Thanks for all the helpful responses!
I did not thought about the impact of the technical implementation of the interface. Even with absurd low buffers like 32 samples @96K I have roundtrip latencies of ~ 7ms.
I read a lot about the recommendations you all and think something like a MOTU hybrid will do the job: It is way cheaper than a RME and seems to be a solid device. If it will be to limited in the future, it would be a nice mobile solution. A pcie solution with all the additional gear will definitly be oversized. You all saved me a lot of $$$...

A couple of month later, I did my decision and bought a new interface. Based on the DAW benchmark results you can find at Gearslutz, I took a RME HDSPe AIO. I did not expect the fantastic results:
- Stable playback of heavy projects with 5ms RTL - I had to freeze some of the tracks at 30ms with the old interface.
- At least 12db less noise via pre-amp or synth
- Now it is fun to use TH3!
- No wired noises anymore (not speaking about Drop-outs, I think the reason was power related through USB)
- Even the phone output is way better.
 
I am glad that I waited and spend a little bit more. 

i've been very impressed with my newly acquired rme babyface pro...
 
granted, it's only got two preamps,
4 inputs in all analog,
but for me that's enough.
 
i can slave my 18 track vs roland machine to it if i need to, but i rarely track more than two tracks at a time.
 

The RME AI/O is primarily used in scenarios for digital I/O.  
Also popular with composers primarily using Virtual Instruments
 
With outboard pre-amps and A/D, you can achieve super low noise-floor.
Prime time outboard can get costly...