ATI GPUs have fine-grained redundancy which means turning off 1 ALU in 17, for example. So every GPU is built slightly bigger than needed and then the vast majority of defects are "captured" by the fine redundancy, resulting in a full-spec GPU. Clock speeds are obviously a separate matter. And apparently there were some HD3690s based on recovered RV670s. Seemingly these have half the memory bus active.
Jawed
But that's only for the ALUs themselves, not the schedulers, the TMUs, the ROPs, the MCs, etc. - so it's not magically going to catch as many defects as having coarse redundancy everywhere as NVIDIA does. On the other hand, the nice thing with their approach is that they can sell more chips at the maximal or near-maximal ASPs, so the gross margins calculations aren't fully obvious.
BTW, it's not just about redundancy. I stumbled about this very nice whitepaper from Synopsys recently:
http://www.synopsys.com/products/solutions/dfm/higheryield_article.pdf - clearly they are focused on reducing htospots which would cause serious yield issues, but another obvious implication is that lower transistor density (not just using more transistors for the job, that's not the point) in general can improve yields through all the main defect mechanisms, for both functionality and variability reasons.
This is also why I expect, for example, the density of the 'unique' part of modern GPUs with no redundancy to be noticeably less dense than the rest of the chip; it's not just about I/O & analogue not scaling, it's also about some parts of the chip not being as dense on purpose. This might in part be able to explain NVIDIA's lower transistor density on 65nm (see: G96 vs G84, for example) but doesn't really explain why their transistor counts per unit went up despite relatively small changes and relatively unimpressive clock speeds. So yeah, it's not all excusable either!
Another fairly obvious consequence of the above is that yields really aren't that much about die size anymore. It still matters, but it's far far from the main factor. How good your designers & your design tools/flow are is likely orders of magnitude for important, especially for a design with inherent coarse redundancy. Sorry for the semi-OT BTW, but I figured that given the yield discussions for GT200 this was all well worth pointing out.
