I'm developing a bit of OpenCL software and would like to optimize my software for both the GeForce 480 and Radeon 5870. I'd also like to generate accurate performance comparisons between the two versions. To that effect I'd like both cards to be able to operate without having to "drive" a monitor (or any calculation other than mine for that matter).
I've already managed this with the GeForce, executing OpenCL software over ssh. But I just got my new Radeon and have spent the last few days trying to figure out how to do the equivalent.
Before I get ahead of myself, it's worth mentioning I'm running Fedora 12 64bit with Catalyst 10.5 drivers.
I've gotten the Stream SDK installed, compiled and all that jazz. Things work smoothly when I've got my monitor hooked up to the Radeon; CLInfo returns both my CPU and GPU devices and FindNumDevices returns 1. However, if I drive the monitor from the nVidia graphics card, both programs fail to recognize the Radeon card. Also, if I try to run either remotely via ssh, neither finds the Radeon (this, or something similar is my eventual goal).
I know that CAL requires a running X server to function properly. I've tried spoofing it with Xvfb, using Xvnc, and managed to get Xorg running (somehow) without an attached monitor. But still, no luck identifying the Radeon card without having a monitor attached to it.
I've tried the "Running ATI Stream Apps Remotely" tips and I've even tried removing all nVidia drivers and hardware from the machine to no avail. I'm out of ideas. Is true (without a spoof dongle, etc.) headless remote execution of CAL/OpenCL code even possible?