• JuanElMinero@alien.topB
    link
    fedilink
    English
    arrow-up
    1
    ·
    11 months ago

    Feels like DDR5 capabilities have been running away from what boards and memory controllers can reasonably support, bar the lowest amount of ranks/channels and non-optimal clock ratios.

    • djent_in_my_tent@alien.topB
      link
      fedilink
      English
      arrow-up
      1
      ·
      11 months ago

      And it’s so painful for me since my workload is completely memory bandwidth bound but does not scale across multiple sockets, much less multiple nodes.

      96 cores per socket? 😎

      12 channels of ram per socket? 🤩

      DDR5-4800? 🫠

        • djent_in_my_tent@alien.topB
          link
          fedilink
          English
          arrow-up
          1
          ·
          11 months ago

          A specialized computational fluid dynamics program. And nominally CFD is great to scale across many sockets/systems but it seems this particular program wasn’t written to be able to do that.

    • topdangle@alien.topB
      link
      fedilink
      English
      arrow-up
      1
      ·
      11 months ago

      well technically the high freq chips right now are out of spec and overclocked. people just don’t care because chances are good that they work. XMP/EXPO are overclocks but people treat them like standard settings.

      so it makes sense that current IMCs and boards are struggling with these speeds since they’re far out of the spec they were designed for years ago.

    • VenditatioDelendaEst@alien.topB
      link
      fedilink
      English
      arrow-up
      1
      ·
      11 months ago

      Ultra-high density chips is exactly what you want to support the largest possible memory amounts and speeds with minimal ranks. With 32 Gib chips, you could build 32 GiB single-rank UDIMMs, or 64 GiB dual-rank.

      That means up to 128 GiB of RAM in mini-ITX!