[QUOTE=LaurV;291213]I finished 4 DCs.
All mismatches. Two of them were done on tesla boards (26166389 and 26176597). My money are on a software bug. Returning to v1.3 and mfaktc.[/QUOTE]

What version did you use?

[QUOTE=flashjh;291196]I'll have a good list of M26026433 residues from start to finish.[/QUOTE]
I also just started 26026433 using a Linux build of version 1.58 on a card that I trust. ETA is 31 hours:
Iteration 10000 M( 26026433 )C, 0x7b8dde83fd216693, n = 1572864, CUDALucas v1.58 (0:43 real, 4.2871 ms/iter, ETA 30:58:26)

[QUOTE=flashjh;291215]What version did you use?[/QUOTE]
1.58, sorry for not mentioning it, I was in hurry to leave for job (such short posts are not my style :P).

Tonight when I reach home I will dld and check 1.61 for another 4 expos (eta about 24 hours for 26M size, that would be about 33 hours from now, here is 10:15AM). I will keep you informed.

Honestly I don't like too much this activity; if I don't report the results then my time is wasted, and if I report them then my "reliability" as a gimps participant decreases as soon as they would be found being bad results. Some would say I intentionally report bad results to rise my credits. But I DID those tests, and the only thing keeping me doing them is the [U][B]hope[/B][/U] that we will have a [B]very reliable CudaLucas[/B] program in the future, which would be eventually faster (or at least same fast) for a good GPU as P95 is on a good CPU. [edit: and maybe, which will implement the shifting too, to make multiple-CL-checks acceptable by primenet, in that case]

Right now I am still... well, not bitchy, but a bit skeptic. Last version of CudaLucas, assuming that everything would work perfectly, could spit out a DC per day, but newest P95 with AVX running on i7-2600k at 4GHz spits out 4 DCs every 3 days (with about 10-12ms per iteration, in 26M range exponents, this is achievable even with lower-numbered, or lower-clocked, CPUs), that means the [B]CPU is 33% faster[/B], at half of the price and a fraction of the consumed power. [edit: I pay here about 10k baht for a CPU in this range, and i5-2500 or some AMD's could even be cheaper, but I have to pay 16-18k for a gtx580 and over 50k for a tesla, so the "double" is just an approximation; OTOH, a gtx560 top OC - the one at 950Mhz - costs about 6k5 to 7k baht, but its CL output is not so "tumultuous"].

So, thus I am deeply thankful to msft, brain and jerry for their efforts, however my enthusiasm is a bit down compared with 4-6 months ago (see my posts that time), and I start to move slowly to the other side of the barricade, into the crowd of the people arguing that GPU's should be only used for TF for now on (last year I was strongly arguing against such a "stupid" remark!)

But to finish in a positive note: I like the output format of the newer versions, with real time and ETA. Please do not change it back as happened in the past :P At least this is an improvement... All developers should use the last source, and if one of you makes changes or improvement, better tell to the others too... [edit: I said this not as irony or sarcasm, but I am sincerely afraid that one of you will find a bug in the last versions and will rewrite everything from scratch starting from an older version, like the good-and-stable v1.2b, or 1.4x, and all the nice cosmetic changes done to the last versions would be lost :P]

[QUOTE=LaurV;291229]1.58, sorry for not mentioning it, I was in hurry to leave for job (such short posts are not my style :P).

Tonight when I reach home I will dld and check 1.61 for another 4 expos (eta about 24 hours for 26M size, that would be about 33 hours from now, here is 10:15AM). I will keep you informed.

Honestly I don't like too much this activity; if I don't report the results then my time is wasted, and if I report them then my "reliability" as a gimps participant decreases as soon as they would be found being bad results. Some would say I intentionally report bad results to rise my credits. But I DID those tests, and the only thing keeping me doing them is the hope that we will have a very reliable CudaLucas program in the future, which would be faster (or at least same fast) for a good GPU as P95 is on a good CPU. [edit: and maybe, which will implement the shifting too, to make multiple-CL-checks acceptable by primenet, in that case]

Right now I am still... well, not bitchy, but a bit skeptic. Last version of CudaLucas, assuming that everything would work perfectly, could spit out a DC per day, but the i7-2600k running at 4GHz spits out 4 DCs every 3 days (with about 10-12ms per iteration, in 26M range exponents, this is achievable even with lower-numbered CPUs), that would be 33% faster, at half of the price and a fraction of the consumed power. [edit: I pay here about 10k baht for a CPU in this range, but I have to pay 16-18k for a gtx580 and over 50k for a tesla, so the "double" is just an approximation; OTOH, a gtx560 top OC - the one at 950Mhz - costs about 6k5 to 7k baht].

So, thus I am deeply thankful to msft, brain and jerry for their efforts, however my enthusiasm is a bit down compared with 4-6 months ago (see my posts that time), and I start to move slowly to the other side of the barricade, into the crowd of the people arguing that GPU's should be only used for TF for now on (last year I was strongly arguing against such a "stupid" remark!)

But to finish in a positive note: I like the output format of the newer versions, with real time and ETA. Please do not change it back as happened in the past :P At least this is an improvement... All developers should use the last source, and if one of you makes changes or improvement, better tell to the others too...[/QUOTE]

Well, if my final test on 1.61 is not good, I'm inclined to move back to 1.49 and update it from there.

I too, agree with you about the problems and I get very irritated with all the problems. I just know that none of these products will mature without people to develop and test them. I gave up on my 100% productivity a while back because I realized that computer problems are going to happen. In this case, we're having to test the versions to get it right. I'm willing to do that as long as I have time. I wish it didn't take so long to find errors, but that's the cost of testing.

My goal is to find a version that is stable and make sure it includes all the features that people want. I don't have the programming skills to implement everything, but I can certainly contribute to debugging and other things.

The best we can get from everyone now is lots of tests on all versions from 1.49 up to find out where the error is occuring. It could be my video card or the code. The more people who post results (good or bad) will speed up the process and very soon we'll have a version that is fast and stable.

Thanks to everyone who has contributed and is posting and testing here.

Anyone here in need of a CUDA-memtest? :smile:

Luigi

[QUOTE=ET_;291273]Anyone here in need of a CUDA-memtest? :smile:

Luigi[/QUOTE]
Nice materials to read (the pdf and docs inside of the zip). Thanks.

P.S. Any win64 builds for cc 2.0?

There are alternatives called [URL="https://simtk.org/home/memtest"]Memtest G80 and MemtestCL[/URL]

[QUOTE=Karl M Johnson;291278]There are alternatives called [URL="https://simtk.org/home/memtest"]Memtest G80 and MemtestCL[/URL][/QUOTE]

But you have to be registered to download... :smile:

Luigi

[QUOTE=LaurV;291276]Nice materials to read (the pdf and docs inside of the zip). Thanks.

P.S. Any win64 builds for cc 2.0?[/QUOTE]

Working it... I have to convert the makefile first :cool:

The makefile indicates it will work with AMD, as well. I'll figure that out later...

[QUOTE=Karl M Johnson;291278]There are alternatives called [URL="https://simtk.org/home/memtest"]Memtest G80 and MemtestCL[/URL][/QUOTE]

[QUOTE=ET_]But you have to be registered to download... :smile:[/QUOTE]

I have the zip file of MemtestG80. Unfortunately, it is 897KB: too big to upload here. The COPYING and README files only disclaim liability and forbid use on a machine which you don't own.

If anyone wants it, give me a place to upload it, or PM and I'll email it.

Use Mediafire or Sendspace to upload it.