Typically, when it comes to geometry heavy scenes I prefer to "link" individual models into the final scene (MXS references) for the faster viewport performance. Basically, I open the individual model, say a tree, texture it, save it, and then link it into my scene, say a landscape (multiple trees, all referenced).
Now, for what it's worth (and if this experience helps your decision process in any way) . . . I think geometry dense Maxwell scenes tend to respond better on systems with more CPU cores.
I use The Grove 3d Trees a lot --'cause I like them
. However, these are dense, geometry heavy trees made for realism. I'm talking hundreds of leaves, hundreds of little berries on those leaves, sometimes, and twigs upon twigs (all instances in Maxwell of course).
One tree in particular, the Manna Gum tree is like molasses to load up in Maxwell on my GTX machine. Spinning wheel . . . Maxwell not responding . . . for minutes . . . every time . . . did I say every time . . . I so much as select a leaf on this tree.
On the other hand, I can open that same tree up on my Quadro machine and it handles it like kid's play. No lag. No Maxwell not responding spinning wheel. Nothing but "I got this".
Here's the thing: my Quadro card is a 2GB card. My GTX card is an 8GB card. The system memory on my Quadro machine is 16GB. The system memory on my GTX machine is 24GB.
So, what gives here? Well, the Quadro machine has dual Xeons --a 12 core 24 thread machine it is. The GTX machine is just a 4 core 8 thread i7 baby of a machine (operating at a faster clock speed than the Xeons --likely because it wants to grow up faster).
If I had to guess, it's those CPU cores doing all the heavy lifting with the Manna Gum tree.
Just a thought.
Hope it helps.