Announcement

Collapse
No announcement yet.

Radeon Memory Boost: Up To 50%+ Better Performance

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

  • #31
    Originally posted by agd5f View Post
    This affects all radeons so it could help on older cards as well. It will also help on APUs and IGPs. Even though APUs don't have dedicated vram, they use a stolen chunk of system memory as a vram pool. This is faster than regular system memory mapped through the gart since it's contiguous an non-snooped.
    What is the situation with Kaveri? How is the dedicated VRAM different from system ram in its case since it has hUMA?

    Comment


    • #32
      In order to get better performance from my patches, you need an app which uses more memory than your card has. You'll mainly notice it on old and low-end chips. If you have enough video memory, you won't see a difference.

      Comment


      • #33
        The only way a GPU can save you money

        Originally posted by GreatEmerald View Post
        I'm aware of that, I'm waiting for stabilisation. As mentioned, it works fine with the blob now, so no reason to jump to unstable land just for that.



        Well, that depends. Newer hardware is often more power efficient, so upgrading can both boost performance and cut down spending (plus you can then sell your old hardware). Or, alternatively, make use of your old hardware in some fashion (like assemble an HTPC ? I did exactly that).
        Is if you have an older high-end GPU, the big ones from 2007-2009 seem to have been the worst power hogs. For smaller cards on machines not run 24-7, comparing a 12-35W power saving card to an older card that may max out at 25W full time I doubt you'd ever get your money back. A lot of energy goes into fabbing and shipping hardware, don't forget that part of the equation. You'd have to be replacing a Pentium 4 Prescott with a nettop to pay for it in electricity savings, you'd do better by shitcanning two or three incandescent light bulbs for LED's.

        Comment


        • #34
          Originally posted by marek View Post
          In order to get better performance from my patches, you need an app which uses more memory than your card has. You'll mainly notice it on old and low-end chips. If you have enough video memory, you won't see a difference.

          Hey Marek,

          What are you cooking for next? What is still missing from the Radeon driver? Any plan for the next couple months?

          Thanks for your work.

          Kind regards,

          Lajos

          Comment


          • #35
            Next job: fixing severe performance regression with new Xserver

            Originally posted by LN17 View Post
            Hey Marek,

            What are you cooking for next? What is still missing from the Radeon driver? Any plan for the next couple months?

            Thanks for your work.

            Kind regards,

            Lajos
            I believe the new version of Xorg is 1.15, I see DRI3 packages in the list of new packages for it. I got a very bad performance regression with it, about a 50% reduction in framerate across the board with the Radeon 6750, a bit less with the Radeon 5570. Worst results were in scorched3d with the Radeon 6750, dropping all the way to 10fps on some screens as the heaviest screens used to be good for about 25fps with 50-60fps typical. With Xorg 1.15 and Mesa 10.2 the game was sometimes unplayable. With X reverted to 1.14 I got much better results but also had to roll back to Mesa 10.1 to get back all the previous performance.

            Comment


            • #36
              Originally posted by LN17 View Post
              What are you cooking for next? What is still missing from the Radeon driver? Any plan for the next couple months?
              You'll see it when it's ready.

              Comment


              • #37
                Originally posted by Ansla View Post
                What is the situation with Kaveri? How is the dedicated VRAM different from system ram in its case since it has hUMA?
                It should still matter in Kaveri, as the GPU has a different link to the memory, one that doesn't have cache coherency.

                Comment


                • #38
                  Still training the AI. It can currently match the LRU, but that isn't really enough. I do have results on the two-ended allocation threshold, which reduced eviction nicely, kernel patch maybe later this week or the next.

                  Comment


                  • #39
                    I am very interested in this patchset. Again thanks Marek!

                    I have an old 4670 521mb on which I play some games, how can I check how much vram they use?
                    I looked into radeontop but it does not seem to provide this information.

                    Comment


                    • #40
                      OK I found gallium_hud, thanks!

                      Comment

                      Working...
                      X