Comment by storystarling

Have you tried running any local LLMs via llama.cpp? I am curious if that high RAM is effectively usable as unified memory for larger models. I wonder if the memory bandwidth is sufficient to get decent performance on something like a 70b model or if it bottlenecks.