Thanks for taking the time to write that explanation. What you wrote makes sense based on the RAM usage scenario you portrayed. Essentially you are saying that, with conventional architecture, any memory usage by the VRAM also requires the same memory usage by RAM. So having separate VRAM doesn't add any additional capacity.
But aren't there use cases that would require significant VRAM without also requiring usage of that same amount of RAM, i.e. where the GPU was using a lot of VRAM for its own internal processing? E.g., suppose you send 1 MB of data to the GPU with a processing request (say, run a simulation on that data and give the result)*. It then processes that internally, requring 4 GB of VRAM to do so. Then once it's done it sends the answer back to the CPU, where the answer is 2 MB in size.
If the CPU needs 16 GB RAM while this is going on, then total memory usage is 16 GB RAM + 4 GB VRAM.
But with UMA (unified memory architecture), total memory usage during this process would be 20 GB RAM.
*Not saying things would actually work this way; just trying to illustrate a concept.
