!)