FX-8120 and FX-8150. Disable one core per module, eliminating the sharing of resources and allowing each core to then use all of the resources within a module, a roughly up to 20-30% increase in performance per Hz per core and a roughly 30-40% drop in power consumption. Overclock the CPU frequency and the CPU/NB frequency (this increases the L3 cache frequency, among other things; letting it default at 2.2GHz regardless of the CPU frequency seems to be a bad move from AMD IMO) and they can compete with the LGA 1155 i5s (even the LGA 1155 i5 K editions when they're given a very high overclock) surprisingly well. AMD competes a lot better than the hype says when you know how to use the CPU. Unfortunately, AMD didn't make such things more default nor even the easiest thing to do if you don't already know how to do it, so it's no surprise that most people don't realize this.
AMD doesn't compete perfectly in power efficiency when this is done, but it does sky-rocket compared to stock FX. This same trick works well on the six-core FX CPUs, but since it halves the active core count, it probably wouldn't work well with the quad-core FX models. The i3s and the i5s have nothing against AMD in performance for the money when this is taken into consideration (especially the i3s versus the six-core FX CPUs), granted Intel would still have the efficiency win. It would be even greater if Intel had Ivy Bridge i3s. Now if there was a K edition i3, then this would be different. Heck, giving the i3s Turbo would at least make a difference, but Intel seems intent on not letting them overclock at all except a minor amount through the BLCK, so oh well.
I can't think of a trick to let FX compete with the i7s directly in performance, especially not the LGA 2011 i7s, but still, this is a pretty significant achievement for AMD. Well, it would be if they would recognize this... Heck, their next Phenom-branded line could be the eight and six core CPUs that are designed like this by default while FX is delegated to the six-eight core models without one core per module disabled. It would be a better setup than having to do this manually.