Announcement

Collapse
No announcement yet.

AMD Tonga & Fiji Open-Source Performance Boosted By PowerPlay Patches

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

  • #51
    Originally posted by drSeehas View Post
    Well, as you can read in this thread, people tried to convince me there is logic in the marketing names. I am just to stupid to see and understand this logic.
    Yes there are great amount of logic in marketing too, if you try to understand target. For that cosumer target, that sort of info is considered "enough" and of course any of them has its limitation.

    In the end if you really need more and more information about particular product not desribed in any publicly available documentation i am afraid you need to start to work for that company
    Last edited by dungeon; 16 November 2015, 04:57 AM.

    Comment


    • #52
      Originally posted by rabcor View Post
      (what good is high bandwidth memory on a low bandwidth bus? it's bandwidth starved)
      FYI, you're supposed to use "real" GPUs like this: you upload all data you need to compute into its fast VRAM and only use high-bandwidth RAM. Which could be order of magnitude faster than system DDR3 RAM if GPU got reasonable GDDR5 RAM on wide bus. And HBM is further step in making RAM bus even more wide. Not to mention PCI-e transfers. You see, GPU and its memory is faster than rest of system by several times. Either it can do on its own, using VRAM for virtually everything, or you have trouble on the way, crashing to performance barely looking like RAMless integrated stuff, rather than high-end GPU thnig. And even fastest bus can't do much about the fact system memory is like an order of magnitude slower than VRAM on graphic tasks, dammit.

      For this reason it is absolutely mandatory for scene data to fit VRAM. And that's why you need more VRAM on higher resolutions. Should it fail to do so... fetching texture over PCIe is like an order of magnitude slower and your FPS going to crash. And system RAM would not cope with it, no matter what, even if you would have blazing fast bus. Don't you mind DDR3 ICs are only being used on low-end cards? And there is good reason why it happens this way.

      So from technical point of view, HBM is EPIC WIN on its own - it gives a lot of headroom for improving RAM bandwidth further. So GPU can have more headroom. And no, you're not supposed to do PCI-e xfers to compensate lack of VRAM. There is no magic. Either you have enough VRAM, or you don't. Anything else is similar to "swapping" and MUCH slower. Your whole system is no match to what GPU and VRAM can do. You can't overcome it. The whole point of dGPUs is they have turbo-fast local memory and shitload of SIMD-like engines. If you can program it properly, it turns to real powerhouse, beating CPU and its memory subsystem by orders of magnitude, literally. And if we assume your main system is comparable to GPU... you do not really need GPU at all and can use main system to do the very same things, right?
      Last edited by SystemCrasher; 17 November 2015, 02:39 PM.

      Comment


      • #53
        I have been eying a new R9 380 as and upgrade. Just need to wait for all the PM patches to be available.

        Comment


        • #54
          Hmmm, something nice happened when I tried the latest powerplay patches with kernel 4.4-rc2: http://openbenchmarking.org/result/1...HA-1510216HA93

          Comment


          • #55
            Does a Ubuntu PPA with a PowerPlay enabled kernel exist? Someone is going to have a 285 under the yule-tree this year :-)
            (Got a 285 for the price of a 370)

            Comment


            • #56
              Not that I'm aware of. I'm building an update now, instructions are roughly:

              git clone git://people.freedesktop.org/~agd5f/linux linux
              cd linux
              git checkout amdgpu-powerplay

              Download the patches from here: http://kernel.ubuntu.com/~kernel-ppa...v4.4-rc3-wily/

              git apply them

              Follow these instructions: https://wiki.ubuntu.com/KernelTeam/GitKernelBuild from step 4 onwards.

              Improvements to that process are welcome.

              Comment


              • #57
                Originally posted by BenPope View Post
                Hmmm, something nice happened when I tried the latest powerplay patches with kernel 4.4-rc2: http://openbenchmarking.org/result/1...HA-1510216HA93
                Nice, that's more like it!! I wonder what that was...

                Comment

                Working...
                X