Announcement

Collapse
No announcement yet.

11-Way AMD Radeon GPU Comparison On Linux 3.12, Mesa 9.3

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

  • #16
    kernel 3.11 + dpm by default is not fast for games in my experience, There is needed to force the performance level to the highest level via sysfs to get it working properly

    Comment


    • #17
      Originally posted by Vim_User View Post
      I am not an "average user". I want the same or at least comparable performance with the same settings as the Catalyst driver uses,
      Exact same level as will never happen. Catalyst has like PER CARD optimizations that would NEVER EVER EVER be accepted in an open source project like Mesa (or the kernel) it creates automatic spaghetti code.

      About MSAA you can modify it, its an environment variable.

      Comment


      • #18
        Originally posted by Ericg View Post
        Exact same level as will never happen. Catalyst has like PER CARD optimizations that would NEVER EVER EVER be accepted in an open source project like Mesa (or the kernel) it creates automatic spaghetti code.
        This is why I said "at least comparable". If we get to 90% I am happy about that.

        About MSAA you can modify it, its an environment variable.
        Only partially true: http://phoronix.com/forums/showthrea...817#post353817

        Comment


        • #19
          Originally posted by Ericg View Post
          Exact same level as will never happen. Catalyst has like PER CARD optimizations that would NEVER EVER EVER be accepted in an open source project like Mesa (or the kernel) it creates automatic spaghetti code.
          So, opensource driver never will be as good as the privative one? so, technically is not possible? Thats sad to heard, I was expecting that day.

          But, in my case (ati hd4200 + mesa 9.2 develop..... + dpm thing + r600 shaders optimization) steam games like team fortress 2 and portal seems to run equal (or maybe faster) than in windows using the free driver, seems than the theory is different than the practice in what I have tested.

          Comment


          • #20
            Originally posted by Tyler_K View Post
            not needed as its enabled by default since 3.6
            yes, as stated and shown on the very first page
            Cool, thanks for the info!

            Comment


            • #21
              Originally posted by edoantonioco View Post
              So, opensource driver never will be as good as the privative one? so, technically is not possible? Thats sad to heard, I was expecting that day.

              But, in my case (ati hd4200 + mesa 9.2 develop..... + dpm thing + r600 shaders optimization) steam games like team fortress 2 and portal seems to run equal (or maybe faster) than in windows using the free driver, seems than the theory is different than the practice in what I have tested.
              It'll be good, and maybe even 90% good but I don't forsee Radeon being 100% as good as Catalyst in all cases. Like I said above, the catalyst devs have even said themselvs that Catalyst has tons of hand-tuning done for various cards and generations, things that would make the mesa code nothing but a tangle of IFDEFS and make the code unmaintainable in the open source fashion.

              I think Bridgman originally said the target was 80% of Catalyst, but that was without a few optimizations (like the SB shader backend for example) and some of the other 'more complicated' work that no one was actually expected to do but someone ended up doing it. So we very well may hit 90% or maybe even 98% of the catalyst code. But EXACT same Catalyst performance, across all cards, in all use-cases...? I doubt it. Its just not feasible without starting hand-tuning everything..

              Comment


              • #22
                Wow, the 6570 seems to be unexpectedly slow. The 5670 isn't tested but I guess I might be still happy with my choice back at the time to take the 5670 instead a 6xxx series since support was better in the free drivers at that time.
                Though I thought the 6570 falls in the mainstream category and iirc. some of the devs said these would be easiest to handle. The small GPUs need lots of attention and care to work with a fair performance, the mainstream is nice but the high ends need again attention since the CPU and others might already be a bottleneck so you can't use their full portential.
                But besides the little cards and the exception all of them show very playable framerates and that is a good thing.

                Comment


                • #23
                  Originally posted by tulioserpio View Post
                  So the real question is:

                  Has the HD 6670 the same problems the HD 6570 has or not?

                  Thanks for the answer.

