I am having big problems when using cufft on my evga titan card. I have an iterative process which uses at each step a few transforms and it always crashes on the system size is large. By comparison my 660 Ti cards and the GT 650 had no issues. When I tried different performance modes and different fan speeds. When I keep the fan speeds up the temp is about 65-70 C and the program just crashes, when I use auto for the fan speed the program first starts to have errors which accumulates and then crashes. I am “lucky” I can see the problem easy, by just looking at a few numbers so I can see imediatley if there is problem. I use double precision and the double precision is enabled in the nvidia-settings.
I already sent on card back and I am planning to sent this one too, but I would rather have a working card than just send the cards to them. I did pay a lots of money and I would like to work on it.
According to this forum http://petachem.com/forum/index.php?topic=295.0 (petachem program), mersenne forum ( http://www.mersenneforum.org/showthread.php?t=17834&page=7 ) and amber forum ( http://archive.ambermd.org/201307/0227.html ) from June there are serious problems with cufft on titan cards which are not just. According got the folks which use amber nvidia was able to reproduce the issue and are working on a fix.
Could someone from nvidia comment on this and tell me what can I do. This bug is not just heating. I am able to reproduce the problem on 2 different cards with different linux one with DP activated and one without DP. The k20 card has no problems. I would be willing to downcloak the card to the k20 speeds if this would solve the problem and provided with an appropriate BIOS.