So this is probably not good news for the MacBook Ultra with 512GB of RAM rumors being..affordable.
What's worse is that this is probably going to get worse. My angel investment group is getting inundated with pitches that amount to building an RX-6000 with 96GB of RAM and installing a local model to do "thing X".
So even if the OpenAI's of the world stop trying to use up all the RAM, you're going to have thousands of start-ups pushing local models.
Makes me really wonder about that new Surface Ultra pricing with the nvidia chip in it.
If Apple can't pull it off with their supply chain weight they can throw around, what is that thing going to be priced at? Microsoft/Nvidia are either going to be subsidizing it or it's gotta be close to $8,000+ at launch.
Compute? Inference doesn't only need memory bandwidth. You need to actually do work with the memory you're loading which needs compute power. Which needs more electricity, which needs more cooling, which isn't practical for something as thin as a MBP.
MacBook Pro has plenty of compute for local LLMs to be usable. I'm getting up to ~150 tokens/s with Deepseek-v4-Flash on a MacBook M5 Max. It's quite capable for coding assistant usage.
In general LLMs are bottlenecked by memory bandwidth rather than raw compute power.
Yes, it's quantized (4 bit). Sure, it's... not quite as good as what's on offer via API. And sure, "up to" does a lot of work (I don't have an average/median for you but it feels fast to me).
But it's usable, fully local, fully private, and has no subscriptions and no operating costs other than electricity.
I mean it’d take minutes of research to realize people are successfully and efficiently running 4-bit quantized GLM 5.2 on MacStudio 512GB M3 Ultras at over 60 tok/s. K2 2.7 is quite literally designed for 4 bit quantization and runs even better.
The integrated GPU. Not enough compute onboard to handle prefill for 100gb+ models, and the decode is constrained by memory bandwidth that's lower than most dGPUs that price.
Apple would be in a much stronger spot right now if they didn't pretend like eGPUs were inconceivable black magic that Macs are incompatible with.
I'm not sure I follow - 614 GB/sec is pretty squarely in dGPU territory (~5070 level). External GPUs can definitely exceed that on the very high end, but it seems pretty competitive, no?
Competitive for 16-24GB dGPUs, but for 100gb+ inference workloads it's going to be a decode bottleneck. For smaller models it'd be fine, but the same goes for the smaller GPUs.
In particular though, the fatal bottleneck is the weakness of the iGPU. Filling a KV cache on a 100gb+ model could take a few minutes, or even hours if you're trying to restore a 256k-to-1m token session.
What's worse is that this is probably going to get worse. My angel investment group is getting inundated with pitches that amount to building an RX-6000 with 96GB of RAM and installing a local model to do "thing X".
So even if the OpenAI's of the world stop trying to use up all the RAM, you're going to have thousands of start-ups pushing local models.