Announcement

Collapse
No announcement yet.

OpenSUSE Tumbleweed Might See Micro-Architecture Packages For Better Performance

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

  • OpenSUSE Tumbleweed Might See Micro-Architecture Packages For Better Performance

    Phoronix: OpenSUSE Tumbleweed Might See Micro-Architecture Packages For Better Performance

    One of the many great programs at SUSE is the roughly annual program where their developers can focus for one week on any new open-source development they desire. SUSE Hack Week has led to many great innovations and improvements since it began in the mid-2000s and for the Hack Week later this month there is one project attempt we are eager to see tackled...

    https://www.phoronix.com/scan.php?pa...-Arch-Packages

  • #2
    Arch is also looking at the feasibility of using LTO and and v2 microarchtitecture level by default.

    https://bbs.archlinux.org/viewtopic.php?id=264185
    https://bbs.archlinux.org/viewtopic.php?id=262571
    https://gitlab.archlinux.org/archlin...rge_requests/4
    https://gitlab.archlinux.org/archlin...rge_requests/2
    https://lists.archlinux.org/pipermai...ate.html#30350

    I think they will eventually be focusing on HWCAPS, like OpenSUSE.

    https://bbs.archlinux.org/viewtopic.php?id=263371

    Comment


    • #3
      openSUSE Tumbleweed in time could potentially be one of the first major Linux distributions offering optimized x86-64 micro-architecture packages to better make use of today's AMD and Intel processors.
      i am not sure about Intel..
      Platinum sponsors of openSUSE

      AMD

      Silver Sponsors of openSUSE

      B1 Systems GmbH
      https://en.opensuse.org/Sponsors

      Comment


      • #4
        Interesting thanks.
        Michael Larabel
        http://www.michaellarabel.com/

        Comment


        • #5
          would be nice for all cutting edge oriented distros to switch to v2 baseline...
          in general i am happy to see from the distro of my choice to offer this optimized paths, offering higher performance out of the box for my scientific loads ...

          Comment


          • #6
            I think this is great, let's see how it goes...
            Alas, Debian will take its sweet time if and when they implement something similar, but that's okay for my workflow.

            Comment


            • #7
              Very cool. Just keep in mind that going beyond level 2 is rather stupid to do globally. All that does is add random lower clocked AVX code all over the place which is usually a net loss.

              Comment


              • #8
                Originally posted by skeevy420 View Post
                Very cool. Just keep in mind that going beyond level 2 is rather stupid to do globally. All that does is add random lower clocked AVX code all over the place which is usually a net loss.
                It depends and I'd like to see some numbers of a v2 versus v3 distribution benchmark comparison. People who care for AVX performance might already set no negative offset at all (which you could consider to be overclocking). But of course not every person does that nor does any chip could run at the higher frequancies within their thermal or power envelope. My 12-Core Xeon 2678 V3 which I unlocked to 3.3 Ghz on all cores downclocks to 3.1 Ghz even with a -80 mV undervolt in Cinebench R23 which uses AVX2 and all cores extensively but games or lighter AVX2 workloads still reach 3.3 Ghz on all cores.

                Comment


                • #9
                  As some people like to keep their museum pieces running until they break, I guess the best compromise would be that distributions start to offer several feature-level-aligned flavors of their whole distribution builds? Openmandriva does just that with their x86-64 and Zenver1 builds (for Ryzen users) and their example proves that it shouldn't be too hard to implement and maintain them. But there might be more optimization opportunities on the table to take advantage of newer hardware features in these v2 - v4 builds. The question would be if these can be realized without costing too much in terms of developer effort from diverging from the standard build?!
                  Last edited by ms178; 08 March 2021, 01:40 PM.

                  Comment


                  • #10
                    Originally posted by ms178 View Post

                    It depends and I'd like to see some numbers of a v2 versus v3 distribution benchmark comparison. People who care for AVX performance might already set no negative offset at all (which you could consider to be overclocking). But of course not every person does that nor does any chip could run at the higher frequancies within their thermal or power envelope. My 12-Core Xeon 2678 V3 which I unlocked to 3.3 Ghz on all cores downclocks to 3.1 Ghz even with a -80 mV undervolt in Cinebench R23 which uses AVX2 and all cores extensively but games or lighter AVX2 workloads still reach 3.3 Ghz on all cores.
                    I wish I could find numbers. I just know that Clear doesn't build with AVX/numbers globally anymore. It's an anecdotal suggestion I've seen in quite a few places over the years, I just don't remember where because I've only had an AVX capable CPU for two weeks so it wasn't really that important to me until now.

                    Comment

                    Working...
                    X