Announcement

Collapse
No announcement yet.

Modern AMD Switchable Graphics Detection Issues

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

  • Modern AMD Switchable Graphics Detection Issues

    Hello!

    I've recently acquired one of the 2020 "AMD Advantage Edition" notebooks, in particular the ASUS Strix G15 Advantage Edition - Model G513QY from BestBuy. This notebook has the display, and HDMI port routed through the onboard Vega iGPU found within the Ryzen 9 5900HX. This is the only notebook I'm aware of that is shipping with the RX 6800M GPU. The 6800M is routed to the HDMI and Display via a Mux with the APU. Only the USB-C port, which is wired to allow Display Port Alt-Mode v1.4, is directly able to access the raw 6800m without involvement from the iGPU.

    Onto the issue then:

    Unlike an Nvidia+Intel configuration, which allows you to flag what programs to run on what GPU from the Nvidia Control Panel, AMD appears to favor the Windows 10/11 Graphics interface to handle this. I'm still running Windows 10, at the moment just for context. This appears to be the core of the issue. While running on explicitly the 6800M only, via DisplayPort Alt-Mode, my scores are significantly higher in both 2D and 3D, which led me to the AMD performance overlay which confirmed my suspicion. The 6800M was only being used for maybe 1 or 2 of the normal tests while it had the option to choose between GPUs for rendering.

    With some extremely rare exceptions, Windows normally assigns these fairly accurately. Minecraft was the only other notable exception in my month of ownership. Dozens of other programs all seem to work fine, though Minecraft is notorious for tripping up Nvidia's Optimus as well. To try and combat this, I opened up the Win10 "Graphics settings" panel and added every .exe file I could find related to PE10.2 from the installation directory. Once added, I set them all to "High Performance (6800M) and ran PT10.2 again.

    This appears to run the majority of the tests on the 6800M, but not all of them unfortunately. While I would normally just run everything without the APU MUX in place (via the 6800M DP out), I'm trying to determine the penalty cost invoked by the Mux switch on the 6800m. This is a measurable affair after a quick browse around the web regarding "performance loss from mux switch" including by AMD themselves, so I was trying to nail out the difference with my exact setup.

    Any suggestions are welcome, and I'll gladly pass along any information you may be curious to see. In the past I've helped out with some DX11 testing before PT8 came out if I recall (still have that email chain somewhere I believe), so I'm not afraid to run test builds or whatnot if it helps out. Dual AMD GPU notebooks are still a little rare these days, though, they are anticipated to pick up steam big time this year if AMD is to be believed.

    Duncan

  • #2
    This appears to run the majority of the tests on the 6800M, but not all of them unfortunately.
    Which tests still seemed to be running on the integrated card?

    Comment


    • #3
      Sorry for the delay Tim, I did a bit more digging than was necessary! Below are my condensed notes from the last few hours of testing things.

      ---------------
      Vega iGPU (Mux Active, All .exe set to "prefer 6800M")

      2D:
      Unbuffered Solid Vectors
      Back Buffered Solid Vectors
      Unbuffered Transparent Vectors
      Back Buffered Transparent Vectors
      Fonts and Text
      Windows Interface
      Image Filters Rotation
      Image Filters Brightness
      Image Filters Grayscale
      Image Rendering

      3D:
      Passmark OpenCL Particles Simulation (50% load)

      ---------------
      RX 6800M (Mux Active, All .exe set to "prefer 6800M")

      2D:
      Fonts and Text using D2D
      DX11 Rotation
      DX11 Brightness
      DX11 Grayscale
      DX11 Image Rendering
      Direct2D Bench (50-80% load)
      PDF Renderer
      Direct2D SVG

      3D:
      DirectX9 Test (82% load at start, then swings wildly between 18-80% throughout remainder of run) - Score of 42
      DirectX10 Test (100% load)
      DirectX11 Test (100% load)
      DirectX12 Test (100% load)
      Direct Compute (100% load)
      NBodyGravityCS11 (100% load)
      Direct Compute QJulia4D (100% load)
      Passmark OpenCL Particles Simulation (50% load)



      ---------------
      The behavior regarding DX9 doesn't change if I bypass the iGPU Mux either. I have a Surface Pro 6 with an UHD620 that scores 23, and a GTX 1060 Max-Q Dell notebook that scores 148... so something is definitely off there. I "think" this has something to do with AMD's Navi Boost/Downclocking algorithm being hyper aggressive, but that's over my head at the moment. Not sure if there is a flag that can be set to force/trick the GPU into a higher power state to combat that? My Desktop's 6900XT (also RDNA2), of all things, appears to be affected by the same downclocking bug (it scores 1017, but my old 1070Ti scored 1166!). This DX9 test appears to be the second such program I've seen actually affected by it so far, with Minecraft being the other.

      With everything configured according to my description, and the behavior laid out above with the DX9 test, the only things that appears to run on the iGPU are the non-DX based 2D items, and some strange utilization behavior on the Particle Simulations I can't explain.
      ---------------
      ---------------
      Below are the tests that the 6800M correctly runs with the Mux active, and Windows Graphics Settings set to "Let Windows Decide" on all Performance Test executables. Everything not listed below ran on the Vega iGPU.

      6800M ("Let Windows Decide")

      2D:
      Fonts and Text using D2D
      DX11 Rotation
      DX11 Brightness
      DX11 Grayscale
      DX11 Image Rendering
      Direct2D Bench (50-80% load)
      Direct2D SVG

      3D:
      DirectX9 Test (same results as above)
      DirectX10 Test


      Hope this helps!

      ----Quick Side Tangent - Did want to quickly mention that there is appears to be a bug in PT10.2 B1001 where on both of my Radeon 6000 PCs (Vega+6800M & 6900XT) freeze a frame of the DX9 test onscreen on top of everything once it ends, with the only thing that can display over it being a Fullscreen 3D application. This remains until Win+D is pressed, or PT is closed via Task Manager, or the PC is powered off.

      Comment


      • #4
        For the OpenCL test if you launch PeformanceTest in debug mode, run the GPU Compute tests and then send us the log files we can check what video cards the OpenCL test is seeing.

        For the rest of the 2D tests if they aren't hardware accelerated then likely the Windows graphics settings will be ignored, I don't believe there is really anything that can be done in this instance with the way the dual graphics work in this instance.

        id want to quickly mention that there is appears to be a bug in PT10.2 B1001 where on both of my Radeon 6000 PCs (Vega+6800M & 6900XT) freeze a frame of the DX9 test onscreen on top of everything once it ends,
        This sounds like a driver bug.

        Comment


        • #5
          Attaching the file requested, as well as a few others. Filenames of all the included logs and PTX files should provide all the necessary details and context. Curiously, on this batch of runs on the notebook, I was unable to replicate the DX9 test being stuck onscreen. I suspect that may actually be due to another software I had running at the time. For all attached runs, all ancillary testing and monitoring software** was not running.

          While reviewing the SysInfoLog.txt file, nothing particularly helpful stood out to me, though I did find a typo.


          16.125s - SysInfo: GPU ATI 8 adpaters found



          **Edit - The extra software being MSI Afterburner, RivaTuner, and HWiNFO.
          Attached Files

          Comment


          • #6
            Thanks for that, we've made a change in the OpenCL test to try and choose the non-integrated option if there is more than one card of the same platform, could you please unzip this updated version into your PerformanceTest install directory and let us know hoe it goes.

            Comment

            Working...
            X