                  No, it hasn't. I posted my 6670 results while ago.

                  Comment


                  • #24
                    Originally posted by Ericg View Post
                    I think Bridgman originally said the target was 80% of Catalyst, but that was without a few optimizations (like the SB shader backend for example) and some of the other 'more complicated' work that no one was actually expected to do but someone ended up doing it. So we very well may hit 90% or maybe even 98% of the catalyst code. But EXACT same Catalyst performance, across all cards, in all use-cases...? I doubt it. Its just not feasible without starting hand-tuning everything..
                    The original estimate I mentioned was 60-70%, but that wasn't a *target* -- just an estimate from our architects of the performance we could expect from a "vanilla" implementation with a simple shader compiler/translator. It was more a reflection of the amount of developer effort expected to be available. As Ericg said, the SB backend and the use of LLVM have already invalidated some of those assumptions so performance limit is really a matter of developer time and keeping the code clean.

                    Unfortunately optimization is one of those "exponentially expensive" things, where each incremental improvement tends to be both smaller and more time-consuming than the last one.

                    The primary "advantage" Catalyst has is that significant chunks of the code are shared across most of the PC market (ie not just Linux) and therefore can justify more developer time, with the caveat that the resulting code usually needs to be kept closed because of requirements from the other OSes it supports.

                    Comment


                    • #25
                      @Vim_user

                      If the game offers the option to disable MSAA in its own menus, that option will work. It's only driver-level forcing that is not complete.

                      Comment


                      • #26
                        Originally posted by bridgman View Post
                        The primary "advantage" Catalyst has is that significant chunks of the code are shared across most of the PC market (ie not just Linux) and therefore can justify more developer time, with the caveat that the resulting code usually needs to be kept closed because of requirements from the other OSes it supports.
                        Wasn't the reason gallium was so cool that it was very portable?
                        "Gallium drivers have no OS-specific code (OS-specific code goes into the "winsys/screen" modules) so they're portable to Linux, Windows and other operating systems."
                        "Gallium3D is designed to support multiple API's, multiple GPU's, and Multiple OS and windowing systems."
                        http://www.freedesktop.org/wiki/Software/gallium/

                        And here are sometimes threads about gallium3d being used in proprietary drivers, most recently http://phoronix.com/forums/showthrea...ietary-drivers

                        So since mesa is MIT licensed, significant chunks of the open source driver could theoretically be used on most of the PC market (ie not just linux) too, right?

                        Comment


                        • #27
                          Originally posted by ChrisXY View Post
                          So since mesa is MIT licensed, significant chunks of the open source driver could theoretically be used on most of the PC market (ie not just linux) too, right?
                          Yes, if we wanted to "start over" and discard most of a $20-50M investment (I don't know the exact numbers) in the current driver... but where is the justification for doing that ? Going with gallium3d, for example, would just mean replacing our existing hardware layer (hwl) with a different one a bit further down the stack. Having the hwl further down the stack does make for a smaller and more maintainable hwl driver and allows relatively more of the upper level code to be common, but also tends to make performance tuning more difficult once you get out to the edges.

                          If we were writing new drivers from scratch (which AFAIK was the scenario VMware faced when writing gallium3d-based drivers for their emulated SVGA hardware) *and* were looking for a good compromise between maintainability and performance then using gallium3d would make a lot of sense. In the case of both AMD and NVidia, however, where workstation and gaming markets both require "every last scrap of performance" to win sales, going with something more specialized is usually the only option.

                          The key point though is that it's not "lack of use of open source code" that keeps the closed driver closed, it's the need to design it around non-Linux operating systems. We could make a closed-source driver using some open source code but it would still have to be closed source.
                          Last edited by bridgman; 09-24-2013, 08:50 AM.

