3 GPUs and 3 TB RAM.

What I am trying to say is, take 3 GPUs and 1TB each memory. Load the data from SSD, calculate LLM and check the total time. Now add 1 TB of memory each and see the change in time. Keep on adding the memory until there’s no performance improvement. Don’t add the GPUs at all. Of course someone needs to write a program for doing all this.