2
After a kernel fails due to "unspecified launch failure", the only way I am able to use the device again is after a system reboot. Is there another way to do this? If I attempt to to run again after the ULF, the program hangs on the first CUDA call.
The Nvidia card is attached to a Windows 7 Professional 64-bit Virtual Machine running on a Parallels Extreme Workstation. I connect to the system via Windows' Remote Desktop Connection.
0
If you were using this on a non-virtual machine then the unspecified launch failure (ULF) would prevent you from using the GPU again in the same process but not cause problems for any other processes.
Since you are using this within a virtual machine it sounds like an issue either within Parallels or in the driver, so I would suggest that you contact Parallels to investigate the issue further. If you are a registered developer on the NVIDIA site you could also submit a bug there.
Tom
Posted 2011-09-02T00:34:30.933
Reputation: 101
1I thought you couldn't use Remote Desktop Connection to run CUDA as this method wouldn't make the graphics cards available to use. On XP at least I can't use RDC and CUDA. I currently use VNC but was it changed for Windows 7? – None – 2011-09-02T14:18:08.947
Does calling
cudaDeviceReset()(it's only present in CUDA 4.0) after ULF helps? However, such behavior indeed is wrong. @jmsu There is a way to use nVidia GPUs via RDP. It requires having Tesla and doing some changes in registry (or you can use older version of driver, which does this trick with any nVidia GPU). – aland – 2011-09-06T17:28:08.287