Announcement

Collapse
No announcement yet.

Mesa 9.2 & The R600 SB Back-End Are Good For AMD APUs

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

  • #11
    Originally posted by Hamish Wilson View Post
    Considering everything that has happened, I think in terms of drivers both AMD and Intel can now (or at least soon) be considered to be in similar states, now that we have dynamic power management and video acceleration for radeon chips. I wonder if this will make anyone rethink their next APU purchase, or least make their choice a lot more difficult next time.
    I've had problems with drivers every time I've ever purchased ATI/AMD graphics cards, which is the main reason that I've chosen NVIDIA. I'd really like to support open-ness, but AMD binary drivers bothers me, and the open source ones, not quite on par with the binary drivers (until now?), Intel HD-graphics doesn't quite suffice, and noveau, while I believe and hope they'll gain on binary drivers, it seems a long way away.

    Comment


    • #12
      Originally posted by ObiWan View Post
      Given that the article was using an APU at 4.7Ghz, and yours was at 4.1Ghz, that's pretty impressive. The other main difference that I saw was in memory allocated to the GPU (768MB in article, 2048 on yours).

      I don't see the system logs attached, so I can't say whether your test results were with S3TC/FP Textures, but even then... mighty impressive. I'm looking forward to direct 3.9/3.10/3.11 comparisons in the coming weeks.

      Comment


      • #13
        Originally posted by ObiWan View Post
        Wow... nearly four times faster with dpm!

        Comment


        • #14
          Originally posted by Veerappan View Post
          Given that the article was using an APU at 4.7Ghz, and yours was at 4.1Ghz, that's pretty impressive. The other main difference that I saw was in memory allocated to the GPU (768MB in article, 2048 on yours).
          It's within expectations, if you look at the GPU clocks. On APUs, without DPM, the driver will only use a low default clock that's about 40-45% of the maximum clock. It gets worse if the GPU supports boost states. E.g. on E-450, you'll only get 200 MHz clock without DPM. With DPM however, the GPU can clock to 500 MHz and use the boost state of 600 MHz most of the time. So if you are not memory bandwidth and/or cpu bound, you might see 3x speedups!

          Comment


          • #15
            Originally posted by UraniumDeer View Post
            *snip*.
            Well, your post really had nothing to do with what I said, as I was comparing AMD to Intel, but anyway...

            Comment


            • #16
              A couple of those results look odd.

              The Nexuiz 2.5.2 results are basically identical (to 4 significant figures!) between the 1920x1080 and the 2560x1600 resolutions. Even if that game were CPU bound (which at those frame rates it isn't) you'd expect way more variance run to run.

              I'd say it was a copy/paste error, except a) that the numbers differ at the 4th significant digit, and b) The Reaction Quake numbers show the same behaviour.

              Any suggestions as to what's going on?

              Comment


              • #17
                One issue with these benchmarks is that newer versions of Mesa default to the R600 LLVM compiler, and just setting R600_DEBUG=sb doesn't disable LLVM. R600_DEBUG=sb,nollvm yields better performance based on my tests so R600sb performance should be even better (although this may vary application to application)

                Comment


                • #18
                  Originally posted by Veerappan View Post
                  Given that the article was using an APU at 4.7Ghz, and yours was at 4.1Ghz, that's pretty impressive. The other main difference that I saw was in memory allocated to the GPU (768MB in article, 2048 on yours).

                  I don't see the system logs attached, so I can't say whether your test results were with S3TC/FP Textures, but even then... mighty impressive. I'm looking forward to direct 3.9/3.10/3.11 comparisons in the coming weeks.
                  it's with S3TC

                  Comment

                  Working...
                  X