So far, Stanford has not released a native Linux GPU folding client. I am running Stanford's Windows XP 32 bit GPU folding client in Arch i686 through WINE. To enable usage of the stream processors in the 8000 and 9000 series Nvidia GPUS for parallel processing, NVIDIA video drivers which employ the CUDA package must be installed.
I am running Arch i686 on an AMD AM2 64 bit platform with all testing updates installed. This includes nvidia-180.22 and nvidia-utils-180.22. When I try to start the Windows GPU folding client through WINE, I get the following error message:
initializing Nvidia gpu library cudaMalloc CUDAStream::allocate failed feature is not yet implemented.
Here's where it gets interesting. From a console screen, I shut down the X-server, and uninstall nvidia-180.22 using pacman -Rs nvidia. Then I downloaded NVIDIA-Linux-x86-180.22-pkg0.run which is the exact file that the PKGBUILD for nvidia-180.22 downloads and uses to extract the source code from. I then compile and install the 180.22 kernel module as such:
Reboot, and the Windows GPU client runs perfectly. I realize that this method of installing the Nvidia video driver is NOT recommended, which is why I am trying to figure this out.
Judging from the error message, it looks like CUDA is broke in the testing repo nvidia-180.22. I looked at the PKGBUILD and didn't see anything that I thought would exclude CUDA.
Anybody got any ideas why compiling and installing using the NVIDIA-Linux-x86-180.22-pkg0.run file works and the nvidia-180.22 package in the testing repository doesn't?
Last edited by Pudge (2009-01-15 02:52:08)
I went to file a bug report on this, and one already existed. Duh, I should have checked for a bug report before starting this thread. Sorry.
In the bug report they had a work around
ln -s /usr/lib/libcuda.so.1 /usr/lib/libcuda.so
to create a missing symlink. The symlink fixed the problem.