Well as long as the requirement is under 10% I can't imagine any non-CPU miners could have a problem.
With FPGA the issue is of course that they are new and the designers haven't quite got the idea about how GPU mining works to make sure you minimise lost work
With the BFL - the current device doesn't reply with any shares until it's finished processing the full nonce range.
In that case the efficiency would usually be close to normal on a normal pool.
However, of course, with anything like P2Pool with short LPs (10 seconds) that's gonna mean a lot of aborted work - probably 33% of getworks will never give a share.
If you issue extra LP's (e.g. for NMC mining) then that will of course lower the BFL efficiency %
With the Icarus - the current device only returns one nonce per request and then stops mining.
This means you need to make sure you queue up enough work so that if you get a few quick shares it doesn't end up being idle.
It is also aborted before it finishes the nonce range to avoid going idle (the current device doesn't say when it finishes if it doesn't find a share)
These typically show around 50% efficiency on cgminer due to the fact that of course on average you process under half of each work request.
I could have given you info on the ZTEX also except the guy making it was unhelpful when I requested info so I gave up bothering to add that to cgminer once I found it works very different to the rest and I'll let someone else do that
They are the only 3 I've have any contact with.