[QUOTE=aaronhaviland;245691]I can't find 0.14 anywhere. Is it out?[/QUOTE]

No, not yet.

[QUOTE=aaronhaviland;245696]So true! Switched to a GTX 460 yesterday (clocked at 875MHz), still seems better without MORE_CLASSES, until I get to the 2^69-2^70 ranges.

Also, I've found that I need to lower SIEVEPRIMES much lower...
I reduced the lower limit and found that it liked to hover around 4500, if I'm not also using the CPU for other intensive tasks. If I am, it likes it around 500.[/QUOTE]

Please don't change SIEVE_PRIMES_MIN/MAX unless you're really sure what the consequences are. E.g. [B]Increasing MAX above 100k will produce false results in some cases[/B] (24bit overflow in k_tab[] possible) and lowering MIN that small is usually just a waste of energy. Perhaps I'll lower MIN a little bit in feature releases if this is really wanted. If a single core of your CPU isn't able to feed your GPU fast enough just dedicate another core (another instance of mfaktc) or just accept the fact that you GPU isn't utilizated 100%.

Oliver

mfaktc 0.14
 

Hello,

find attached mfaktc 0.14. :smile:

Highlights of this version: The [B]barrett79 kernel[/B] is a little bit faster and no longer limited to single bitlevels. On my GTX 275 (compute capability 1.3) this kernel is [B]3-4% faster[/B] than before and on my GTX 470 (compute capability 2.0) it is [B]~10% faster[/B] than before. The README is updated, too. It includes instructions how to get work and report results from/to the primenet server.

Fixed a potential application crash (exponents just above 1M and SievePrimes close to 100k) and the debug code received some changes, too.

The other issues which James has recently discovered have to wait until mfaktc 0.15, sorry (I've allready started coding for 0.15).

Oliver

P.S. for those nice persons who build Windows binaries and provide them to others here, please include the following files with your binaries:
- COPYING
- mfaktc.ini
- README.txt
(- Changelog.txt)

A win-64 binary of mfaktc 0.14 (complete with all files specified above :smile:) can be found here. win-32 binary to follow.

[url]http://www.sendspace.com/file/x0sijm[/url]

... and the win-32 binary.

[url]http://www.sendspace.com/file/48y6i3[/url]

I have also attached below the win-32 Makefile that I use as this is not yet included in Oliver's tarball.

Mfackt 0.14 bench, gtx460; model name GV-N460OC-1GI (gigabyte model), core 2 duo 8300; on Win XP.

Should i get another instance, both run at 80 M/s

here is with 2 instance running. should i change the number of CPUstream? ( atm 3, Numstream too)

I cant execute Amphoria's latest x64 binary.
I get this error on start: cudaStreamCreate() failed for stream 0 .
However, 32 bit binary works like a charm !
Ideas ?
They both use same mfakt.ini , and both were fed the same worktodo.txt file, so it's not the config that causes such behavior.
Also, I get the same error with and without any cudarts in the folder, so it seems it crashes before calling it.

Hello,

[QUOTE=firejuggler;248807]here is with 2 instance running. should i change the number of CPUstream? ( atm 3, Numstream too)[/QUOTE]

2 instances look reasonable. You can try to adjust SievePrimes manually in mfaktc.ini (try ~6000 on your system) and set SievePrimesAdjust to 0.

Looks like the Win32 binary has some issues with the screen output, those lines should start with "xxxx/4620" and not those big numbers you get. But this is just cosmetic...

[QUOTE=Karl M Johnson;248901]I cant execute Amphoria's latest x64 binary.
I get this error on start: cudaStreamCreate() failed for stream 0 .
However, 32 bit binary works like a charm !
Ideas ?
They both use same mfakt.ini , and both were fed the same worktodo.txt file, so it's not the config that causes such behavior.
Also, I get the same error with and without any cudarts in the folder, so it seems it crashes before calling it.[/QUOTE]

Ideas *hmm* as usall: driver- and CUDA-version, what does the "CUDA version info" output of mfaktc say?

Oliver

Amphoria, what were the drivers you compiled mfakt under ? The toolkit/sdk was 3.1, for sure.

No complaints: Amphoria's latest x64 binary works perfectly here.

Then Amphoria must be using 26x.xx Forceware then.
And, it seems, 64 bit CUDA apps, which were compiled using 26x.xx and any toolkit/sdk, cant be executed properly on machines with < 26x.xx using cudart shenanigans.