I've had my rpi (still using 2012-12-16-wheezy-raspbian) running for a few days now with 1xMMQ and 1xBFL without any trouble
(with only one manual restart coz I rewrote a part of the MMQ driver)
I noticed that the MMQ was using about 3% CPU per daughter board (so about 12% for an MMQ)
So I changed the MMQ code to only poll every 100ms (instead of every 10ms) and check the temperature every 500ms (instead of every 300ms)
This has reduced it to roughly 0.5% CPU per daughter board (so about 2% for an MMQ)
On top of that of course there is the main cgminer code - and that's using somewhere around another 0.5% CPU Stratum mining on OzCoin
(and the BFL is using a bit under 0.2% CPU)
This does bring up some information in advance for ASIC
The BFL SC polling for results will probably be only a few times every nonce range time - so for 80GH/s it will be ~54ms per nonce range so maybe poll once every 1/3 of that (or maybe only once per nonce range will be enough?)
Anyway I haven't decided yet and of course wont until I actually have some hardware - but the reality is that it may well only be as much USB CPU usage as the MMQ code before I changed it.
The amount of data sent to do the work would be of course be one queued element per result (based on the current BFL spec) thus it may even be realistic to consider the USB CPU usage of the BFL SC Single to be less than the old MMQ code.
However, the main cgminer code will be processing 70~100 times the number of 1diff shares.
So I'd guess that the rpi should be able to handle a single 60~80GH/s ASIC, but not sure how that will go with more than one.
Of course until I get some ASIC hardware and can see the real CPU usage, this is all just guess work
P.S. my new MMQ code is in my 'hotplug' branch here:
https://github.com/kanoi/cgminer/tree/hotplughmm maybe the MMQ needs an --mmq-options so you can specify the poll time ...
(something like Icarus has for specifying options)