With sparse MoE it's worth running the experts in system RAM since that allows you to transparently use mmap and inactive experts can stay on disk. Of course that's also a slowdown unless you have enough RAM for the full set, but it lets you run much larger models on smaller systems.