[QUOTE=Prime95;273903]Feel free to use the manual assignments page. One of two things will happen:
1) The result is recorded, you get CPU credit, it doesn't count as a double-check even if it matches.
2) The server rejects the result as a duplicate.

Let us know what happens[/QUOTE]

I think that (2) should be the right and logical to happen. Happening of (1) would allow easy cheating, by sending the same result many times. That is, I can get a LL residue from somewhere (done with P95, CudaLucas, or taken from known results of other guy, etc) and send it from all the computers I have access, as "CudaLucas result". It won't be right to get credit for all of them.

Problem unsolved
 

[QUOTE=Brain;273907]What a pity for the second CUDALucas user on submit...[/QUOTE]
The basic question is still unanswered: How can I avoid DCing an exponent with CUDALucas which had its first time test also via CUDALucas?

The only way I know is as follows:
1. Goto exponent status page
2. Goto exponent's history entries
3. Check that residue contains any SMALL [a-f]
4.a) If 3. evaluate to yes, skip exponent as it was done via CUDALucas.
4.b) If 3. evaluate to no, check that it does contain capital letters and if yes, manual assign it for DC.

Problem is: There's no guarantee that CUDALucas residue contain small letters.

Suggestion: Please print the full submit message as this should contain the submitting client. This is the existing style for history mfaktc results.

One problem stays: Users who don't check the exponent's status before getting DC assignments.

Why all this? Users who submit DCs won't be credited and disappointed. I cannot test this last statement on my own. Correct me if I'm wrong.

The same problem applies to any third party client. PrimeNet will not credit more than one third party result per exponent, and third party clients do not necessarily use lower case residues.

IMHO, the only feasible option is to restrict yourself to NO-LL and LLERR-only results, which you will have to scavenge for manually.

[QUOTE=Ethan (EO);273828]If Serge still has his interim checkpoint files ([URL]http://www.mersenneforum.org/showpost.php?p=187713&postcount=92[/URL]) and there is a way to utilize them, this triple-and-a-half check could happen relatively quickly.[/QUOTE]
What was the FFT size used for these two GPU-DCs? 8M, 6M?

I remember that in 2009, this number was run with a bland 6M FFT (there were no other sizes around), but when I now started a small proof-of-concept re-run - from 99M to the end, the FFT size was picked to be 5376K by v26.6. Fancy! (it has both radix-3 in one stage and radix-7 in the other!) ...it reran thus far and it matched.

I was too lazy to manually rotate the binary contents of the savefiles, but this is what I always meant to have done - see the 2009 discussion. However, different FFT size is also sort of a diversion from the old routine; the machine is also different, the old one is pushing up the daisies these days.

[QUOTE=Batalov;274156]What was the FFT size used for these two GPU-DCs? 8M, 6M?[/QUOTE]

CUDALucas is currently power-of-2 FFT sizes; the GPU-DCs were run at 8M FFT.

Ethan

How much memory needs for 100M exponents?

For example:
70000141 - is about 248 Mb
290000033 - is about 728 Mb

I have GTX460 768 Mb and i want to buy new card with more memory ...

[QUOTE=Lorenzo;275293]How much memory needs for 100M exponents?

For example:
70000141 - is about 248 Mb
290000033 - is about 728 Mb

I have GTX460 768 Mb and i want to buy new card with more memory ...[/QUOTE]
Ehm, I read that msft wrote that you cannot do 100M expo as max supported FFT size is 8M. I read that max expo is around 159,200,00. Is this still correct? Never tried it. I prefer DCs.

[QUOTE=Brain;275327]Ehm, I read that msft wrote that you cannot do 100M expo as max supported FFT size is 8M. I read that max expo is around 159,200,00. Is this still correct? Never tried it. I prefer DCs.[/QUOTE]

All done works!

[CODE]Iteration 10000 M( 290000033 )C, 0x9e2181b8405a2511, n = 16777216, CUDALucas v1.2[/CODE]

FFT size 16M!

But unfortunatly a 332M Exponent processed with a 1024 MB GTX560Ti causes:

CUDALucas.cu(527) : cudaSafeCall() Runtime API error : out of memory.

According to [URL="http://www.wolframalpha.com/input/?i=332+million+bits+to+MB"]Wolfram Alpha[/URL], that size exponent should only require around 50 MB. Current wavefront exponents are 6 MB.

You're measuring the size of the memory required to write down the candidate number N; what's relevant for CUDALucas is the amount of memory required to perform an FFT-based multiply of numbers the size of N, which is a good deal bigger (firstly because you can only fit about eighteen bits of N into each 64-bit floating-point number, and secondly because the CUDA FFT uses a lot of scratch space to allow it to use hundreds of threads).