• ranzispa@mander.xyz
    link
    fedilink
    English
    arrow-up
    1
    ·
    30 days ago

    AI factories have CPUs and GPUs. Either way, TPUs can be used for other kinds of calculations as well. And definitely, given enough availability of TPUs, specialized software will be optimized to run on them.

    But either way, no this is not the case. I know people who work with LLMs and they mostly use GPUs, even when they work on US computing clusters.

    • Hirom@beehaw.org
      link
      fedilink
      English
      arrow-up
      1
      ·
      30 days ago

      Fair point, GPUs are often used too, and have other uses for HPC workloads.

      There is still a major misallocations of resources for AI. This infrastructure is badly needed by big tech for their AI growth plans. No one else need this much.

      There’s no good reasons to build at that scale and so fast, for any reason.

      There’s a need to lower overall energy and resources usage, and improve efficiency. Building many large datacenters for LLMs go in the wrong direction.

      • ranzispa@mander.xyz
        link
        fedilink
        English
        arrow-up
        1
        ·
        29 days ago

        I work in research and I have dire need of GPUs. I do not work on LLMs.

        Recently we got granted access to 500 GPUs for 2 months; all of them are spinning - and this is the work of 5 researchers. A normal research HPC institute normally has some 100-200 GPUs. In this sector there are startups which need to execute simulator calculations.

        Yes, GPU infrastructure is very much needed and lacking in Europe.