I haven't been game to push the memory harder as there is no heatsinks or active cooling on them. :P

All I can say is that memory overclocking in the 7000 series makes a massive difference.  I have found that if you average the +Core and +Mem, it equals your gain in performance.  On top of that, overclocked memory uses FAR less extra power and creates far less heat than overclocking the core.

Case and point:  My 1165/1805 overclock beats my 1205/1650 overclock and uses 30w less of power.  Look into it... ;) 

All I can say is that memory overclocking in the 7000 series makes a massive difference.  I have found that if you average the +Core and +Mem, it equals your gain in performance.  On top of that, overclocked memory uses FAR less extra power and creates far less heat than overclocking the core.

Case and point:  My 1165/1805 overclock beats my 1205/1650 overclock and uses 30w less of power.  Look into it... ;) 

Is the GCN architecture really that bandwidth starved ? I thought that wasn't too much of an issue seeing as how they already have a massive amount of bandwidth. :P

Call it what you want!  I see similar results from most Nvidia cards.  Also keep in mind this "Starved" card nets me a 30%+ performance boost over stock.  Yes a $300 7970 can compete with a $650 780...

Huh, I guess you can say that most recent cards performance is limited by the memory performance rather than raw shading power.

Yeah I just checked my GTX 765m (A GTX 650 Ti), and when I overclock my core by 13% and memory by 26%, I get an average of a ~20%+ performance boost.  Granted the 650 Ti is probably one of the most bandwidth starved cards out there, but it just goes to show that the average between the core and mem is genarally your performance boost overall.  That is why I can never understand why so many overclocking reviews ignore the memory half the time...

It would be nice if someone tested this "Theory" on a 500 or 6000 series or older card to see if it holds true for older architectures... 

It would be nice if someone tested this "Theory" on a 500 or 6000 series or older card to see if it holds true for older architectures... 

I have a pair of 6950's unlocked in my other machine.

Some games I might see a 5% performance boost from *only* unlocking the shaders and maybe 1-2% increase with a 10% memory overclock.
Some games you get roughly a 1-2% increase from unlocking the shaders but almost a linear increase in performance with the memory clock. (Crysis being a prime example.)

When I get back from my trip, I can run some hard numbers in a week if you want, because those cards will be sold/donated/used as paper weights soon.