Edit: Never mind, I didn't realize you guys were in some sort of academic debate over architecture. Nevertheless, my comment stands. You compare products at any given point in time. That is what has relevance in the marketplace to non-academics who buy cards.
There is no debate on that; I completely agree with your entire post #159. The discussion started when it was insinuated that you can compare the efficiency of
architectures and completely ignore the advantages that a node shrink brings. How do you conclude that GCN architecture is more efficient than Fermi architecture is when the two are made on completely different nodes? If you figure it out, here is a cookie (oops, we don't have such an emoticon) ^_^
Imho, it was, is, and will remain fair to compare chips built on different nodes against each other, if they are positioned against each other in the marketplace (similar price and purpose), at any given moment in time.
I agree with you that GTX580 is the current competitor to the HD7970 given their prices.
You should read the last 2 pages of this thread. That's not the what's being discussed. It was insinuated that GCN is more efficient than Fermi is. We can only hypothesize this to be true, but we cannot confirm it to be true.
You can ONLY conclude with certainty that:
1) GCN on 28nm node is more efficient than Fermi is on 40nm node, and
2) HD7970 is more efficient than GTX580 is (in performance/watt, performance/mm^2 of die space)
You can't conclude anything about the efficiency of Fermi vs. GCN architectures because the node shrink difference brings 3 major advantages. For more explanation, please see
Post #144.
For example, GTX580 = 3B transistors, HD7970 = 4.3B transistors. With 43% more transistors, HD7970 is only 25% faster on average. From here, one might be tempted to say that Fermi might actually be more efficient per transistor, but even that might be misleading since NV and AMD count transistors differently! (and plus would you even trust that after the Bulldozer transistor adjustment?). The transistor measurements themselves are just a best guestimate, not a physically verifiable characteristic. But even then, I still wouldn't conclude anything about the efficiency of Fermi vs. GCN architectures.
You can compare SKUs, chips, products, whatever you want across different nodes because the performance differences between them are both a function of the architectures and the nodes. But because the node itself impacts the efficiency (because it impacts power consumption, transistor switching speeds and transistor density), you now introduce 2 variables, which in turn blurs the causation of which of them brought the greater performance increase - the node or the architecture, or a combination of the two. You need to be able to isolate the variable you are comparing (in this case the architecture), while keeping everything else constant, if your goal is only to compare the efficiency of the architectures.
--->
The Scientific Method.
If you had GCN on 40nm and Fermi on 40nm or if you had GCN on 28nm and Fermi on 28nm, you would be able to say with certainty which
architecture was more efficient since you would completely isolate for the changing variable, the 28nm node, which allows for 60% more transistors at 50% lower power consumption.....
And if you really wanted to be scientific about it, you'd have to compare the same 28nm process (so you would need to make both chips at either TSMC factory or at GF, etc.). 28nm process at TSMC isn't the same as it is at GF.