[Dylan16807]

Get there by what mechanism? In the near term a good model pretty much requires a GPU, and it needs a lot of VRAM on that GPU. And the current state of the art of quantization has already gotten us most of the RAM-savings it possibly could.

And it doesn't look like the average computer with steam installed is going to get above 8GB VRAM for a long time, let alone the average computer in general. Even focusing on new computers it doesn't look that promising.

[SirHumphrey]

By M series and amd strix halo. You don't actually need a gpu, if the manufacturer knows that the use case will be running transformer models a more specialized NPU coupled with higher memory bandwidth of on the package RAM.

This will not result in locally running SOTA sized models, but it could result in a percentage of people running 100B - 200B models, which are large enough to do some useful things.

[Dylan16807]

Those also contain powerful GPUs. Maybe I oversimplified but I considered them.

More importantly, it costs a lot of money to get that high bus width before you even add the memory. There is no way things like M pro and strix halo take over the mainstream in the next few years.