The problem is that I don't think the Bobcat design can actually do anywhere close to 3GHz under any real conditions.
You're right, I did some cursory research and it seems the max OC I've seen online is about 2.4 ghz; all the design choices such as the fairly short pipeline would probably prevent it from being a speed demon in the future too, so max single threaded performance of any bobcat design is likely to lag behind BD rather significantly.
My main issue is that they didn't seem to use their die space in their current BD implementation very efficiently for the performance they're achieving, so I was toying with what AMD's alternatives had for this generation. For server / high throughput workloads, it might have been a good decision for AMD to stick lots of bobcat cores onto a die running at their stock 1.6 ghz, and they could fit in a lot:
http://www.chip-architect.com/news/AMD_Ontario_Bobcat_vs_Intel_Pineview_Atom.jpg
The 8mm^2 cpu die size at 40nm is very impressive! They could easily fit 64 cores onto a BD sized die at 32nm and have room for other odds and ends. (Not that it's a trivial thing to execute.) As for desktop performance, they probably would have gotten better performance by fabbing a die w/ 6-8 husky cores (< 300mm^2 at 32nm) and clocking it as high as possible.
BD in its current iteration seems to be an uncomfortable compromise that uses way too much die space for what it accomplishes. Many indications seem to point to process issues and much higher initial target clocks. There are cache and software optimization problems too but judging from AMD's future roadmap, most of the 10-15% yearly improvements will come from process refinements rather than architectural change; see paragraph 3:
http://www.tomshardware.com/reviews/fx-8150-zambezi-bulldozer-990fx,3043-9.html
In the the long run, k10.5 doesn't have much clock speed / threading headroom and bobcat doesn't have server features like virtualization built in, but the BD design in its current iteration still seems like it was released just a touch too early.
Edit: It seems like bobcat does come w/ a large part of hyper-v built in and there has been talk of bobcat cores in servers so maybe this is a route they will explore in the future. Another possibility for AMD's poor diespace usage w/ BD is their increased reliance on software synthesization in design. From a business standpoint, It might better for AMD to switch to a more synthesizable core design which allows them to push something through the door quickly and reap some revenues. Intel's vastly larger R&D budget allows them to take on some diminishing returns for much more polished result.
