I’d suggest that Omnisphere is a friend , not a problem. Imagine that you want to have three parts each having different voices (patches) and you were using a monotimbral softsynth like DUNE. You would need three instances of DUNE on separate tracks to do this.
If in fact the voices could be provided by Omnisphere, then you only need one instance of Omnisphere. Although the three voices will consume resource (CPU and memory) because of the shared usage of function, there will be less overall resource usage.
Say DUNE uses R1 units of resource (CPU / memory) for one instance, then there will be 3xR1 used for the three voices
If Omnisphere uses R2 units, the total for the three voices would be R2 + 3m, where m is the marginal extra resource for adding a voice. Bottom line is that if an instance of DUNE consumed the same amount as an instance of Omnisphere (i.e R1=R2) then Omnisphere would win hands downs in resource usage.
Probably Omnisphere would need more than DUNE for a single instance, then rapidly move ahead with extra voices added - it’s all about what code/memory is shared - none in the case of DUNE, some in the case of Omnisphere
Later on I’ll post a simple Omnisphere example with routings.
Warning, if you are not familiar with ‘sysex’ BOME might be a step too far. It’s very powerful, and I believe that Ableton used BOME products in the development of PUSH.
A better bet might be to take a look at Canatabile Cantabile - Software for Performing Musicians
What is doesn’t say is that they have a free version as well, which I used initially before going for the USD 69 version. This can do some of what BOME does in a more user friendly way.
See Playing with multiple MIDI ports (post 1) - (very) poor man's divisi to get some idea.
Now I’ve confirmed my understanding of your goals, I’ll have a look later today (I’m London based) , and post something this evening if possible - depends how long it takes me to experiment.