• catloaf@lemm.ee
    link
    fedilink
    arrow-up
    0
    ·
    2 months ago

    Usually people are against just throwing more hardware at a problem.

    They’re going to keep making more powerful hardware either way, since parallel processing capability supports graphics and AI just fine. But if they can use a novel software solution to drastically increase performance, why not?

    • tunetardis@lemmy.ca
      link
      fedilink
      arrow-up
      0
      ·
      2 months ago

      They’re going to keep making more powerful hardware either way, since parallel processing capability supports graphics and AI just fine.

      It’s not quite as simple as that. AI needs less precision than regular graphics, so chips developed with AI in mind do not necessarily translate into higher performance for other things.

      In science/engineering, people want more—not less—precision. So we look for GPUs with capable 64-bit processing, while AI is driving the industry in the other direction, from 32 down to 16.

      • catloaf@lemm.ee
        link
        fedilink
        arrow-up
        0
        ·
        2 months ago

        For science and engineering, workstation cards like the A6000 aren’t going anywhere.

      • averyminya@beehaw.org
        link
        fedilink
        arrow-up
        0
        ·
        2 months ago

        It’s funny because we don’t even need GPU’s. There’s tech that offloads the model’s “search” to an analog computer which is ~98% accurate for a fraction of the energy.

        I imagine NVIDIA isn’t too excited about that side of AI, though.