Since I do not, never have, and 99.9999999% probably never will work at AMD, I can't say for sure what was going through their heads. I think they realized that they:
a). leaned too heavily on k8's future
b). got hurt badly by the failure of the "k9" uarch that they never released
c). got hurt badly again by Phenom (Agena)
That Phenom II/Deneb worked as well as it did, and that Thuban and its server-based cousins worked as well as they did, was a small miracle for AMD. Regardless, it was still behind Nehalem, and they could probably see the handwriting on the wall: Intel was going to outpace them slowly but surely. AMD probably couldn't keep up the way things were going, just patching Stars chips to keep up. Intel was going to keep patching Core (and they did), and they were going to maintain or extend their lead doing so.
Then there's the issue of the 45nm-32nm transition and Llano. The die-shrinking process was not a stellar success for that chip, though admittedly, that happened well after Bulldozer was in the development pipeline.
In any case, they tried something radically different to gain ground, even if it meant making some compromises.
In retrospect, it would have worked out much better if the design hadn't launched with as many bugs as it had, and if they had supported FMA3 and AVX exclusively from the beginning (rather than FMA4 and xOP). There's also the issue of how splitting AVX instructions affected performance on Bulldozer, though now we're delving into some of the more-arcane aspect of the processor, and I am not really able to render the issue accurately.
Some of that stuff is fixed in Piledriver, and even more of it is fixed in Steamroller, which makes you wonder what they would have needed to do from the start to avoid all those mistakes in the first place.
There's a lot of coulda woulda shoulda in there.
Today we have the benefit of hindsight. Bulldozer had some flat-out design mistakes in it. It appears to the semi-casual observer that some die shrinks and uarch updates could have gotten Stars much closer to Haswell-like performance by 2014/2015 than anything we see from AMD now. That seems plausible, but it also assumes that Intel would have rested on their desktop laurels to the extent that they have over the last 1-2 years. Broadwell/14nm might have turned out quite a bit differently had AMD presented a more credible threat.
Facebook
Twitter