Apple

Out of context: Reply #3721

  • Started
  • Last post
  • 3,882 Responses
  • ********
    -2

    "Apple unveils new Mac Studio, the most powerful Mac ever, featuring M4 Max and new M3 Ultra"

    https://www.apple.com/newsroom/2…

    ********
    -2Permalink
    UpvoteDownvote
    Flag
    • M3 Ultra goes up to 512GB RAM
      ********
    • But will wait for next 1TB RAM versions then running LLMs locally will be great
      ********
    • i mean a 512GB could theoretically run a model of 200 million parameterskingsteven
    • Llama for sure, but for R1 I think you need more, so 1TB RAM next iteration should do it to have good performance
      ********
    • RI is 671B parameters, so you'd really need 2TBkingsteven
    • although, the more, the better.kingsteven
    • Found the specs to run R1 models:

      https://apxml.com/po…
      ********
    • yeah 1,543 GB so would require 2TBkingsteven
    • looks like DeepSeek-R1-Zero = 671B ~436 GB would run on the M3 Ultra 512GBkingsteven
    • crazy specs though if it can compete with 6 x A100 80GB in any way - thats a datacenter with $200k of hardwarekingsteven
    • Saying that, i've seen people build systems to run R1 locally for £2k and it looks to be fast enough for personal use.kingsteven
    Show [[ numHiddenNotes ]] more notesAdd Note
    SaveCancel

View thread