Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
Message boards :
Generalized Fermat Prime Search :
The upcoming n=21 search
| Author |
Message |
|---|
|
|
|
I felt the need to post, because the N=19 double-check to b=110k is over halfway there, unless I've misread the statistics, or read the wrong statistics. So N=21 is next.
Michael G, in the other thread, wrote:
D) We will begin n=21
Our plans for testing n=21 are rather nebulous at this point beyond "n=21 is next". No decisions on deadlines or whether this will be a GPU only or GPU plus CPU project have yet been made.
Please let CPUs have a go at this. I've obviously not got half the idea that you have about how many people are running slow CPUs, like older Core2s, or AMD CPUs... but whenever I look at "wingmen" for my WUs, it seems like most are on an i7 or i5 of some sort. Hopefully with a "strict but fair" deadline, and a massive bold notice on the user preferences page, the number of people trying to run this and SoB on their Pentium 4 M can be reduced? I've been a bit hyperbolic there, I know, but currently there are still projects which slow computers (by cutting edge number-crunching standards) can run happily on PG.
I know it must be a difficult decision for PG admins - but I wanted (well, I still do want) the chance to run a GFN-WR task on a CPU. I do realise, however, how that could seriously hamper that particular project. But SoB is CPU-only - surely GFN n=21 WUs will "only" take a similar amount of time on a CPU?
I can imagine there will be many timeouts and people aborting after many days when they're shocked that it's taking so long, but as you all obviously know, GPU WUs regularly error out, throw up weird errors, and whatever else similar. With people running the CPU app, who at least half-know what they're doing, you could look at it as a bit of stability.
I know you all must have debated this internally, and maybe until you're sick in response to the mere mention of the subject, but I feel it's important to state the CPU's case here.
Just an idea: I was going to say that presumably a cpu app would be created at some point, anyway, for when the GPUs start to hit their b limit. Would it be efficient to initially have the cpu app crunch only tasks of, and above, the rough b limit of OpenCL*** GPUs? Then once that's done, you could have the CPUs join the GPUs from wherever they're up to. That way, you sort of have one decision removed for you, i.e. it's not possible for GPUs to do some tasks in the range, so the only option is CPU.
Anyway, I'm rambling now, cheers.
***Edit: I see CUDA can go slightly further than OCL with N=21. | |
|
 Michael Goetz Volunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14037
ID: 53948
Credit: 477,051,011
RAC: 285,770
                               
|
|
About 20% of the n=19s have been sent out; about 14% are done.
Nothing has been decided yet.
____________
My lucky number is 75898524288+1 | |
|
|
|
|
Ah, so I did mis-read the stats re: n=19, cheers.
But you know where I stand with regards to the n=21 search, anyway. Will they have a similar runtime to the current SoB WUs? I don't expect you to do any accurate calculations or anything else, just wondering if you had a rough idea. | |
|
HonzaVolunteer moderator
Volunteer tester
Project scientist
Send message
Joined: 15 Aug 05
Posts: 1963
ID: 352
Credit: 6,402,773,857
RAC: 2,541,172
                                   
|
Ah, so I did mis-read the stats re: n=19, cheers.
As JimB corrected - there are two b ranges in n=19:
18-110K
700K-895K
Note that second range is more than twice demanding since it has quite large b values.
Will they have a similar runtime to the current SoB WUs? I don't expect you to do any accurate calculations or anything else, just wondering if you had a rough idea.
You can extrapolate from difference between n=18 and n=19, then n=19 and n=20 and then n=20 and n=21 (use similar b value).
This would apply more to CPU; GPUs are doing better with higher n's.
Or you can use benchmark.
Or you can try to do a test yourself for a desired b value to see if you CPU would be up to the job.
____________
My stats | |
|
|
|
|
I was hoping for some help with understanding the difference in length, hence my post/question.
[I didn't and certainly do not expect anyone to do any tests, though].
This isn't meant to sound sarcastic - it's just a literal statement - I assumed one or more of the admids would have done some advance planning by" now".
But importantly, when I made that assumption, I was looking at only Michael's older post which said we only had to do one range, which we are currently more than halfway through. So I apologise for not doing more complete reading-up and posting a bit prematurely.
I would still like to see n=21 available for CPUs, though. Skylake has been discussed slightly on the Number Crunching forum but mostly before the full details were known - it isn't just the CPU's features which will matter, it's a lot of other things like the PCH and chipset(s), which I suppose should probably be discussed in the Number Crunching forum in the Skylake thread when people actually have it.
| |
|
Michael GoetzVolunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 14037
ID: 53948
Credit: 477,051,011
RAC: 285,770
|
|
Command line: genefer64 -q 500000^2097152+1
Priority change succeeded.
Testing 500000^2097152+1...
Using FMA3 transform
Estimated time remaining for 500000^2097152+1 is 149:52:52
Does that help? That's a Core i5-4670K at stock speeds, with only 1 core running.
You can start up a Genefer test on your own computer to see how fast it will run on your specific hardware. The time estimate is displayed shortly after startup.
I was hoping for some help with understanding the difference in length, hence my post/question.
Considering how easy it is to answer this question for yourself, I sort of expected if you wanted the answer to this question you'd go ahead and see how your hardware performs on these numbers. You don't have to run full length tests, after all. It only takes a moment to see how long a full test will take.
____________
My lucky number is 75898524288+1 | |
|
|
|
|
Cheers Michael - it does help. I'll reply slightly more fully later - nothing important though, just an explanation of why my initial reply might have seemed rude, but was not intended to be. Bit of a complicated one I suppose. | |
|
Message boards :
Generalized Fermat Prime Search :
The upcoming n=21 search |
