GPU to 72 status... - Page 207

chalsall · 2013-05-15, 15:28

Quote:

Originally Posted by petrw1

Granted well under a percent different but it might be something you choose to look into when you are bored.

I'm simply using PHP code provided by James (derived from C code provided by George) converted to Perl for the credits and estimates. It is possible there's a slight discrepancy between this and what PrimeNet uses.

I would say that a direct comparison between PrimeNet and GPU72 should not be expected to be exact (although it should be close); PrimeNet is canonical.

Aramis Wyler · 2013-05-15, 17:17

primenet only gives partial credit if you submit a factor, but gpu72 will give you the given number of GHzDs credit whether you ran the whole 3 hour job w/o a factor or found a factor inside 5 minutes. Also sometimes I think primenet misinterprets my TF factors as some other type of work.

So there are some differences, though they are more minor with LL work than factoring or p-1. Depends on how many bits of factoring/p-1 you get with your LLs, maybe.

kracker · 2013-05-15, 17:24

Quote:

Originally Posted by Aramis Wyler

primenet only gives partial credit if you submit a factor, but gpu72 will give you the given number of GHzDs credit whether you ran the whole 3 hour job w/o a factor or found a factor inside 5 minutes. Also sometimes I think primenet misinterprets my TF factors as some other type of work.

So there are some differences, though they are more minor with LL work than factoring or p-1. Depends on how many bits of factoring/p-1 you get with your LLs, maybe.

Credit is totally same no matter how much P-1 or TF you do on it.

EDIT: On LL tests, of course.

chalsall · 2013-05-15, 17:44

Quote:

Originally Posted by Aramis Wyler

primenet only gives partial credit if you submit a factor, but gpu72 will give you the given number of GHzDs credit whether you ran the whole 3 hour job w/o a factor or found a factor inside 5 minutes.

Yes. I had asked James how to deal with this, and he explained (IIRC) that Prime95/mprime searches "bottom up", and mfakt* searches "top down".

I decided that since finding factors was our goal, there wouldn't be great harm in giving the Worker credit for the full "bit" range in such cases (to be clear, if someone pledged to TF to 74, but they found a factor at 71.x, they get credit for TFing to 72). Plus, it made my job easier!

Quote:

Originally Posted by Aramis Wyler

Also sometimes I think primenet misinterprets my TF factors as some other type of work.

A known bug, which has a known workaround -- submit a no-factor found result first (even if the result has already been submitted). Also, I believe James is working on this on PrimeNet.

Quote:

Originally Posted by Aramis Wyler

So there are some differences, though they are more minor with LL work than factoring or p-1. Depends on how many bits of factoring/p-1 you get with your LLs, maybe.

Estimates and Credits for P-1 work is going to be where people see the most discrepancy. When GPU72 gives "Estimated GHz Days" on the assignment form, it can't know what B1 or B2 values are going to be used. Once it observes this work completed, it can't (without extra work, and only in the case of a "no factor found" result) determine what B1 or B2 values were used.

IMO, at the end of the day, this doesn't really matter. We're here to do work. And the rankings between individuals (when viewed within the context of GPU72 or PrimeNet) are using the same metrics.

James Heinrich · 2013-05-15, 18:17

Quote:

Originally Posted by chalsall

Yes. I had asked James how to deal with this, and he explained (IIRC) that Prime95/mprime searches "bottom up", and mfakt* searches "top down".

On www.mersenne.ca I've now implemented a tweaked credit calculation that takes into account what class mfakt* would find the factor in, which adds a third possibility into the mix.

Prime95 · 2013-05-15, 18:26

Also, James estimates LL CPU credit based on the FFT size it expects prime95 to use. Prime95 may actually choose a different FFT size. When prime95 reports an LL result it also reports the FFT size used which Primenet uses to calculate CPU credit. There is no way for GPU72 to get this FFT size info to match Primenet's credit calculations.

lycorn · 2013-05-15, 22:06

Quote:

Originally Posted by Prime95

Depending on your budget, wait for a Haswell CPU due in three weeks.

