Announcement

Collapse
No announcement yet.

AMD Radeon HD 8670D: Gallium3D vs. Catalyst

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

  • #51
    Originally posted by Azrael5 View Post
    If you state that no improvements are possible because I'm only a "troll" you state that linux has limits which cannot exceed. Do you think linux cannot improve and mesa reach the same target. I think it is possible.
    I never said that you were a troll or that Linux and Mesa could not improve; they have been improving by leaps and bounds over the past few years. But I do not think you truly understand why the driver model we have works the way it does, or comprehend the benefits that are derived from it. I certainly would not want it to change, to have to go back to what happens with XP where you need to go and hunt for an assortment of hard to find drivers to get any of your hardware working. Granted, I have heard this has become a lot more automated in later versions of Windows, but it is XP we are discussing here.

    Comment


    • #52
      Originally posted by orome View Post
      having a non-root access would classify as information leak security bug (at least dmidecode requires access to /dev/mem).
      Ideally would be nice if the kernel provided the memory details over sysfs/debugfs for others too, not asking for /dev/mem access from unprivileged user-space.
      Michael Larabel
      https://www.michaellarabel.com/

      Comment


      • #53
        Originally posted by Azrael5 View Post
        Benchmark shows to all what are the real capabilities of hardware and drivers as well.

        If you state that no improvements are possible because I'm only a "troll" you state that linux has limits which cannot exceed.
        No, you only need to benchmark properly, and the results improve significantly. The hardware and the drivers already run much faster than this benchmark shows.

        It's you who doesn't understand drivers, or what was being benchmarked here. I suggest you read a bit and test it yourself before posting again.

        Comment


        • #54
          Originally posted by Kivada View Post
          I've seen the 3D mark results of the 8670D being OCd to 1.2Ghz on DDR3 2.4Ghz, he said that at that speed the performance hit a wall due to lack of ram bandwidth, but did not say if he tried or even could push the ram speed higher for whatever reason.

          Said results show that at those enough to keep pace with the Iris Pro 5200, with more ram speed it may even overtake it. Though many reviews are showing that the HD4600-Iris pro GPUs seem to have very high frame latency compared to the AMD APUs causing a noticeable stutter in gameplay.

          Link
          http://forums.anandtech.com/showthread.php?t=2323903
          The results showed it to be still lagging the slower of the 5200s. More, some commenters in the thread seemed to think that there is a problem with the memory controller and that simply oc'ing the ram more won't help (maybe a 5% gain was the speculation).
          Really amazing job by intel.

          Comment


          • #55
            Originally posted by liam View Post
            The results showed it to be still lagging the slower of the 5200s. More, some commenters in the thread seemed to think that there is a problem with the memory controller and that simply oc'ing the ram more won't help (maybe a 5% gain was the speculation).
            Really amazing job by intel.
            Not quite, the closest thing to the 8670D in a dedicated GPU is the HD5670 which has 64Gb/s of bandwidth, the 8670D with DDR3 2400 isn't even at 3/4ths of that bandwidth. Since the 8670D is using a more efficient GPU design the only thing holding it back is it's memory speed. Haven't looked for the tests yet but HWBot says that the fastest the 8670D has been pushed on AIR is 1204Mhz/2980Mhz, though ram that can handle those timings is going to be either quite rare or quite expensive.

            The AMD parts still have better frame latency as well as are actually available to the public at a cheap price. the Iris pro is currently only found in a single chip that can only be used soldered onto the mobo and costs as much as the entire A10-6800K/DDR3 2133 system just for the chip even when sold in lots of 1000.

            For the money that Intel wants for chips that have a semi competitive GPU you can but a system with a quad core FX and a much faster dedicated GPU.

            If you are burning money on Intel hardware for their GPU tech you are paying far more then you should for what you are getting.

            Comment


            • #56
              Originally posted by liam View Post
              The results showed it to be still lagging the slower of the 5200s. More, some commenters in the thread seemed to think that there is a problem with the memory controller and that simply oc'ing the ram more won't help (maybe a 5% gain was the speculation).
              Really amazing job by intel.
              The Iris Pro has additional embedded RAM to gain memory bandwidth. It is really all about bandwidth in this game. Embedded RAM is expensive, though. AMD wants to use GDDR5 instead, which is a good choice, too.

              Comment


              • #57
                Originally posted by Kivada View Post
                Strawman much? They also don't undermine the OS they are running by intentionally using outdated versions of the drivers or OS unless they are being paid to quite laughably try to show a bias toward competing hardware in a comparison review.

                They also usually try to find out why they have large discrepancies in their numbers on the same or similar hardware with different software or drivers.

                They go into detail about the hardware and settings they use, they also often try multiple tweaked hardware settings, they provide heat and noise tests, they post about custom tweaks that can help with heat or noise.

                Furthermore they also review with a much broader selection of software, they don't build their own super limited automated review software so they can pull a Ron Popeil and "Set it and forget it" style of review.

