the sweet spot is $9999 USD for 512GB unified memory
but that seems like an M5 maybe even M6
Currently its $5000 USD for 128GB on M3 I believe.
There just isn't anything good LLMs to run with current unified memory limits, especially 4090/3090s that can barely run anything special.
Many enterprises have bought m3s, setup their own sub 100gb VRAM gpu rigs but all in all report very little economic value being derived from it.
Even if local LLaMA's were as good as OpenAI or Anthropic, there is a disappointing lack of applications for it currently speaking with different users.
Curious to know if anybody has any counter anecdotes but really feels like AI winter is coming this winter.
> Many enterprises have bought m3s, setup their own sub 100gb VRAM gpu rigs but all in all report very little economic value being derived from it.
Literally who? I cannot name a single business that would be dumb enough to go "let's spend tens of thousands of dollars on AI but not the Nvidia hardware" and then act surprised when they get no ROI. Apple Silicon is not AI datacenter hardware; it is consumer silicon sold at consumer silicon markup. This only happens if you rely on organizational structure that allows you executives to override decisions vetoed on a technical basis. Smart orgs don't make these kinds of mistakes.
AI is not and has never been a driver for Mac adoption. This idea that Unified Memory is somehow destined to be the savior of bad LLM models is unfounded and repeated without evidence. Nvidia currently ships server hardware with interconnect speeds comparable to M3's bandwidth, and we've got just about nothing special to show for it. If Apple Silicon was better value for AI operations, we'd all know by now. Surprise: it's not.
If it's the right machine for you at the right time you should buy it. If it's not there's no harm in waiting. I don't know what this post is supposed to be about.
It's mostly some benchmark summaries and notes that the older ARM Macs are still good. Yes they are, so buy the right computer for you.
but that seems like an M5 maybe even M6
Currently its $5000 USD for 128GB on M3 I believe.
There just isn't anything good LLMs to run with current unified memory limits, especially 4090/3090s that can barely run anything special.
Many enterprises have bought m3s, setup their own sub 100gb VRAM gpu rigs but all in all report very little economic value being derived from it.
Even if local LLaMA's were as good as OpenAI or Anthropic, there is a disappointing lack of applications for it currently speaking with different users.
Curious to know if anybody has any counter anecdotes but really feels like AI winter is coming this winter.