Shall we take it as a hint that you´re planning some new improvements for Prime95 on this architecture? Due to AVX2, or FMA3, or...?

Prime95 · 2013-05-15, 22:16

Quote:

Originally Posted by lycorn

Shall we take it as a hint that you´re planning some new improvements for Prime95 on this architecture? Due to AVX2, or FMA3, or...?

FMA3 should provide a performance boost. I can't quantify it yet. It could be a lot as Haswell doubles the theoretical peak FLOPs rate --- or it might be small because we are still memory bandwidth limited. This has been discussed somewhat in the Haswell thread.

chalsall · 2013-05-16, 16:24

Quote:

Originally Posted by davieddy

I've got a really crackpot idea:
Take expos (>62M and not yet assigned for LL) in ascending order, and TF them to 74.

As a demonstration that I don't dogmatically reject your suggestions, I had a deep think about this, ran some numbers, and concluded that we can and should do this. (Although, I do find it a bit ironic that this suggestion came from you, after the all the "debating" about how we couldn't sustain taking 63M and above to 74 bits...)

So everyone knows, "What Makes Sense" is now "Lowest Exponent" to 74, starting from 62M. The LL P-1 form (and proxy) will now only assign work which is already TFed to at least 74 bits.

To be fair to P-1 workers who get their assignments directly from Primenet, most of the ~6,000 candidates in the 62M range TFed to "only" 73 bits will remain with Primenet until we've returned enough TFed to 74 to satisfy their requests for work (should be about a week).

In order to not starve Primenet's LL workers, those candidates (~3,500) with a P-1 run completed will remain with Primenet for the time being. We can decide if we want to bring those in for processing once enough candidates that have been TFed to 74 and P-1'ed are available to satisfy the request load.

As always, comments welcome.

James Heinrich · 2013-05-16, 16:38

Quote:

Originally Posted by chalsall

... most of the ~6,000 candidates in the 62M range TFed to "only" 63 bits will remain with Primenet until we've returned enough TFed to 64 to satisfy their requests for work (should be about a week).

Presumably that was supposed to be "73 bits" and "74" respectively?

chalsall · 2013-05-16, 16:41

Quote:

Originally Posted by James Heinrich

Presumably that was supposed to be "73 bits" and "74" respectively?

Yeah....

("Dyslexics of the world! Untie!")

Corrected. Thanks for pointing that out.

2013-05-15, 17:17	#2268
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 2³×53 Posts	primenet only gives partial credit if you submit a factor, but gpu72 will give you the given number of GHzDs credit whether you ran the whole 3 hour job w/o a factor or found a factor inside 5 minutes. Also sometimes I think primenet misinterprets my TF factors as some other type of work. So there are some differences, though they are more minor with LL work than factoring or p-1. Depends on how many bits of factoring/p-1 you get with your LLs, maybe. Last fiddled with by Aramis Wyler on 2013-05-15 at 17:18

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Status	Primeinator	Operation Billion Digits	5	2011-12-06 02:35
62 bit status	1997rj7	Lone Mersenne Hunters	27	2008-09-29 13:52
OBD Status	Uncwilly	Operation Billion Digits	22	2005-10-25 14:05
1-2M LLR status	paulunderwood	3*2^n-1 Search	2	2005-03-13 17:03
Status of 26.0M - 26.5M	1997rj7	Lone Mersenne Hunters	25	2004-06-18 16:46

2013-05-15, 18:26	#2272
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 2×3²×419 Posts	Also, James estimates LL CPU credit based on the FFT size it expects prime95 to use. Prime95 may actually choose a different FFT size. When prime95 reports an LL result it also reports the FFT size used which Primenet uses to calculate CPU credit. There is no way for GPU72 to get this FFT size info to match Primenet's credit calculations.