Hacker News Reader

Comment by ActorNightly | Hacker News Reader

Honest question - why are you so stuck on Macs for local inference?

A 4 GPU linux box with 3090s, which are $1500 a piece right now, will blow this thing out of the water. Even 2x3090 rig will run most of the good local models like Gemma4:31b at 100+ tok/sec

The VRAM of the GPUs are MUCH faster than the unified ram within Apple Silicon. The only difference is the initial model load, which takes longer from disk to VRAM due to PCIE limitations, but once the model is loaded, GPUs can prefill and and generate tokens way faster than any Apple Silicon.

So is your desire to upgrade to mac because you just aren't aware of how to set up a GPU rig, or is it something else?

View on HN

That's 6 YEARS of their cloud inference!

And, you have to include power. Where I am, the power for a rig with 2x 3090, running for four hours a day, would be around $70/month!

It's kinda like datacenter only exist to optimize compute cost through oversubscription of hardware/time share, cheaper business energy rates, and bulk discounts, compared to self hosting. ;)

Running local models isn't about the cost, its about owning your own data and being able to run uncensored models. If you just want to code with an assistant, local inference is by far not worth it.

You also don't need to run the models 24/7, nor is your computer gonna be doing inference all the time when you are coding.

No, you're mixing unrelated concepts here. Cloud does not mean you don't own the data, and can't do what you want. It means it's running on commoditized server hardware, powered with cheap commercial energy costs, oversubscribed so nothing sits idle.

AIUI you can only connect two 3090s at a time with NV link? So you’d only have 48GB of fast combined memory? That’s not tremendously interesting as you’re still limited to the smaller models which, while impressive in their own right, are still IMO too limited to use as your only model.

There are kernel patches to enable P2P communication via PCIE on 3090 which is almost as fast as NVLink for vllm

Please correct me if I'm wrong, but:

* 3090 memory bandwidth: 936 GB/s

* Maximum NVLink bandwidth between two 3090s: ~56 GB/s one way

* Maximum PCIe v4 bandwidth between two 3090s: 31.5 GB/s one way

* M5 Ultra memory bandwidth: 1.2 TB/s

I know that memory bandwidth is only one factor influencing LLM performance, but this seems like a major problem if your goal is to run larger models that won't fit on one 3090 - and even those fitting on two 3090s with NVLink will be pretty restricted.

Upgrading RAM is still probably cheaper than spending $6000 on a linux box. You're correct that inference will be much faster on the Linux box. But, the mac's unified memory can load larger models. And as OP mentions, it's still hard to beat cloud pricing at home.

Yes, GPUs are much better for dense models. On Macs, MOE models run better, so I agree Macs are more limited and expensive. I have a Linux laptop with a 10GB 1080 GPU, dated, but I should add even more system RAM and try that.

A Mac mini running an LLM is quiet

A PC with similar capabilities is going to sound like a jet taking off.

>A PC with similar capabilities is going to sound like a jet taking off.

Not at all. Airflow with big fans is quiet. What I do hear is coil-whine. In fact my PC is quieter than my Macbook when both are running top speed. But one has 4000 AI TOPS.

A Mac Mini is easy to match by a PC in terms of inference, and the PC will win without getting noisy.

Reply on news.ycombinator.com