650 Ti + 430 problem


I’ve recently bought another graphics card (650 Ti) and installed into my system alongside my old 430.
Currently I’m using the 650 Ti as my display card and I thought of experimenting with my 430 as “computation” card. I plan to switch ( 430 - display; 650 - computations ).
First problem:
430 still has it’s watchdog enabled ( or so does cudaDeviceProp tell me ). I somehow thought that the card that doesn’t have a monitor attached to it will automatically have it’s watchdog disabled, but I started to dig around and apparently it’s a Windows feature and now I’m just confused on this issue.
Second problem:
My program multi-threads. The main thread gets all the MFC stuff, initialization and whatnot and my worker thread only does the kernels. I do this so as to not make the program appear as if it has frozen when I’m executing the kernels. This works fine when I select my primary card ( 650 performs as expected ), but when I select the 430, the kernels called from the worker thread end with cudaUnknownError. Previously to installing the 650, the 430 worked fine in this situation ( initialized from one thread and did the work as expected from the worker thread ).

Has anyone encountered a similar problem?

Also, GPU-Z for some reason shows that CUDA is disabled ( an empty checkbox ) when showing the properties for my 430.

You don’t mention OS, I’m assuming Windows Vista/7.

Any card under WDDM model will perform slightly worse than under Linux. If TCC mode if available – only on professional (e.g. Tesla) cards, that also avoids the overhead on Windows.

To avoid problems with kernels that take a long time to execute, you can disable the TDR (timeout detection and recovery) feature via the Windows registry. I suspect your second problem will disappear when you accomplish this. Place the contents below into a new text file, and save it as disable_TDR.reg, and run it. Make sure to reboot your system afterwards to have it take effect.

Windows Registry Editor Version 5.00


Disregard the GPU-Z output of CUDA checkbox, the author of the program isn’t the best at giving precise info for all the cards.