• Spacehooks@reddthat.com
    link
    fedilink
    English
    arrow-up
    58
    arrow-down
    1
    ·
    23 hours ago

    I saw a random AI video where guy said to use this model you need 39gb of vram. Like wtf are these ppl running at home?

    • Domi@lemmy.secnd.me
      link
      fedilink
      arrow-up
      35
      arrow-down
      1
      ·
      22 hours ago

      39 GB is very small, DeepSeek R1 without quantization at full context size needs almost a full TB of RAM/VRAM.

      The large models are absolutely massive and you will still find some crazy homelabber that does it at home.

      • 87Six@lemmy.zip
        link
        fedilink
        arrow-up
        15
        ·
        edit-2
        19 hours ago

        All that RAM for the idiot AI to tell me what I can find on stackoverflow with one startpage search.

    • waz@feddit.uk
      link
      fedilink
      English
      arrow-up
      11
      arrow-down
      2
      ·
      edit-2
      23 hours ago

      I put (regular RAM) 64gb on my home pc, because that was the max my board would take. My old Mac Pro, 96gb because it was the most it could run at max speed, total could have been 128. Both only for 8gb gfx cards. Both because, I might want to open 400 tabs on a browser or something, maybe casual gaming, lol

        • DrDystopia@lemy.lol
          link
          fedilink
          English
          arrow-up
          2
          ·
          16 hours ago

          No, it’s both - Offloading to system RAM is normal for regular users with consumer level hardware.

        • FlexibleToast@lemmy.world
          link
          fedilink
          English
          arrow-up
          6
          ·
          22 hours ago

          Mac has unified RAM. It can use the system RAM as vRAM. The AI line of AMD processors can kind of due that too. Granted these aren’t as fast as dedicated GPUs, but they’re the most affordable way to get huge amounts of vRAM.

        • waz@feddit.uk
          link
          fedilink
          English
          arrow-up
          2
          ·
          19 hours ago

          The meme differentiates between shortage of GPUs and shortage of memory, so I thought it was about mobo ram, but I get that the comment I replied to mentioned vram.

        • Dudewitbow@lemmy.zip
          link
          fedilink
          arrow-up
          1
          ·
          edit-2
          20 hours ago

          vram in the context of an igpu like an apple chip or strix halo is the same thing as system ram. its shared memory

          its why strix halo and apple m4 chips are popular with users running local ai models, because those will cost you 2-4000$ for 128 gb ram, while the closest nvidia alternative is the RTX 6000 blackwell with 96gb vram costing 2-4x more.