                          Comment


                          • #28
                            Originally posted by Vim_User View Post
                            I am not an "average user". I want the same or at least comparable performance with the same settings as the Catalyst driver uses, I don't buy expensive hardware to get sub-par performance. If in a game in 1920x1080 with all settings to maximum the Catalyst driver has room to enable visual enhancements like MSAA and AF then the radeon driver should give a similar performance with the same visual quality.
                            My point was that in all of the Phoronix benchmarks 60fps@1920x1080 is attainable with the open source drivers. Arguing that 250fps isn't acceptable because the hardware can theoretically do 280fps is really irrelevant when your monitor can't display that and your game is vsyned to 60fps (yes, I know the profesional gamers prefer no vsync for lowest latency, but 99% of people don't because it results in frame tearing) . If you need more than 60fps@1920x1080, then either you game at higher resolution than 1920x1080 (probably multi-screen, perhaps SLI), or you have a 120hz monitor, or the Phoronix benchmarks are not challenging enough to represent your game+graphics settings (in which case you need to run your own benchmarks).

                            If open source drivers get 85% of the fps of closed drivers, it only impacts your gaming if your card would otherwise be running at 60-70fps with closed drivers, because that means the open source driver is going to fall below 60fps whereas closed drivers wouldn't. But if you would get less than 60fps with the closed driver, then you need a better card anyway. At 80%+ performance, there's a very small number of game/card/resolution configurations where you're going to exceed the 60fps threshold with a closed driver, but not with the open one.

                            So you can't just consider average fps - the real metric is whether avg or min fps exceeds the hardware limitation of the monitor for the particular game+settings that you are using. The monitor will never display more than a fixed fps (usually 60). Excess frames are never even transferred to the monitor. You can't see them.

                            (Incidentally, I'm using 60fps as the upper limit because that is the max refresh rate of the majority of monitors. In reality, 30fps is acceptable for most people - most PS3 and Xbox 360 games are 30fps, including the $1billion GTA5.)

                            Comment


                            • #29
                              recentl byed XFX 6570.. it was crap..

                              subj.. decided to replace my old 4xxx series amd gpu to newer.. Here you go: ubuntu didn't support this card from box, proprietary drivers wasn't ready either.. testing icon, all sorts of overheating/artifacts despite huge radiator + 3xadditional system coolers. Switched back to old 4xxx card..
                              Some black voodoo magic in area of AMD/nvidia products + recent linux kernels in field. I hope i can fix all issues with system, still i can't say i would recommend amd products + linux as stable system currently to anyone for now. Having all kinds of issues with amd hw ranged from newest amd fm2 gpu-cpu hw to older systems on modern linux software(actually fm2 have hw issues on ms systems too).

                              Comment


                              • #30
                                Originally posted by Asariati View Post
                                That is such a big bullshit, it's not even funny. You don't hire a bunch of people and they are immediately productive. You don't hire software engineers and they are immediately capable of writing solid, high performance drivers. You don't hire software engineers and they know every nook and cranny about a until-then-unknown hardware.
                                If you have a recipe for training people as fast as you claim, please share it or open a development company and earn big money. Then you can fund all those open source projects.


                                Why should any company give up a competitive advantage?


                                Business are to earn money for the shareholders. Nothing else.


                                So for your or any modern car you are getting a complete repair manual with cheap tools?
                                Don't think so or let me know model and brand. Though this will make a lot of mechanics very unhappy.


                                AMD is putting in a lot of money (seen from their revenue perspective) and time and you're buying Intel. Yeah, that is a hell of a support for AMD ...
                                Well the real question is better for them to cut costs and share risks or to get competitive advantage Openstack participants think otherwise.
                                In defence industry yes in industry where competitive advantage should matter much more risk and cost sharing is really common (Russia is cooperating with France and Italy)

                                Business are to earn money by any means necessary for the shareholders. Nothing else.

                                Well the problem is that closed source drivers in Linux just suck. Try to install them in Suse Tumbleweed and youll get the picture. Not to mention other problems

                                Comment

                                Working...
                                X