This has been a problem with Microsoft's Remote Desktop implementation since the very beginning of CUDA - I believe it unloads the NVidia driver and substitutes one of its own, which is not CUDA-aware.

However, people have reported success with other remote access implementations, such as the various flavours of VNC.

I use the TeamViewer, free for private use and not commercial, it's good to remotely control Windows systems.
____________

I have remotely connected to a number of computers that have GPU's crunching for GPUGrid.

On every occasion the GPU is reported as no longer in a working condition. Usually the running tasks crash!

The only way to resolve this is to log on locally, shut down Boinc and then Start Boinc again. On one occasion I had to restart the system.

Remotely connecting to a computer that does not have a GPU does not cause a problem, just computers with GPUs that crunch.

The big problem with fast user switching is that GPUGRID tasks immediately end with computation errors, and then more work is fetched, then each new task fails as well. I quickly use up my daily quota and am out of work for a day. Isn't there any way for GPUGRID tasks to detect this condition and/or just go to sleep when CUDA is not available?

think windows remote desktop loads generic vga driver for the session and that crashs the cuda aps, VNC viewer ftw :)
____________

I use TightVNC with about 98% success rate. Somtimes from the master PC, I have to shut done Boinc on the master to remote into other computers.

Also I've had multiple failures in a row in quick sucession but have never been limited to 2 WU's in a day even thogh "Messages" said I had reached my limit of 2. Maybe the server needs to catchup or ....

UltraVNC is what I use for remote connection.

VNC works very well - it won't unload the NVIDIA drivers. TightVNC is what I use.

UltraVNC is great. I use it as well. I've also got Logmein installed, but I've not tried it on my gpu cruncher yet, *cross fingers*