Originally Posted by scotthulbs
Again if you look at the Valley 1.0 thread you will see the fastest 660 ti's, 670' and 680's Have huge Memory overclocks a couple of the 680 Lightning guys have their memory pushed up around 8,000 mhz Effective. That brings their memory bandwidth from 192.2 Gb/s up over 240 Gb/s. I think its safe to say that memory bandwidth is holding these cards back. It may be only holding them back slightly or perhaps quite a bit. Unfortunately we likely will never know.
Again I am just wondering how these cards would perform with more memory bandwidth. Cards with huge overclocked memory seem to have a healthy %20 performance increase over cards running stock memory. I think is is possible that the next gen 700 series could simply increase the memory bus and bandwidth to achieve the roughly %30 performance increase most are expecting. They could up the base core clocks and boost clocks at least %10 also.
I think GK104 is more powerful than it is in its current form, its memory needs to be running on a bigger highway to get up to it max speed.
Yes, the bandwidth for those cards falls extremely far below the curve of power / bandwidth. What I'm telling you is the titan / 680 isn't really that far under the curve. I've noted some gains on my 670 with memory OC and some smaller gains on my titan with memory OC, but overall the higher end cards are closer to the sweet spot of power to bandwidth ratio than the lower end cards you are talking about.
The lightnings you mention aren't really to be compared to the rest of the series, as they have much beefier VRMs and the core clocks are also likely to be in the 1400mhz range using extreme cooling to reach those speeds. You need to be careful to look at how the card is cooled and whether its a 24/7 gaming clock vs a suicide DICE / LN2 run for max score in benchmarks. The standard cards with standard cooling simply do not utilize the higher bandwidth you are talking about.
GK104 being bandwidth limited is not something we don't know, it was the midrange chip that did so good against GCN that they didn't need to release GK100 as the 680 / 670. Arguments can be made to say GK100 production process wasn't ready with yields so they clocked up the GK104 instead. Regardless, as stated before the chip runs out of power, and capacity more often than bandwidth for real use scenarios.