so, avoid 960/970/980/titan and 7.0 cuda drivers, got it.

[QUOTE=firejuggler;402575]so, avoid 960/970/980/titan and 7.0 cuda drivers, got it.[/QUOTE]

CUDA 7.0 capable drivers seem to be OK, mfaktc binary compiled with [B]CUDA Toolkit 7.0[/B] triggers the bug (inside the driver?)
At least for Titan X it will be impossible(?) to find a driver not capable of CUDA 7.0.

To make a long story short: use CUDA 6.5 binaries and you don't have to care about this bug.

Oliver

[QUOTE=firejuggler;402575]so, avoid 960/970/980/titan [COLOR=Red]X[/COLOR] and 7.0 cuda drivers, got it.[/QUOTE]
FTFY. The plain titans and blacks are ok (in fact, my titan is faster with 5.5, no reason to use 7.0)
[edit: just to clarify, titan Z should also be ok, as it is in fact two plain titans fuzed together]

GTX 960/970/980 and Titan X are OK, too (indeed they are very nice cards!).
Just don't use CUDA Toolkit 7.0 when compiling mfaktc.

Oliver

subcc bug in cuda 7.0
 

This is a bug confirmed by Nvidia in CUDA Toolkit 7.0, they have a fix that will be released in the next toolkit release after 7.0.
The bug concerns sub with carry -- the carry being sometimes set when it shouldn't.
I suspect it may be wrong assembler generated by ptxas (as the PTX is correct).

The workaround is to compile with CUDA toolkit 6.5.

where/when did they confirm? My bugreports last action was just
[QUOTE]Status changed from "Open - pending review" to "Open - in progress"[/QUOTE]

I'm not even sure whether this is a bug in the CUDA toolkit or the driver itself, last time they fixed it with an updated driver.

Oliver

subcc bug in cuda 7.0
 

My bug report to them is here: [url]https://developer.nvidia.com/nvbugs/cuda/edit/1642061[/url] but I suppose only I can view that bug (unfortunately).

Here's a brief on that bug report:

The code below generates wrong behavior for the multiword subtraction with borrow sub() routine, where a spurious borrow takes place. Basically:
{0, 0, 0, 0, 1, 0} - {0, 0, 0, 0, 0, 0} returns {0, 0xffffffff, 0xffffffff, 0xffffffff, 0, 0}.

Here is the source for reproduction: [url]http://pastebin.com/Ab7d8YAh[/url] compile with: nvcc -std=c++11 --gpu-architecture sm_50 bug.cu I ran it on GTX 980.

On that bug this is the last update from Nvidia:
"This issue has been fixed in our development versions, and the fix would be available for you in the next CUDA release that following of 7.0. Thanks again for the reporting!"

There is usually some display lag when mfaktc is running. However, my computer is running a lot smoother now even though I did not change any mfaktc settings. Could this be related to the latest driver update? Anyone else seeing the same thing?

[QUOTE=ixfd64;402917]There is usually some display lag when mfaktc is running. However, my computer is running a lot smoother now even though I did not change any mfaktc settings. Could this be related to the latest driver update? Anyone else seeing the same thing?[/QUOTE]

Lag still comes and goes for me, with driver 350.12. If I want the screen to be responsive enough for playing video or editing images, I restart the 580 that runs the display with something like 'GPUSieveSize=40'.

subcc bug persists in CUDA 7.5 RC
 

CUDA 7.5 RC is available for registered developers.

Using [B]cuda_7.5.7_rc_linux.run[/B] (includes driver 352.07 and nvcc 7.5.6) the problem is [B][U]NOT fixed[/U][/B].

So for now:[LIST][*]don't use CUDA toolkit 7.0 for [B][U]compiling[/U][/B] mfaktc[*]don't use CUDA toolkit 7.5 RC for [B][U]compiling[/U][/B] mfaktc[/LIST]
I repeat myself: using a CUDA 7 capable [B]driver[/B] is okay for mfaktc!

Oliver

Hi everyone,

just a little update: It is not only "the subcc"-bug, there are at least two issues with CUDA 7.0 and 7.5RC in regard to mfaktc.
Nvidia told me that the subcc stuff is fixed in their internal built and will be included in the final CUDA 7.5 package.
Nvidia told me that they have fixed the other issue(s) in their internal built, too. Unfortionally this fix won't be included in CUDA 7.5... They'll fix this in a "feature release".

So my previous [URL="http://www.mersenneforum.org/showpost.php?p=405539&postcount=2562"]post[/URL] is still valid and there is a high chance that CUDA 7.5 (final) needs to be added to the list. :sad:

The good news are that there is (right now) no need for CUDA 7.x in mfaktc:[LIST][*]no new GPUs supported by CUDA 7.x.[*]I didn't notice any stuff which might increase the performance of mfaktc (such as the "funnel shift" in CUDA 5.0)[/LIST]
Oliver