4 Replies Latest reply on Nov 1, 2010 9:16 PM by laobrasuca

    Combining ATI & NVIDIA cards under Linux

      Are there any gotchas I should be aware of?

      Hi all-


      I have in the past programmed for CUDA using NVIDIA GPUs, to good success. Now I'd like to expand into OpenCL, using ATI cards, and I'm trying to educate myself. A few intro questions:

      Is it possible to have both an NVIDIA card and an ATI card in the same machine, and have their linux drivers play nice together?

      Does it matter which card is providing the video output? Can I do all my work remotely via SSH?

      Do I absolutely have to restart after installing the ATI linux driver?

      Thanks for any words of wisdom / anecdotes, and I look forward to joining the community!

        • Combining ATI & NVIDIA cards under Linux

          Yes, it is possible for ATI and NVIDIA to work on the same host. However, I have not been using both cards (HD 5870 and GTX 480) concurrently. That is only as I have been purely focused on kernel optimization (rather than solving real computational problems) so use each card separately. Sorry that I can not say for certain there is no interaction at runtime.

          Video output does not matter. Note that some mainboard BIOS'es do care. The system won't boot depending on the which GPU is the video output. It's a BIOS issue. I have a BloodRage and had to upgrade to a patched Foxconn BIOS to get it to boot with the GTX 480 in the primary PCIe slot. I recall reading of similar issues with ATI cards.

          So yes, headless works with SSH. That's how I work. No monitors on the GPU host. I work off a laptop and ssh into the GPU host.

          No, you do not have to restart after installing the ATI Linux driver. I never do. Note however that I install both the NVIDIA and ATI drivers in a headless configuration, i.e. without a running X server. So I see a curses installer application. My guess is that if you install using an X GUI installer, then you may have to restart X or unload kernel modules in order to pick up whatever the installer installed?

          Here is an old thread about this: http://devforums.amd.com/devforum/messageview.cfm?catid=328&threadid=117913&highlight_key=y

            • Combining ATI & NVIDIA cards under Linux

              sry for double posting, but the other discussion is on the ATI Stream section of forum and i'm pretty sure that most of people who check opencl forum don't do the same for the ATI Steam. Here's my reply to Ceq in the older thread:


              Originally posted by: Ceq Try using Win7 x64 instead of WinXP, it offers less GPGPU performance but it worked for me. I installed the Radeon's driver first using the installer and then the GeForce, but using the device manager instead of the installer. PhysX didn't work, however OpenCL and CUDA do, so I don't mind at all about PhysX. 

              hi Ceq,

              I was trying to install both HD5870 and GTX460 on my Win7 x64 system just as you mentioned, but unfortunately ever since I run the

              status = clGetPlatformIDs(0, NULL, &numPlatforms);

              to get platform info, the code crashes with an access violation in a given reading location (0x0000013c). What driver versions you used?

              I have hooked both cards in my system and each one is cabled to a different monitor. After boot, windows prepared the cards and I rebooted. After reboot, I cleared every Radeon/Geforce software (display drivers, opencl sdk and so on). After rebooting, 1- I installed the ccc 10.10 with display drivers and the included opencl sdk (v2.2), 2- I went into the device manager and updated the gtx 460 driver using the 260.99 version.

              Any suggestions?

              ps: i had previously installed two amd cards together (HD5870 and HD5770) successfully (could run opencl on both cards separately).

                • Combining ATI & NVIDIA cards under Linux

                  About NVIDIA and ATI card on one host on linux. I use both concurrently and they work pretty well. when i run CUDA on NVIDIA and OpenCL on ATI applications these is no penalty on speed.

                  I work only via ssh.

                  But i my always installed  NVIDIA driver, tools and SDK first then ATI driver and SDK. After i rebooted system, and did aticonfig --adapter=all --initial. And  i checked PATH and LD_LIBRARY_PATH, and installed ICD for ati. Then i ran X in screen.

                  My system is OpenSUSE 11.2 x64 kernel, asus suprecomputer X58, i7,12GB DDR3,2x ATI HD5870,NVIDIA GTX470.

                  One display connected to first ATI card(1st PCIe slot) no one to NVIDIA (3th PCIe slot) and dummy VGA dongle to second ATI card(7th PCIe slot). It is possible to connecte dongle to 1st ATI card too.

                  I use GPGPU only for scientific research so no graphic output.