Announcement

Collapse
No announcement yet.

Next-Gen AMD EPYC Changes To EDAC Driver Sent In For Linux 5.2 Kernel

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

  • Next-Gen AMD EPYC Changes To EDAC Driver Sent In For Linux 5.2 Kernel

    Phoronix: Next-Gen AMD EPYC Changes To EDAC Driver Sent In For Linux 5.2 Kernel

    The notable change with the "EDAC" changes for Linux 5.2 comes down to the "Zen 2" support for the new AMD EPYC processors launching later this year...

    http://www.phoronix.com/scan.php?pag...en-2-EPYC-EDAC

  • #2
    Unified Memory Controllers ... looks like Rome will indeed be a UMA design, not complex NUMA. Fun times ahead.

    Comment


    • #3
      Off Topic:

      Lenovo is finally offering Ryzen on their top of the line T-series Thinkpad: https://www.notebookcheck.net/Lenovo....418208.0.html

      Comment


      • #4
        Originally posted by pegasus View Post
        Unified Memory Controllers ... looks like Rome will indeed be a UMA design, not complex NUMA. Fun times ahead.
        The statement really says nothing about the system memory topology.
        Only that they have increased the possible number of unified memory controller per _die_ from two to eight, not per socketed CPU or any higher.
        I'm not saying that you are wrong. But this could mean that you can potentially have 8*4 channels NUMA in a 4-chiplet CPU for example.

        Comment


        • #5
          Originally posted by milkylainen View Post

          The statement really says nothing about the system memory topology.
          Only that they have increased the possible number of unified memory controller per _die_ from two to eight, not per socketed CPU or any higher.
          I'm not saying that you are wrong. But this could mean that you can potentially have 8*4 channels NUMA in a 4-chiplet CPU for example.
          I really doubt that NUMA is going away. The more cores you have the more sense it makes. This especially when you consider the markets AMD is going after. This could mean lots of things though from more core per chiplet to special function blocks on each chiplet.

          In any event i love love to see AMD being resurrected from the dead. In some ways I see their tech as the better solution for a variety of users.

          Comment


          • #6
            Originally posted by wizard69 View Post
            I really doubt that NUMA is going away.
            It's not going away for multi socket but Rome is definitely UMA for single socket. All chiplets have identical paths to memory with the same latency (chiplet -> I/O die -> memory). There is a small possibility that the chiplets that are physically further away from the I/O die have slightly higher latency but I assume that's insignificant.

            I pretty much rely on this will fix the Windows scheduling problems with the 24 and 32 core TRs. Since Microsoft obviously don't care.

            Comment


            • #7
              This patch does not work for the new Ryzen 3000 Zen2 series of AMD CPUs.

              I got the Ryzen 3900x with ECC memory and this EDAC patch does not support the new ryzen as far as I can tell... the new ryzen I have is Familiy 17h and model 71h, and the patch was for for F17_M30H (0x30 to 0x3F only). So there is currently no ECC support in the linux kernel with the new Ryzen that have been released. I managed to patch the linux kernel (5.2.1) by changing the PCI device IDs which appear to be different from all other devices so far in the AMD EDAC driver (different than F17 M30H as well). It appear to load the EDAC driver on boot and detect all ECC DIMM properly, but it does not report any ECC CE or UE errors although they appear to be happening based on my memory overclocking test and being corrected when ECC is enabled.

              It is a bit disappointing that it does not work, when I read this some time ago I felt that this would support all new Zen 2 CPUs, not just EPIC.

              Comment

              Working...
              X