Announcement

Collapse
No announcement yet.

Qualcomm Aiming For Snapdragon X Elite GPU Support In Linux 6.11

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

  • #61
    Originally posted by drastic View Post
    Geekbench is a crap benchmark.

    Cinebench, on the other hand, is much more serious. But, it is heavily optimized for x86, because that's the market of Cinebench. And, it is mostly a floating-point test, tells you nothing about integer performance.
    Does Cinebench make use of AVX-512? Quick Google says no.
    For integer performance benchmark, I would probably pick a blend of SPECint single-threaded and 7-zip single-threaded. I don't really know of any other good integer benchmarks, they are all crappy (but I don't really track the benchmarking software ecosystem).
    It's just a matter of running multiple tests that are off real world applications. As you can see in the video that while these new Snapdragon chips do really well in 3Dmark, but they do terrible in any real application. It's not an ARM specific test, but it points out the need to run real world benchmarks.

    Comment


    • #62
      Originally posted by Dukenukemx View Post
      Does Cinebench make use of AVX-512? Quick Google says no.

      It's just a matter of running multiple tests that are off real world applications. As you can see in the video that while these new Snapdragon chips do really well in 3Dmark, but they do terrible in any real application. It's not an ARM specific test, but it points out the need to run real world benchmarks.
      I was primarily interested in integer benchmarks.
      For me, integer single-threaded is more important than integer multi-threaded.
      Cinebench is floating point. Perhaps Cinebench is fair on ARM vs x86, perhaps not, but Apple M3 is already beating x86 in Cinebench.
      3Dmark is an GPU benchmark.

      What benchmark is the best for integer performance? I don't know. 7-zip single-threaded is OK, but it measures only a specific case with heavy branching. So, something else must be added to the mix. I suggested SPECint, a standard benchmark.

      Comment


      • #63
        Originally posted by drakonas777 View Post
        When pro-ARM people appeals to SPECint single thread you know that there is no immediate threat to x86 in the practice yet
        Well, the crowd here is certainly impossible to please.
        When I comment on RISC-V performance, then people criticize me that I'm pro-x86.
        (here: https://www.phoronix.com/forums/foru...ptop?p=1471002 )

        When I comment on ARM vs. x86 performance, then I'm pro-ARM.
        It's true that I am cheering for ARM, because I dislike the AMD/Intel duopoly (i.e. the x86 monopoly).
        On the other hand, I try to be fair in my estimations (for my own sake), but I acknowledge that I can make mistakes.

        Comment


        • #64
          Originally posted by drastic View Post
          It's true that I am cheering for ARM, because I dislike the AMD/Intel duopoly (i.e. the x86 monopoly).
          Keep in mind that going ARM does not make this a more competitive market. Like I said, who can make ARM chips for MacOS? Nobody else can make ARM chips for MacOS but Apple. That's not a duopoly but essentially a monopoly. Why you think the US and EU are suing Apple right now? You may think that on Windows anyone can make an ARM chip, but the reality is that only Qualcomm is allowed to make ARM based chips for Windows as per agreement with Microsoft. Out of the many previous attempts that Microsoft has tried to bring Windows to ARM, they've all been through Qualcomm. Who are the major ARM manufacturers in the world? Can't count Apple because they don't sell chips, but there's Qualcomm, Samsung, Rockchip, and Allwinner. To be competitive in the market you can't just use what ARM has, so the good ARM chips are custom made designs, like Apple, Qualcomm, and I think Samsung maybe? It's less than a duopoly on ARM due to what platform it's on. We know Apple doesn't, and so far Apple has donated less than nothing to Asahi Linux. We're assuming that Microsoft won't do to Windows ARM as Android ARM has done with locked down bootloaders. x86 does have a 3rd competitor from VIA, but CPU manufacturing is such a cut throat market that you rarely find a 3rd active competitor. Look at AMD and Nvidia when it comes to GPU's, where we finally got a 3rd competitor from Intel. The only reason Intel is now finally making GPU's is because Nvidia is taking away their server market share. That and you can't make a CPU in 2024 without having at least a decent GPU in it. This is also assuming that Softbank who owns ARM doesn't file for bankruptcy yet again. At some point Softbank will have to sell ARM and Gaben help you if it's Nvidia. You'll see how quickly Nvidia would rip up those permanent license agreements that ARM has with Apple. Probably Samsung too if they have cheap license.
          Last edited by Dukenukemx; 26 June 2024, 06:29 PM.

          Comment


          • #65
            Originally posted by Dukenukemx View Post
            Keep in mind that going ARM does not make this a more competitive market. Like I said, who can make ARM chips for MacOS? Nobody else can make ARM chips for MacOS but Apple. That's not a duopoly but essentially a monopoly.
            Apple MacOS is a monopoly, but only for 10% of the laptop market: the MacOS hostages.
            Intel/AMD x86 is a duopoly-monopoly, but for 90% of the laptop market: the x86 hostages.

