I've been reading up a lot about silicon fab processes, yields, and technologies. Leading into all sorts of scholarly/engineering analysis of the Dark Silicon problem. There's tons of info out there, it's becoming an entire discipline in itself, but internet resources are always stonewalled by proprietary secrets and NSA barriers (silicon yield figures, and even the most trivial approaches which can increase them, are highly sensitive topics which apparently equate to corporate/industrial megaprofits and military/aerospace superiority).
We all know that a complex part is designed. And a batch is produced. And not every particular piece in any given production run will meet 100% spec, some parts will have defective cores or defective cache or defective functions or electrical issues or can only operate at slower speeds or whatever. So the parts are binned, there is the "perfect" top-end part, there are a variety of "imperfect" lesser parts which have fewer cores, less cache, slower speed, etc. Some parts are outright defective, nonfunctional, or simply fail to perform at minimum thresholds. The proportion of perfect vs imperfect vs defective parts is referred to as the yield.
The actual binning tier is somewhat arbitary. For example, Intel's latest Haswell-E parts come in 5960X (3.0-3.5GHz/8-core/20MB/40PCIe), 5930K (3.5-3.7GHz/6-core/15MB/40PCIe), and 5920K (3.3-3.6GHz/6-core/15MB/28PCIe*). So it's possible that a particular 5920K might, for example, happen to have some combination of up to 3.5/3.7GHz, 6-8 working cores, 15-20MB cache, or 28-40PCIe*, just not enough to qualify for a proper 5930K or 5960X rating. Sometimes (very rarely at first, but more and more often as a new cutting-edge lithography process matures) a company like Intel will have very high yields and be in a position where it competes with itself, it will actually bin higher-rated parts as lower-rated parts so that it can still supply low- and mid-end pricing tiers. Thus we have happy overclockers who sometimes luck out and get a lesser part to perform as well as (or even outperform) a superior part.
(* I am aware that the 5920K is a deliberately feature-crippled part locked down to 28PCIe lanes, basically just included to round off Intel's initial Haswell-E offerings with a low-priced entry. But I'm ignoring that here since I only used the Haswell-E group example because other chip families include too many members to list without more wall of text.)
I've observed similar patterns with earlier Intel chips, AMD's Tahiti chips, and NVidia's Kepler chips. And have not found any answer for the question I originally asked:
Why does it seem like parts with fewer active cores (and/or less integrated cache/functionality) can operate at higher clock specs? They appear to have the same electrical parameters and same TDP spec. The same die with the same transistor count and density (although some entire "bad" blocks of transistors are deactivated). To me this means they have the same "power budget" - meaning that while they should work at higher speeds given the same power level, the better (slower, more complex, more functional) parts should be comparable when the "power budget" is increased. Yet, this doesn't seem to consistently be the case. Why can the simpler parts work faster, since they're basically just underspec complex parts?
[Edit]
Is the answer just based on the manufacturer's arbitrary product structure? In my above example, is it just that Intel discards 5920K and 5930K parts which fail to perform at higher-than-5960X speeds so that the consumer faces a more "balanced" purchasing choice? Or perhaps that yields of faster 5960X parts are too low, so they've lowered the threshold? Such draconian marketing, destroying some of your own good products just to control pricing tiers, has been observed with Intel before, especially when they dominate a niche. And, no doubt, Intel can repurpose semi-functional Haswell-E dies (or their components) in some other product line.