                So yeah, I don't like how Larabel does what he does because he does it so badly he makes Linux look bad and Linux users look stupid.
                I thought you were being an asshole with some of your previous posts (and maybe you were), but if the above is what you were trying to communicate earlier then I'm forced to change my mind and agree with you.
                I think that if Michael spent more time for both setting up and analyzing the reviews he would get more interested readers (while I appreciate the news I think much of it could be provided by readers in a separate news section letting him focus on reviews).
                I don't agree with your last sentence, however, b/c I don't think he does either of those things. He has a legitimate point in testing OOTB configurations, but I agree that he should also include tweaked settings as a matter of course. What would be ideal, IMHO, would be a system a PTS that tweaked sysfs values and ran each of them to find what provides the ideal combination for a given test. Yeah, that would take awhile, but I'm not aware of anyone having ever done something like that and think it would be valuable info that could lead to distros changing their default configurations.

                Originally posted by Kivada View Post
                Not quite, the closest thing to the 8670D in a dedicated GPU is the HD5670 which has 64Gb/s of bandwidth, the 8670D with DDR3 2400 isn't even at 3/4ths of that bandwidth. Since the 8670D is using a more efficient GPU design the only thing holding it back is it's memory speed. Haven't looked for the tests yet but HWBot says that the fastest the 8670D has been pushed on AIR is 1204Mhz/2980Mhz, though ram that can handle those timings is going to be either quite rare or quite expensive.

                The AMD parts still have better frame latency as well as are actually available to the public at a cheap price. the Iris pro is currently only found in a single chip that can only be used soldered onto the mobo and costs as much as the entire A10-6800K/DDR3 2133 system just for the chip even when sold in lots of 1000.

                For the money that Intel wants for chips that have a semi competitive GPU you can but a system with a quad core FX and a much faster dedicated GPU.

                If you are burning money on Intel hardware for their GPU tech you are paying far more then you should for what you are getting.
                I was only responding to the link you provided in the AT forums (IIRC), and I stand by it (though I may be misunderstanding your point).
                Also, the intel gpu isn't discrete so I think it completely fair to compare it to any of the apus, but that the Intel system performs like a mid-range discrete system I think that somewhat justifies the cost (though I'm sure they could easily cut their margins but the market doesn't work that way so they don't have to). It's OpenCL performance, OTOH, puts it above many of the mid-range chips, IIRC. So, I don't know of another combination of cpu and gpu that provides at least as much performance at the given energy budget at the same cost (though I admit I haven't looked very hard).

                Originally posted by brent View Post
                The Iris Pro has additional embedded RAM to gain memory bandwidth. It is really all about bandwidth in this game. Embedded RAM is expensive, though. AMD wants to use GDDR5 instead, which is a good choice, too.
                I realise that, and certainly either choice can be justifed given tradeoffs between latency and bandwidth, but from what I've read intel seems to have made the correct one given their gpu constraints. IOW, they don't seem bandwidth limited by that 128MB of embedded ddr3.
                Last edited by liam; 07 July 2013, 05:33 PM.

                Comment


                • #58
                  Originally posted by liam View Post
                  The results showed it to be still lagging the slower of the 5200s. More, some commenters in the thread seemed to think that there is a problem with the memory controller and that simply oc'ing the ram more won't help (maybe a 5% gain was the speculation).
                  Really amazing job by intel.
                  I have come to an understanding that Intels APU uses an integrated memory controller that is bottlenecked by the integrated northbridge. Theoretically increasing memory frequency should yeild improvements, but in practice it doesnt because of the bottleneck. AMD's APU has the same bottleneck but to a lesser extent, and it can be alleviated some more by increasing the integrated northbridge clock.

                  Comment


                  • #59
                    Originally posted by liam View Post
                    I think that if Michael spent more time for both setting up and analyzing the reviews he would get more interested readers
                    Unfortunately that's not feisible when I am doing everything by myself and need to keep the constant flow of content in order to sustain ad revenues.

                    Originally posted by liam View Post
                    What would be ideal, IMHO, would be a system a PTS that tweaked sysfs values and ran each of them to find what provides the ideal combination for a given test. Yeah, that would take awhile, but I'm not aware of anyone having ever done something like that and think it would be valuable info that could lead to distros changing their default configurations
                    PTS already supports automatically testing any abstract system with a list of tunables to find ideal testing configurations from a list of sysfs values to kernel config options.
                    Last edited by Michael; 07 July 2013, 05:41 PM.
                    Michael Larabel
                    https://www.michaellarabel.com/

                    Comment


                    • #60
                      Originally posted by liam View Post
                      I thought you were being an asshole with some of your previous posts (and maybe you were), but if the above is what you were trying to communicate earlier then I'm forced to change my mind and agree with you.
                      Well I'm always an asshole and it's a point I like to make quite often.



                      Originally posted by liam View Post
                      I was only responding to the link you provided in the AT forums (IIRC), and I stand by it (though I may be misunderstanding your point).
                      The point I was making there is that the AMD APUs are bandwidth limited. I used the HD5670 as the closest analogue to HD8670D in terms of GPU core performance, comparing it's availible bandwidth to that of the HD8670D shows that while the GPU core of the HD8670D should in theory be capable of spitting out frames far faster then an HD5670 it's held back by being starved for bandwidth even with quite fast ram.

                      Intel tries to squeeze by with a very expensive eDram chip on the GPU, allowing very low latency. If AMD had done the same it's fairly obvious that the 5200 would not be able to keep up based souly on the GPU core tech.

                      I stand by my statement that for the cost of the 4950HQ/Iris Pro 5200 is going to be $657 for just that chip and nothing else. At least in the US you can build a nice non iGPU system with a GPU that blows the doors off of the 5200 for that much cash.

                      Comment

                      Working...
                      X