Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

Adept I

AMD OpenCL Driver Installation on Display-less node


I wish to clarify if in the Catalyst 10.2 driver install step, is the aticonfig --initial step necessary? My usage configuration is of a 1U server system with two 5800 series GPUs connected for OpenCL development. A seperate display adapter is used for viewing. I am using CentOS 5.3 

When after installing catalyst, I call aticonfig and reboot the system, the system goes to a blank screen when I try and start the Xserver using "init 5", I assume this is because the aticonfig sets up a xorg that makes the 5800 as a display.

I have done the ICD stuff and the OpenCL works for CPU, The cal examples dont work and Findnumdevices returns 0.

I ask this because,  the cal runtime libraries are not present in /usr/lib 

Can some one please suggest a work-around that allows me to use the ati gpus without a display on them.

I already did the steps that describe running stream apps remotely,

6 Replies

yes you need runing Xserver. and you should have libatical*.so in /usr/lib


Hello Nou,

Thanks for your prompt response. I changed our server's KVM switch and then ran ati-config --initial. 

Now I am using one of the ATI GPUs as one of the display adapters and running GNOME. This removed all the old problems of segfaults and missing stuff. The GPU OpenCL examples work now

However our server has two GPUs but FindNumDevices only shows me one. How can I make FindNumDevices return two GPUs? I tried aticonfig --initial=dual-head but no luck...

Do I actually need to hook up the other GPU to a display?

aticonfig --list-devices returns two adapters with "0" being the default one.

This is almost at the home stretch  

Thank You for your help, I am going to write up how to setup ATI GPUs in a cluster soon

-- Perhaad 



To enable both the gpus do aticonfig --adapter=all --initial

and export the DISPLAY env to :0

export DISPLAY=:0


Just to clarify before I proceed, Is that step different from the "case export" line in the /etc/bashrc that we need to add when setting up Xwindows for remote access?

Sorry but as visible Xwindows isnt one of my strengths.




normaly DISPLAY is set to :0.0 you can select which GPU you will use setting to :0.1 will use second GPU :0.2 third etc. if you set DISPLAY to :0 you will select all GPUs.

you can place export DISPLAY=:0 to .bashrc or somewhere else.


Thank You everyone,

The surprising part of all this was that even the remote stuff worked without having to modify the two files in the gdm directory after I added the export DISPLAY=:0 in my personal .bashrc file.

We now have both GPUs visible devices on each server in our OpenCL AMD cluster, Yaay....