well the card definitely doesn't use up all of the x16 bandwidth, but it might be the problem here. lets say just for proving a point, that the the card uses up x10 of the x16 allocated bandwidth. so if you use 2 of those cards in sli, then you are loosing out on x2 of the cards potential. I don't know if this is true, it would take research and such to prove it, but its possible.
also its true generally with any type of sli or crossfire, that the rule is its better to have the next model up then 2 of the model below.
so it would be better to have 1 gts/gtx/ultra then to have 2 gt's. I know its true with the gt vs gtx, not sure about the gt vs gts though.