Announcement

Collapse
No announcement yet.

Just a quick question

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • Just a quick question

    I read somewhere that AMD (I think it was AMD) graphics cards require at least one CPU core (on a multicore system) to receive data from in order to draw whatever is on screen. I read this is usually the case on Linux more than it is on Windows.

    Does anyone know if this is true or not?

  • #2
    Unless you're talking about actually drawing things on the screen (which would need CPU to prepare commands for the GPU) the only similar thing that comes to mind is an IGP connected via HT to a discrete CPU. In that case the memory controller on the CPU needs to be active (along with any other bits between MC and HT) in order to let the iGPU's scan-out logic access data from the CPU's memory.
    Test signature

    Comment


    • #3
      Originally posted by bridgman View Post
      Unless you're talking about actually drawing things on the screen (which would need CPU to prepare commands for the GPU) the only similar thing that comes to mind is an IGP connected via HT to a discrete CPU. In that case the memory controller on the CPU needs to be active (along with any other bits between MC and HT) in order to let the iGPU's scan-out logic access data from the CPU's memory.
      I was just curious since in regards to playing TF2 on Linux, I have to force the game to render on a single core (mat_queue_mode 2 causes stuttering and huge graphics lag). This isn't the case on Windows, since I have TF2 running with mat_queue_mode 2 and -threads 4.

      I mainly just read that the CPU->GPU feeding data is heavier on Linux than it is on Windows in regards to running a dedicated AMD GPU. Like if I have a game that can use 4 cores, Core0 would be dedicated feeding GPU while the game renders on Core1,2,3 to prevent the GPU from lagging behind in drawing the scene. I read that NVIDIA GPU doesn't have this problem and doesn't require a single dedicated CPU core just to feed data to the GPU.

      I had this some time ago, wish I had bookmarked the site which I read it on. Basically wondering if this could be the issue with a lot of performance problems in regards to, for example, Source engine games and AMD GPU on Linux. Source engine is heavily CPU dependent and to not have stuttering forcing the game to render on one single core (mat_queue_mode 0) could be impacting performance since it may be rendering on the same core that is trying to feed the GPU.

      I am also, of course, referring to the respective proprietary drivers and not the open source ones.

      Comment


      • #4
        That is sounding vaguely familiar, but I thought the context was open gfx driver rather than Catalyst.

        IIRC the issue was something to do with CPU power management, ie that with the work piled onto a single core the clocks were kept higher than when the same work was spread around multiple cores. Take this with many many grains of salt though...
        Test signature

        Comment


        • #5
          Originally posted by bridgman View Post
          That is sounding vaguely familiar, but I thought the context was open gfx driver rather than Catalyst.
          It may or may not be an issue present in both Catalyst and radeon drivers. I would test it out, but sadly my HD6450 won't work with the radeon drivers since it doesn't support KMS and I have to have nomodeset (which leaves me stuck on crappy vesa drivers), otherwise I get severe graphics corruption.

          Catalyst is the only option for me currently. :/

          Comment

          Working...
          X