@cm0002 No good answer for RAM right now but with respect to GPUs, consider Intel Arc. I purchased a B50 Pro for a whole of $300, it's turned out to be a reasonably solid performer. I have SR-IOV working on it now a vf for a windows vm, and one for LLAMA inference, I'm using a mamba model so unlike the big boys my LLM learns on the fly. And this thing draws a whole of 50 watts so you don't need your own nuclear power plant to run it.