Well the obvious first course of action would be to switch out one of the plugs for something else and see if you get the same issues. I don't have a lot of experience with the midi buffers, I changed them once and everything has been working fine for years but I guess I'd go changing them in increments and seeing if anything improves, though you mention having done that already.

Just remember that the results you may be getting using the HDSP9632 with Sonar for example on your system is not an indicator of how this card may perform on another system with another program.  But of course you have to work with what you have.
 
I am getting excellent results with this same card under Studio One on a Win 7 machine.  Plus the fact now we have the ability to set high output buffers with Dropout protection and low input buffers for super fast response means that all these things are possible under different circumstances.  And we also now have Low Latency monitoring added in.
 
The audio engine in the software also makes a difference.  I can loop over sections with total gapless performance with no CPU spikes while looped sections are repeating in large sessions with many synths playing at once.
 
I still have not tested the FF800 yet but will do so shortly.  It is also a great audio interface as well.
 
I also use the Midi IN/OUT on the HDSP9632 as an extra Midi port driving external hardware and it is also working super well too.  It has very low jitter on Studio One as well. 
 
Despite the fact this card is an older model it still works well even today and when the software changes the conditions under which it is operating then it too appears to go along with the changes and allow these new features to exist, rather well in fact.  It is interesting that what an audio interface can and cannot do can be tied to the software that it is connected to it rather than only doing its own internal thing.
 
Have you checked and updated the firmware.  I think you mentioned before you have. 

occide
Been following the Studio Tour of JunkieXL on Youtube. He's got 6 VST-Servers, the main machine running Cubase has 5 DSP cards with dedicated plugins that run on the cards, a separate machine with ProTools that handles all cinematic stuff when he's working on scores. Not to mention estimated 100qm² of hardware synths & samplers.
 
Pretty much any computational task you can "outsource" from your DAW he has outsourced. It kinda speaks for it's own.
 
EDIT: I wonder what it might take to have an Amazon EC2-Server as VST-Server. VST-Power on demand, you only pay what you use, scalable to infinity. Dafq this would be awesome. Let's create this! Latency might be a small issue
 
EDIT: Opening a lot of bottles here, but what about CUDA and OpenCL? I thought this was in the making already?

Occide I ran my splat with an rme 9652 ( same thing as yours minus da) without a hitch for years no problem. BUT if you have a copy of cubase around you can try that.  Many pluses and minuses with that daw.  I'm using it with my fire face and it works great.  Still have some electrical noise though.

Occide, I run a 9632 PCI with Splat and it is rock solid.  There is a helpful RME user forum here:
 
https://www.forum.rme-audio.de/
 
Best,
 
Tom