I just sold mine on Ebay this morning. (At least I
think it wasn't
 shared memory. Hardly remember anymore. Anyway, it's gone...) 
It's not that easy to get anything to Russia :)
  Is it possible to do 8in/8out and 64-sample
buffers with RME Multiface
 CardBus on a laptop with shared video memory? 
 Depends on the laptop but possibly. Why so stringent about 64/2? Do
 you _really_ require sub-3mS?
 
 Processing live inputs better be done with lowest latency.
Since 64/2 is only a part of total latency, increasing this value is
noticable. Especially when playing guitar.
 Possibly, but even shared memory probably won't kill you if you're
 just running Ardour, etc.
 
I hope it will not kill me for running some live processing stuff  -
SooperLooper, Jack-rack, etc.