ever since we started the AA, I have noticed (on my two 200 series cards) that the GPUgrd WU's seem to get a bit of run time each, swapping between one and the others, as opposed to finishing one off before starting the next.
I wonder if it is worried that they wont get done befor the deadlines or soomething?.
Also, I'm sure I have seen the message (I think on the Q6600 with the DGT9600) "reached limit of two tasks per host " or similar... but the other two machines have half a dozen each...?