There are ultra low powered LLMs but even then you’re looking at at least 2GB, and the most typical graphics card has 8, so it’s going to be some impact at least. Their intelligence/capability scales hard with memory usage too, so for most things you might want to use it for the smallest ones likely would not be good enough.
For example there’s a Rimworld mod that adds locally generated flavor text dialogue, using such a low powered model. But it’s a really simple feature that doesn’t affect actual gameplay at all. Games where the gameplay interacts with LLM output in any way are going to have higher hardware requirements, to the point where they will need to use the graphics card more for that part and less for the actual graphics; it’s enough of a bottleneck that anyone wanting to do this will basically need to design the game around it.
Basic tasks like that aren’t super resource heavy, so I don’t think it would impact hardware requirements all that much.
Training the models is the most resource intensive part.
There are ultra low powered LLMs but even then you’re looking at at least 2GB, and the most typical graphics card has 8, so it’s going to be some impact at least. Their intelligence/capability scales hard with memory usage too, so for most things you might want to use it for the smallest ones likely would not be good enough.
For example there’s a Rimworld mod that adds locally generated flavor text dialogue, using such a low powered model. But it’s a really simple feature that doesn’t affect actual gameplay at all. Games where the gameplay interacts with LLM output in any way are going to have higher hardware requirements, to the point where they will need to use the graphics card more for that part and less for the actual graphics; it’s enough of a bottleneck that anyone wanting to do this will basically need to design the game around it.