Announcement

Collapse
No announcement yet.

AMD Radeon RX 5500 XT Linux Performance

Collapse
This topic is closed.
X
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • #61
    Originally posted by Danny3 View Post

    Who says it's a 1920x1080 card ?
    Who says you can't hook it up to 4K monitor or TV and watch a movie on it ?
    I honestly don't buy GPUs just for games.
    I want to be able to watch movies with them with the latest quality features turned on like HDR (High dynamic range) and HFR (High frame rate).
    Also I want in the future to buy a [email protected] or more when they will become available.
    Why should I need a new GPU just for this just because they chose the cheaper path of putting a very old version of the HDMI instead of a somwhat newer one?
    HDR and HFR is in no way "quality features", all the movies shot above the proper 24fps (and note here that this card will support 4K60Hz) looks like ugly tv soap operas. However if that really is your fancy then yes you will have to look for a different card, even if they had added 2.1 support I have a hard time believing that the GPU on it would be able to push 48Gbps, let along decode any codec on that bandwidth.

    For anyone interested in why 24fps is better for movies, please see this Youtube by Filmmaker IQ:
    Last edited by F.Ultra; 14 December 2019, 03:45 PM.

    Comment


    • #62
      Originally posted by atomsymbol View Post
      I would like to see an expression that is more accurate. So please post it here.
      I'm not really sure what you were trying to show with your calculation, but I don't think it made any sense.

      7nm / 12nm = 58%.
      251 mm2 / 445 mm2 = 56%.
      (251/10.3)/(445/10.8) = 59%.

      It all seems to line up pretty closely to me.

      That said, you are correct that different features take different sizes of silicon and doesn't necessarily scale directly with the # of transistors. I just don't think your calculations had any relation to that fact, if that's what you were trying to show.
      Last edited by smitty3268; 14 December 2019, 03:54 PM.

      Comment


      • #63
        IMO, it's good that they managed a 128-bit card that can hang with their previous-generation 256-bit cards, but the value is not there. RX 570 performs nearly as well, and if you look at the current sale prices, in particular, its performance per $ is way better. Plus, Polaris has no caveats around support for GPU compute.

        Over time, RX 5500 cards should come down in price, hopefully to the RX 560's bracket. However, at least until RX 570 stocks run dry, the 5500's will be challenging to recommend.
        Last edited by coder; 14 December 2019, 04:16 PM.

        Comment


        • #64
          Originally posted by F.Ultra View Post
          HDR and HFR is in no way "quality features", all the movies shot above the proper 24fps (and note here that this card will support 4K60Hz) looks like ugly tv soap operas.
          It basically boils down to the same arguments of film vs. digital. Old people, stuck in their ways, and used to having a certain look. Also, I think actors don't like how HFR exposes their weaker performances.

          I saw Gemini Man in 60 FPS HFR 3D and it looked amazing. Like looking through a window. Most action scenes in movies are a blurry mess, but the motorcycle chase scene was totally clear. Yes, I'll have some more of that, please.

          And yes, I also use my TV's motion interpolation, when watching 24 fps content. Once you get over the initial adjustment period, a fair-minded person cannot deny that it looks better (assuming a competent implementation with minimal artifacts).

          Originally posted by F.Ultra View Post
          even if they had added 2.1 support I have a hard time believing that the GPU on it would be able to push 48Gbps, let along decode any codec on that bandwidth.
          You only need 25.8 Gbps to reach 4k @ 120 Hz (8-bit), though I think you might be right that the link probably has to run at 4x12 Gbps to deliver that. Still, if you're talking about the memory bandwidth required by the RAMDAC (or whatever you call its modern equivalent), it's peanuts - just 3.2 GB/sec. The card is rated at 224 GB/sec.

          So, the only question is whether the decoder can manage. According to this, the decode block used in Navi 10 (the 5700 cards) is only capable of 4k @ 120 Hz for H.264. At H.265, it's limited to just 4k @ 60 Hz.

          https://www.techpowerup.com/256481/a...ngine-detailed

          Comment


          • #65
            Originally posted by smitty3268 View Post

            I'm not really sure what you were trying to show with your calculation, but I don't think it made any sense.

            7nm / 12nm = 58%.
            251 mm2 / 445 mm2 = 56%.
            (251/10.3)/(445/10.8) = 59%.

            It all seems to line up pretty closely to me.
            You are comparing a distance ratio (7 nm / 12 nm) to an area ratio (251 mm2 / 445 mm2). This doesn't make sense. Neither did atomsymbol's calculation:

            Look at it this way: it's a comparison of the total size of idealised transistors on the dies to the actual sizes of the dies, the idealised size of a transistor being (7 nm)² and (12 nm)², respectively. atomsymbol took the root of each side to get a linear metric to compare to the process size (a 1-dimensional size). But he didn't take process size and number of transistors on the chips into account in his calculation which makes it rather nonsensical.

