Basic steps :
Install nvidia Proprietary Drivers (mainly CUDA-oriented instructions)
Install additional SDK stuff for OpenCL
Prove that the install worked
Download and install the nVidia driver
For the official links, see :
The 64-bit linux driver takes me to :
So that I can download :
NB : This has to be downloaded via a browser, so, since the target machine is (essentially) headless,
I have to download to a regular machine and then rsync it over.
Prepare the Machine
Before this will work on a ‘virgin’ Fedora Core 20 machine :
Install the nVidia driver
Need to stop X to install drivers, and make the script executable :
During the running of the script :
Rewrite the initramfs to prevent
nouveau being loaded
Having done that :
NVidia GPU Computing SDK
The SDK includes the OpenCL interface stuff, though it doesn’t advertise it much.
Go to :
The latest versions of the GPU Computing SDK and the CUDA Toolkit (which is required to compile the SDK code samples) are available on the CUDA Downloads Page.
Now, sync up git
So further steps can be created remotely and synced seamlessly.
As local user :
Amend ~/.bash_profile to include :
Install the CUDA samples in personal directory (for fun)
Appears to work!!
- … 15mins elapses for all of compilation (on a i7-4770) …
Run some of the provided scripts as a (CUDA) test
Useful resource :
Check that the correct ICD files appear to be installed in :
Now install the generic
opencl headers, etc :
./cl-demo 1000000 10
Entirely unhelpful side-track ( Please ignore )
The following seemed like a blind alley : There’s no need to do this…
- From https://developer.nvidia.com/opencl “Device Query”