Hacker News Reader

LLMs run on the GPU, and the unified memory of Apple silicon means that the 64 GB can be used by the GPU.

Consumer GPUs top out at 24 GB VRAM.

llama.cpp can run LLMs on CPU. iGPU can also use system memory, the novel thing is not that, it's that the LLM inference is mostly memory bandwidth bound and memory bandwidth of a custom built PC with really fast DDR5 RAM is around 100GB/s, nVidia consumer GPUs at the top end are around 1TB/s, with mid range GPUs at around half that. M1 Max has 400GB/s, M1 Ultra is 800GB/s, but you can have Apple Silicon Macs with up to 192GB of 800GB/s memory usable by GPU, this means much faster inference than just CPU+system memory due to bandwidth and more affordable than building a multi-GPU system to match the memory amount.

It'd be really nice to have good memory bandwidth usage metrics collected from a wide range of devices while doing inference.

For example, how close does it get to the peak, and what's the median bandwidth during inference? And is that bandwidth, rather than some other clever optimization elsewhere, actually providing the Mac's performance?

Personally, I don't develop HPC stuff on a laptop - I am much more interested in what a modern PC with Intel or AMD and nvidia can do, when maxxed out. But it's certainly interesting to see that some of Apple's arch decisions have worked out well for local LLMs.

Reply on news.ycombinator.com