            One could calculate a linear metric like this: (10.8e9 × 12 nm × sqrt(251))/(10.3e9 × 7 nm × sqrt(445)) ≈ 1.35. I interpret this as showing that the transistors take up a larger fraction of the total die area in the 12 nm chip. Multiply by 7 nm to see that if it was possible to just scale number of transistors per die area with (inverse) process size, the RTX 2060's die area and transistor count would correspond to a 9.45 nm process.

            Comment


            • #66
              Originally posted by smitty3268 View Post

              I'm not really sure what you were trying to show with your calculation, but I don't think it made any sense.

              7nm / 12nm = 58%.
              251 mm2 / 445 mm2 = 56%.
              (251/10.3)/(445/10.8) = 59%.
              58% / 59% = 0.98

              It all seems to line up pretty closely to me.

              That said, you are correct that different features take different sizes of silicon. I just don't think your calculations had any relation to that fact.
              f(x)=x and g(x)=x² have different shapes. A transformation needs to be performed in order to be able to compare f and g. In this case the correct transformation is sqrt(g(x)).

              For RX 570, 14nm, 5.7e9 transistors, 232 mm², your logic yields:

              7nm / 14nm = 50%
              (5.7/232)/(10.3/251) = 60%
              50% / 60% = 0.83

              For R9 390, 28nm, 6.2e9 transistors, 438 mm², your logic yields:

              7nm / 28nm = 25%
              (6.2/438)/(10.3/251) = 34%
              25% / 34% = 0.74

              For Radeon HD 6790, 40nm, 2.64e9 transistors, 389 mm², your logic yields:

              7nm / 40nm = 17.5%
              (2.64/389)/(10.3/251) = 16.5%
              17.5% / 16.5% = 1.06

              For Radeon HD 4890, 55nm, 0.959 transistors, 282 mm², your logic yields:

              7nm / 55nm = 12.7%
              (0.959/282)/(10.3/251) = 8.3%
              12.7% / 8.3% = 1.53

              If the logic you used was a valid method for relating transistor density (mm²) and process technology (nm¹), then instead of the ratios 0.83, 0.74, 1.06 and 1.53 we would we get a value close to 1.0 in all 5 cases instead of just (accidentally) in 1 or 2 cases.

              The other option of course is that the transistor numbers and the die sizes have been measured/reported incorrectly, but in my opinion this option has a lower probability to be true.

              Comment


              • #67
                Originally posted by bitnick View Post
                atomsymbol took the root of each side to get a linear metric to compare to the process size (a 1-dimensional size). But he didn't take process size and number of transistors on the chips into account in his calculation which makes it rather nonsensical.
                The number of transistors on the two chips (RX 5700, RTX 2060) was taken into account. Approximately their ratio is 1:1.
                Last edited by atomsymbol; 14 December 2019, 04:47 PM. Reason: Fix grammar typo

                Comment


                • #68
                  Originally posted by bitnick View Post
                  You are comparing a distance ratio (7 nm / 12 nm) to an area ratio (251 mm2 / 445 mm2).
                  I actually don't think 7nm and 12nm are distance ratios, FWIW.

                  I'm admittedly not an expert in this area, but my understanding is that these are basically marketing terms that are used to describe the process, not actual measurements. And that it's based on the average transistor density the manufacturer thinks their process can provide - presumably it's based on actual numbers somewhere back down the line on older processes. But that's why Intel's 10nm process is described as similar to TSMC's 7nm - because they don't use actual distances, so they aren't directly comparable to one another.

                  Comment


                  • #69
                    Originally posted by atomsymbol View Post
                    The number of transistors on the two chips (RX 5700, RTX 2060) was taken into account. Approximately their ratio is 1:1.
                    The RTX 2060's TU106 has 10.8 B transistors, while the RX 5700's Navi 10 has 10.3. So, the Nvidia GPU has about 5% more. Considering it also has ray tracing and tensor cores, that's not bad. However, Navi 10 has about 28% more raw fp32 performance, if we compare the stock RX 5700 XT to the RTX 2060 Super, at base clocks.

                    Of course, even the RTX 2060 Super doesn't fully-enable the TU106. It's still down on "cores", by about 6% vs the base RTX 2070, which also uses that GPU. That should approximately cancel out the transistor-count differences vs. Navi 10, perhaps giving some idea of the fp32 performance Nvidia is leaving out, in favor of tensor cores and ray tracing.
                    Last edited by coder; 14 December 2019, 05:39 PM.

                    Comment


                    • #70
                      Originally posted by coder View Post
                      And yes, I also use my TV's motion interpolation, when watching 24 fps content. Once you get over the initial adjustment period, a fair-minded person cannot deny that it looks better (assuming a competent implementation with minimal artifacts).
                      You know the ultimate solution to the "24 FPS is too low for smooth motion" problem?

                      "mpv" with these interpolation settings:
                      tscale=box
                      tscale-window=quadric
                      tscale-clamp=0.0
                      tscale-radius=1.1
                      The result?

                      No artifacts while having absolutely smooth motion, all while completely avoiding the 'soap-opera' effect!

                      Try it out; else your missing out!

                      Comment

                      Working...
                      X