• MNByChoice@midwest.social
      link
      fedilink
      arrow-up
      1
      ·
      5 hours ago

      that can also do GenAI work for a similar “hardware cost per output”? No

      FYI, the server hosts for the cards often have eight of the cards each. The power draw becomes the host server’s RAM and CPU, plus eight times 750w (or whatever). It scales up quickly.

      • Harbinger01173430@lemmy.world
        link
        fedilink
        arrow-up
        1
        ·
        5 hours ago

        Seems like optimization issue.

        If they can’t train and run a big ass ai model on igpu power at very fast speeds then they are useless as developer companies.

        So much bloat.