Announcement

Collapse
No announcement yet.

Linux 3.5 Can Massively Boost AMD Radeon Graphics

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

  • #11
    This is a really impressive performance increase. Still not enough to compete with Catalyst, but more than just noteworthy. I'm curious to know if this affects significantly older generations of GPUs, like the pre-HD2000 series cards.

    It's too bad Debian sid doesn't like to stray from the 3.2 kernel, I guess I won't be experiencing this for a while.

    Comment


    • #12
      So what's the bottleneck for the 6450? Call it "crippled" all you like, those results are anomalous and should probably be investigated.

      Comment


      • #13
        Alex would know for sure, but my guess is that it only has a single memory channel so improving the way accesses are mapped across memory channels doesn't help.

        Comment


        • #14
          The issue that was fixed only affected cards with >2 render backends (rv740, cypress, juniper, barts). Other chips were not affected. Parts like 5450 and 6540 are generally memory bandwidth limited as they have fewer memory channels and often only have ddr memory rather than gddr memory.

          Comment


          • #15
            Originally posted by Wyatt View Post
            So what's the bottleneck for the 6450? Call it "crippled" all you like, those results are anomalous and should probably be investigated.
            Not really if you read the article, or patches.

            The fix is only for certain GPUs which were misconfigured previously, rv740 is the only r600/r700 card, and then a bunch of evergreen/ni cards, but only in the higher end GPUs.

            Its not meant to speed up all cards, its more that these cards were broken before now.

            Dave.

            Comment


            • #16
              Originally posted by airlied View Post
              Not really if you read the article, or patches.

              The fix is only for certain GPUs which were misconfigured previously, rv740 is the only r600/r700 card, and then a bunch of evergreen/ni cards, but only in the higher end GPUs.
              I have read the article twice, but was unable to find a link to the patches. I'm going to head over and check out the RC1 notes. Is there a summary available with more information than "fixed cards with >2 rendering pipelines"?

              F

              Comment


              • #17
                Originally posted by schmidtbag View Post
                It's too bad Debian sid doesn't like to stray from the 3.2 kernel, I guess I won't be experiencing this for a while.
                You can use another kernel like aptosid's or the liquorix kernel when 3.5 is released.

                Comment


                • #18
                  Originally posted by russofris View Post
                  I have read the article twice, but was unable to find a link to the patches. I'm going to head over and check out the RC1 notes. Is there a summary available with more information than "fixed cards with >2 rendering pipelines"?

                  F
                  The two relevant patches are:
                  http://cgit.freedesktop.org/~airlied...9b1cc684f5d702
                  http://cgit.freedesktop.org/~airlied...4b2dc0b84d9106

                  Comment


                  • #19
                    Given that the low-end cards like the 6450 are already GPU-limited on mesa, they should show any improvements in gpu driving efficiency first.

                    Question, is that part already as good as it gets?

                    Comment


                    • #20
                      Originally posted by curaga View Post
                      Given that the low-end cards like the 6450 are already GPU-limited on mesa, they should show any improvements in gpu driving efficiency first.

                      Question, is that part already as good as it gets?
                      For bandwidth limited parts, things like tiling and hiz should improve things somewhat. Tiling uses memory bandwidth more efficiently and hiz reduces the bandwidth required for certain Z operations.

                      Comment

                      Working...
                      X