If local LLMs become mainstream then you want as much memory bandwidth as possible. For regular home and office use two channels of DDR4 is more than enough.

You really start taking into account bandwidth when you need 64GB or more (which is rarelly).

If it's audio/video, spawning VMs, it doesn't matter much. If it's for generative software, it might become an issue.

Unironically, small LLMs

LLMs and fast GPUs without having a giant PC spring to mind.