• Electricd@lemmybefree.net
    link
    fedilink
    English
    arrow-up
    15
    arrow-down
    2
    ·
    edit-2
    2 days ago

    I want to run LLMs locally, or things like TTS or STT locally so it’s nice but there’s no real support rn

    Most people won’t care nor use it

    LLMs are best used when it’s a user choice, not a platform obligation

    • ulterno@programming.dev
      link
      fedilink
      English
      arrow-up
      1
      ·
      14 hours ago

      I guess an NPU is better of being a PCIe peripheral then?
      And it can then have their specialised RAM too.

      • Electricd@lemmybefree.net
        link
        fedilink
        English
        arrow-up
        1
        ·
        8 hours ago

        Sorry, I’m not a hardware expert at all

        When you’re talking about the PCIe peripheral, you’re talking about a separate dedicated graphics card or something else?

        I guess the main point of NPUs are that they are tiny and built in

        • ulterno@programming.dev
          link
          fedilink
          English
          arrow-up
          0
          ·
          7 hours ago

          When you’re talking about the PCIe peripheral, you’re talking about a separate dedicated graphics card or something else?

          Yes, similar to what a PCIe Graphics Card does.
          A PCIe slot is the slot in a desktop motherboard that lets you fit various things like networking (ethernet, Wi-Fi and even RTC specialised stuff) cards, sound cards, graphics cards, SATA/SAS adapters, USB adapters and all other kinds of stuff.

          I guess the main point of NPUs are that they are tiny and built in

          GPUs are also available built-in. Some of them are even tiny.
          Go 11-12 years back in time and you’ll see video processing units embedded into the Motherboard, instead of in the CPU package.
          Eventually some people will want more powerful NPUs with better suited RAM for neural workloads (GPUs have their own type of RAM too), not care about the NPU in the CPU package and will feel like they are uselessly paying for it. Others will not require an NPU and will feel like they are uselessly paying for it.

          So, much better to have NPUs be made separately in different tiers, similar to what is done with GPUs rn.

          And even external (PCIe) Graphics Cards can be thin and light instead of being a fat package. It’s usually just the (i) extra I/O ports and (ii) the cooling fins+fans that make them fat.