A beefy GPU which can't hold models in VRAM is of very limited use. You'll see 16 GB of VRAM on gamer Nvidia cards, the RTX 5090 being an exception with 32 GB VRAM. The professional cards have around 96 GB of VRAM.

The thing with these Apple chips is that they have unified memory, where CPU and GPU use the same memory chips, which means that you can load huge models into RAM (no longer VRAM, because that doesn't exist on those devices). And while Apple's integrated GPU isn't as powerful as an Nvidia GPU, it is powerful enough for non-professional workloads and has the huge benefit of access to lots of memory.

which professional card has 96GB of VRAM?

Like the NVIDIA H100 NVL 94GB HBM3 PCIe 5.0 Data Center GPU for 27.651,20 €

https://www.primeline-solutions.com/de/nvidia-h100-nvl-94gb-...

that's not a professional line card, but a data center card.

LLMs are primarily "memory-bound" rather than "compute-bound" during normal use.

The model weights (billions of parameters) must be loaded into memory before you can use them.

Think of it like this: Even with a very fast chef (powerful CPU/GPU), if your kitchen counter (VRAM) is too small to lay out all the ingredients, cooking becomes inefficient or impossible.

Processing power still matters for speed once everything fits in memory, but it's secondary to having enough VRAM in the first place.

Transformers are typically memory-bandwidth bound during decoding. This chip is going to have a much worse memory b/w than the nvidia chips.

My guess is that these chips could be compute-bound though given how little compute capacity they have.

It's pretty close. A 3090 or 4090 has about 1TB/s of memory bandwidth, while the top Apple chips have a bit over 800GB/s. Where you'll see a big difference is in prompt processing. Without the compute power of a pile of GPUs, chewing through long prompts, code, documents etc is going to be slower.

nobody in industry is using a 4090, they are using H100s which have 3TB/s. Apple also doesn’t have any equivalent to nvlink.

I agree that compute is likely to become the bottleneck for these new Apple chips, given they only have like ~0.1% the number of flops

I chose the 3090/4090 because it seems to me that this machine could be a replacement for a workstation or a homelab rig at a similar price point, but not a $100-250k server in a datacenter. It's not really surprising or interesting that the datacenter GPUs are superior.

FWIW I went the route of "bunch of GPUs in a desktop case" because I felt having the compute oomph was worth it.

4.8TB/s on H200, 8TB/s on B200, pretty insane.

That’s wild, somehow I hadn’t seen the B200 specs before now. I wish I could have even a fraction of that!

VRAM capacity is the initial gatekeeper, then bandwidth becomes the limiting factor.

i suspect that compute actually might be the limiter for these chips before b/w, but not certain

> Transformers are typically memory-bandwidth bound during decoding.

Not in case of language models, which are typically bound by memory size rather than bandwidth.

nope

I assume even this one won't run on an RTX 5090 due to constrained memory size: https://news.ycombinator.com/item?id=43270843

sure on consumer GPUs but that is not what is constraining the model inference in most actual industry setups. technically even then, you are CPU-GPU memory bandwidth bound more than just GPU memory, although that is maybe splitting hairs

Why are industry setups considered actual while others are not?

What's more important isn't how beefy it is, it's how much memory it has.

These are unified memory. The M3 Ultra with 512gb has as much VRAM as sixteen 5090.

VRAM is what takes a model from "can not run at all" to "can run" (even if slowly), hence the emphasis.

No, with limited VRAM you could offload the model partially or split across CPU and GPU. And since CPU has swap, you could run the absolute largest model. It’s just really really slow.

The difference between Deepseek-r1:70b running on an M4 Pro (48 GB unified RAM, 14 CPU cores, 20 GPU cores) and on an AMD box (64 GB DDR4, 16 core 5950X, RTX 3080 with 10 GB of RAM) is more than a factor of 2.

The M4 pro was able to answer the test prompt twice--once on battery and once on mains power--before the AMD box was able to finish processing.

The M4's prompt parsing took significantly longer, but token generation was significantly faster.

Having the memory to the cores that matter makes a big difference.

You're adding detail that's not relevant to anything I said. I was saying this statement:

> VRAM is what takes a model from "can not run at all" to "can run" (even if slowly), hence the emphasis.

Is false. Regardless of how much VRAM you have, if the criteria is "can run even if slowly", all machines can run all models because you have swap. It's unusably slow but that's not what OP was claiming the difference is.

The criteria for purchase for anybody trying to use it is "run slowly but acceptably" vs. "run so slow as to be unusable".

My memory is wrong, it was the 32b. I'm running the 70b against a similar prompt and the 5950X is probably going to take over an hour for what the M4 managed in about 7 minutes.

edit: an hour later and the 5950 isn't even done thinking yet. Token generation is generously around 1 token/s.

edit edit: final statistics. M4 Pro managing 4 tokens/s prompt eval, 4.8 tokens/s token generation. 5950X managing 150 tokens/s prompt eval, and 1 token/s generation.

Perceptually I can live with the M4's performance. It's a set prompt, do something else, come back sort of thing. The 5950/RTX3080's is too slow to be even remotely usable with the 70b parameter model.

I don't disagree. I'm just taking OP at the literal statement they made.

Sure, this is technically correct, but somewhere there's a line of practicality. Running off a CPU (especially with swap) will be past that line.

Otherwise, you don't even need a computer. Pen and paper is plenty.

For all practical purposes, VRAM is a limiting factor.

Really, really, really, really, really, REALLY REALLY slow.

You can say the same about GPU clock speed as well…

When it comes to LLMs in particular, it comes down to memory size+bandwidth more than anything else.

I was able to run and use the DeepSeek distilled 24gb on an M1 Max with 64gb of ram. It wasn't speedy, but it was usable. I imagine the M3/4s are much faster, especially on smaller, more specific models.