There's one aspect to all this that I really don't think has been adequately explained. Nvidia is trying to pass this off as an accident, based on a breakdown in communication and in their writeups on the matter both Anandtech and Toms seem to accept Nvidia's explanation.
Anandtech ->
http://www.anandtech.com/show/8935/geforce-gtx-970-correcting-the-specs-exploring-memory-allocation
Tom's ->
http://www.tomshardware.com/news/nvidia-geforce-gtx-970-specifications,28464.html
My issue with this explanation comes from the fact that this "feature" of the 970 is entirely new and actually pretty complex. In fact, from a purely technical point of view, it's a pretty impressive accomplishment from Nvidia and a really efficient way of harvesting failed 980 chips with minimal compromise (but not NO compromise!)... again, I just said, "from a purely technical point of view". From Anandtech's deep dive, it involves reasonably complex hardware and software design, and a huge amount of firmware tweaking to try and manage a two-segment memory system by keeping high-frequency data in the faster memory segment. Presumable the whole new system would have required a huge amount of testing and optimising.
In other words, there were lots of lots of people spending lots and lots of time developing a new and pretty innovative memory management system.
If that's all true then, how is it possible that not one single person thought to discuss this brand new, innovative feature with their marketing department for the first ever card to launch with the fruits of their labour? Sure, if this was recycled tech or a minor change, I could see that happening... but significant debut... no one said anything? AND, did not one of those techs/devs/testers read the launch articles to see how reviewers responded to their feature, which many would surely have realised sounds pretty bad if it's not explained properly?
That I find extremely hard to believe. While I agree that the benchmarks are 'true' and some of this hype is overblown, I just can't buy Nvidia's explanation even if the editors of Toms and Anandtech can. Happy to discuss this further (sensibly) if others see things differently.