            Comment


            • #66
              Originally posted by drastic View Post
              Apple MacOS is a monopoly, but only for 10% of the laptop market: the MacOS hostages.
              Intel/AMD x86 is a duopoly-monopoly, but for 90% of the laptop market: the x86 hostages.
              Because Apple's MacOS has only 10% of the desktop/laptop market, then it's not a monopoly? Technically it's not just AMD and Intel because there's still Via, but as it stands right now we'll be giving Qualcomm the entire Windows/ARM market because Microsoft has made an agreement for this. With AMD and Intel we have... AMD and Intel as competitors. This is not OK, this is not better than the x86 duopoly. What you want is an oligopoly, which is far worse. Why you think ARM on Windows has failed in the past? Microsoft immediately tried to lock down the ecosystem so they would be the sole owners of Windows/ARM. This works for Apple because Apple has customers who'll buy anything they sell, but not the rest of the world. As a Linux user, this will also likely create issues for installing Linux, because it's not like Microsoft won't make changes to prevent alternative OS's from being installed if Linux got popular. With x86 we're on the legacy that is IBM compatibles, so it's much harder for Microsoft to do this. AMD and Intel are pushing code to the Linux kernel so that on day1 their new Zen5 and Lunar Lake products work. Qualcomm is just now getting these out? Apple is like, what's a Linux? Sounds like an inferior MacOS. Like it or not, the future of ARM is through Apple and Qualcomm, and these are not companies you want to be the future of anything.

              Comment


              • #67
                Originally posted by Dukenukemx View Post
                Because Apple's MacOS has only 10% of the desktop/laptop market, then it's not a monopoly? Technically it's not just AMD and Intel because there's still Via,
                but as it stands right now we'll be giving Qualcomm the entire Windows/ARM market because Microsoft has made an agreement for this.
                With AMD and Intel we have... AMD and Intel as competitors. This is not OK, this is not better than the x86 duopoly.
                What you want is an oligopoly, which is far worse.
                Why you think ARM on Windows has failed in the past?
                I'm approaching it from a selfish perspective: the x86-duopoly limits my choices when purchasing a laptop. I would like to have a choice of a laptop CPUs not only from Intel, AMD and Apple, but also from other companies. And, I would like my laptop to have a higher autonomy than the standard 6 hours. And, I would like laptop CPUs to be cheaper.

                So, you can say that I prefer oligopolies to monopolies and duopolies. But I'm also cheering for RISC-V: the more players, the better for me and for the entire world.

                Why has ARM on Windows failed in the past?
                Most important: Lack of backwards compatibility (being rapidly improved this year)
                Less important: Worse performance of ARM CPUs (which were primarily mobile SoCs)
                Less important: lack of choice for consumers (insufficient number of laptop models offered)

                Comment


                • #68
                  Originally posted by Shagga, Son of Dolf View Post
                  As an Openwrt user I already was able to get a taste what Qualcomm thinks about supporting Linux. I will stay away from Qualcomm SoCs on laptops for the time being.
                  You should think again because Qualcomm looks dead serious this time. The effort is fully sponsored by them, either done by their own engineers or contract firms, fully in the open, full mainline support with mainline GPU support and all the bells and whistles too.

                  Comment


                  • #69
                    Originally posted by robclark View Post
                    I genuinely like chipsandcheese, but I think he is underestimating the TSO cost. This implies differently: https://lore.kernel.org/lkml/[email protected]/raw (ok, 9% vs the 10% I said earlier.. I mis-remember)
                    I finally got to reading that post.
                    It is important to understand here what was being measured to arrive at that 9% performance difference.

                    The post is referencing this paper:


                    The paper is a report on some benchmark scores. The benchmark is primarily "SPECspeed 2017 Floating Point" suite. It is a floating point test, and it is a heavily multi-threaded test. Therefore, it is similar to CineBench, and I would expect the results to be similar to CineBench scores.

                    It is likely that "SPECspeed 2017 Floating Point" is over-estimating the importance of TSO vs. weak ordering, at least when talking about common laptop use-cases.

                    In this evaluation, we focus on the SPECspeed 2017 Floating Point suite since
                    the utilization of heavy parallelism results in many hardware threads accessing
                    shared memory concurrently, allowing us to evaluate different memory ordering
                    models properly​...
                    How many people in the world are going to be running such algorithms on their laptops? Laptops are primarily about integer single-threaded performance, not floating-point multithreaded.

                    If we are talking multi-threading on laptops, then it's mostly about having 20 browser windows opened.
                    Not only are browsers primarily reliant on integer performance, but also their threads have LITTLE NEED for SHARED MEMORY AT HIGH BANDWIDTH. Therefore, the would be NO DIFFERENCE between TSO and weak ordering in such scenarios.

                    Also, higher speed does not immediately translate into higher efficiency. For example, if one core gets stalled due to accessing memory concurrectly with others, then that core can be power-gated and clock-gated during the stall, so it doesn't spend much extra energy. Therefore, if the performance gain is 9%, then the energy-efficiency gain would be much smaller than 9%.

                    Comment

                    Working...
                    X