• istewart@awful.systems
        link
        fedilink
        English
        arrow-up
        3
        ·
        5 days ago

        Very true. I think one of the possible low-key outcomes of the bubble is a rise both in open-source driver hacking and manufacture-on-demand PCBs to accommodate what would otherwise be high-dollar e-waste.

        • fullsquare@awful.systems
          link
          fedilink
          English
          arrow-up
          3
          ·
          5 days ago

          isn’t openai silicon breaking all the time because it’s so overheated? so they have to replace it a lot of the time? maybe it’s only good for some months

          • istewart@awful.systems
            link
            fedilink
            English
            arrow-up
            6
            ·
            5 days ago

            Or their cooling design stinks, or they/Nvidia are just telling the fab we don’t care about yields, just send us everything that powers on and we’ll figure out which ones are good in production

            • David Gerard@awful.systemsOPM
              link
              fedilink
              English
              arrow-up
              7
              ·
              5 days ago

              basically. Nvidia sometimes has off years, where a processor generation doesn’t work out. Unfortunately, this one’s coincided with a stupid bubble. So they’re shoving out number cruncher cards which are at the limits of what you can do with stacking up the previous generation of chips, and the cards are crappy and have a likely lifetime in months - because they correctly estimate their market doesn’t care.

    • dai@lemmy.world
      link
      fedilink
      English
      arrow-up
      4
      ·
      5 days ago

      Not really, there are methods to use an iGPU as your display out but do the rendering on a dedicated card. GPU Passthrough.

      There are also methods for using two GPUs even from different OEMs to increase framerate, see Lossless Scaling (windows).