On: What has changed in CPU cores in M3 chips?

[treprinum]

The option to have 128GB RAM for stable diffusion and local LLMs.

[moondev → treprinum]

Since 8GB unified memory is "equivalent to a 16GB non Mac" Is 128GB unified memory MacBook basically a 256GB Nvidia GPU(s)?

[lucasyvas → moondev]

They are not equivalent - it's pure marketing BS.

[treprinum → moondev]

Not in the machine learning world ;-) You'd need to switch to marketing workloads to make 8GB = 16GB.

[jug → moondev]

I think that one is mostly marketing but there's a grain of truth in terms of the combined effect of unified memory (which is fast as hell) and macOS memory compression. But I wouldn't put 8 GB past similar to 10-12 GB on another system without these features.

[moondev → jug]

Say I want to launch a vm with 8GB reserved memory. On a 10-12 GB machine I would have 2 to 4GB of memory left to actually allocate to the system. I don't see how unified memory magically fixes this. It's actually worse because even more memory has to be shared with the GPU.

This is like saying a 8GB disk is actually like a 10-12GB disk elsewhere.

[nottorp → moondev]

> Since 8GB unified memory is "equivalent to a 16GB non Mac"

Yeah, right.