top 50 comments

sorted by: hot top controversial new old
[–] 125 points 2 weeks ago* (last edited 2 weeks ago) (3 children)

Basically Apple will be building the perfect computers to run local LLMs.

  • source
  • hideshow 5 child comments
  • [–] 96 points 2 weeks ago (14 children)

    One can only hope that it totally breaks the AI/LLM at industrial scale, so businesses can run their own AI systems with their own data sets.

    No more of this fucking datacanter horseshit.

  • source
  • parent
  • hideshow 15 child comments
  • load more comments (13 replies)
  • [–] 4 points 1 week ago* (last edited 1 week ago) (3 children)

    I guess it depends on your definition of perfect - cheap, good or fast.

    This thing is probably going to cost at least $20K USD.

    Edit:

    "Next year's base M7 processor is expected to arrive in the first half of 2027 and will also upgrade memory bandwidth to about 240 GB/s."

    That's...really fucking slow. What's the goal here - CGI, engineering sims, game dev etc? 1.5TB is cool but at 240GB/s that will crawl for AI use.

    Comparison: this is about $100K, for 7.2TB/s, 252GB VRAM (+500GB system ram, so closer to 750GB total)

    https://www.nvidia.com/en-us/products/workstations/dgx-station/

  • source
  • parent
  • hideshow 5 child comments
  • [–] 3 points 1 week ago

    It’s likely wrong reporting, the m5 ultra gets 614 GB/s today. The m3 pro gets over 800. The rumored m5 pro is 1.2TB/s. Likely this will be 2.4 TB/s, but half the reason for high bandwidth in nvidia chips is somewhat offset in the unified scheme. Mlx needs a lot less copy data around when the gpu can just read it directly.

  • source
  • parent
  • [–] 2 points 1 week ago*

    Something got reported on incorrectly because the M5 Max's have 614 GB/s today, and the Ultra M4 machine's (not laptops) are 819 and that's 3 generations behind a M7 if they make an M7 ultra machine.

    $ for $ you'll get more video ram than paying for a 5090, but it won't be as fast and can't train well.

    Before the ram price decable, you could get a 192gb M4 Ultra for ~10k CAD.

  • source
  • parent
  • load more comments (1 reply)
  • load more comments (1 reply)
    [–] 78 points 2 weeks ago* (3 children)

    A Mac with 1.5TB RAM would be expensive at the best of times. In 2027/2028 it might approach $50k, or even get into 6 figures.

  • source
  • hideshow 6 child comments
  • [–] 29 points 2 weeks ago (2 children)

    It shouldn't be, if Apple gets that import exemption from the Chinese memory manufacturer (ChangXin Memory Technologies) they are asking for. Apparently they're already testing the CXMT chips to put in the phones sold in China, freeing up the orders/stock they've sourced already from "safe" sources, to go into their products sold in the rest of the world.

    Smart move if they can finagle it.

    Hopefully they can get it through before the fuckwits in the administration understand how effectively it can threaten the big AI players that Trump seems to be sniffing around.

    You absolutely BET that he will scuttle any trade deal if it interferes with his own personal agenda WRT his investments in AI.

    He's that much a greedy cunt.

  • source
  • parent
  • hideshow 4 child comments
  • [–] 32 points 1 week ago (3 children)

    1.5 TB of unified memory sounds less like a computer and more like Apple preparing for the moment your local AI starts asking for a raise. Plot twist: by 2028 the RAM upgrade still costs more than the rest of the machine combined.

  • source
  • hideshow 6 child comments
  • [–] 13 points 1 week ago (2 children)

    Remember this is “unified”, it’s not like you can upgrade, nor is it available in the “cheap” packaging we’re used to.

    You’ll get whatever Apple puts on the SoC, and you’ll be happy with it

  • source
  • parent
  • hideshow 4 child comments
  • [–] 12 points 1 week ago (1 child)

    The upside is that unified memory is genuinely different from traditional RAM. The CPU, GPU and Neural Engine all share the same memory pool, so data doesn’t need to be copied back and forth. That reduces latency, improves efficiency and lets AI models, graphics and other workloads access much larger datasets. It also uses less power and saves board space. The downside is obvious: because it’s integrated into the chip, you have to choose the right amount upfront, since it can’t be upgraded later.

  • source
  • parent
  • hideshow 2 child comments
  • [–] 4 points 1 week ago*

    Ya, these high memory amounts and ever increasing memory bandwidth are heavily (but not only) targeting people wanting to run local large AI models like a full deepseek on their machines.

    You might not be able to train as well on them as NVIDIA + CUDA, but for local inference, they're an alternative to NVIDIA and more reasonably priced for the model sizes you can run, and each iteration they get better as the bandwidth increases.

  • source
  • parent
  • [–] 7 points 1 week ago (1 child)

    Apples memory has never been cheap, it's always been a very expensive upgrade.

  • source
  • parent
  • hideshow 2 child comments
  • [–] 2 points 1 week ago (1 child)

    Plot twist: by 2028 the RAM upgrade still costs more than the rest of the machine combined.

    Won't at least China have production capacities ready by then that make the price drop?

  • source
  • parent
  • hideshow 2 child comments
  • [–] 24 points 1 week ago (1 child)

    Can I load all of GTA 6 into RAM?

  • source
  • hideshow 2 child comments
  • [–] 5 points 1 week ago

    Making a RAM drive to load a few minutes of rolling video game footage so it isn’t constantly writing to my SSD or HDD was one of the most “the future is now” things I’ve done lately… and it’s not a new concept, I just never considered it before.

  • source
  • parent
  • [–] 20 points 1 week ago

    Remortgage your house now.

  • source
  • [–] 15 points 1 week ago (1 child)

    Yeah, but can it run Crysis?

  • source
  • hideshow 2 child comments
  • [–] 10 points 1 week ago

    There's going to be a massive boom in local llms if we get there. Already I have replaced ~80% of my paid token usage with a small local model running on my 64gb macbook pro, but being able to run a full-fidelity multi-hundred-billion parameter llm locally would be a game changer for my use case at least.

  • source
  • [–] 9 points 1 week ago

    For $100k probably

  • source
  • [–] 8 points 1 week ago (1 child)

    You'll only be able to afford 640KB, but it CAN go to 1.5TB!

  • source
  • hideshow 2 child comments
  • load more comments
    view more: next ›