@RS we all know that underdog doesn't need to screw-up big time to find themselves in trouble, especially when market-leader executes well.
Tahiti is more overweight than Kepler, but it's very lean compared to NVIDIA total-GPGPU attempt i.e. Fermie
I am actually amazed that AMD did pack all their overhead in only 350mm2.
I said it before and I'll say it again, I think Fermi was underrated. Take Fermi GTX580, apply 28nm shrink to it and it'll be a wonderful card. Obviously I have no way to prove it but Fermi as an architecture was excellent. It didn't have many weaknesses other than TMUs at high resolutions. It's main liability to me was that NV went all on GPGPU + gaming in on
40nm node.
If you think about it Kepler really is just Fermi enhanced. NV didn't do much to revise Fermi with Kepler. They chopped off shader clock, gutted GPGPU/dynammic scheduler and applied 28nm shrink to control leakage and thus power consumption. That 1.4-1.6Ghz shader clock on Fermi was a liability (as Pentium 4 taught us that high frequency is the enemy of power consumption). Then all they did is compensate for the clock speeds with many shader cores, upped the TMUs because gutting GPGPU provided additional space for more functional GPU units. Kepler isn't some revolutionary beast, but just a rebalanced Fermi architecture into a streamlined gaming card and it happens to benefit from 28nm node shrink.
Global Foundries sites, "28nm transistors offer up to 60% higher performance than 40nm at comparable leakage with up to 50% lower energy per switch and 50% lower static power." ~
Source
I can't see TSMC being much worse. Can't discount the massive benefit a full node shrink provides. NV set themselves up nicely since with 294mm^2 die and experience building 500mm^2+ chips, they have ample room to add more functional units and performance. AMD has a monumental task with HD8000 series. They have already ballooned to 365mm^2. There are some theories of how AMD may
improve performance with HD8000 series (Use Google Chrome to translate) but from what I read they all point to even large die size. I am not an engineer but it sounds very difficult to contain HD8000's power consumption, while increasing die size and adding even more functional units. Maybe AMD is counting that 28nm node will mature and they'll be able to have a better stepping on which to build HD8000 series.
What's impressive is how NV made a 180* turnaround in just 2 years since Fermi. In 1 generation, they are now leading in performance/watt, performance/mm^2 of die space for gaming. Now it is AMD who somehow has to figure out a way to build a chip that works in compute and gaming. If NV decides to drop 500mm^2 gaming Kepler, AMD will have no shot unless they drop GPGPU. The question is will NV drop 2880-3072 SP GK110?
http://guru3d.com/news/gk110based-surfaces--has-2880-shader-processors/
GK 110 -- "Big Kepler" specs based on K20:
7.1 Billion Transistors
2880/3072 Shader processors :wub:
15/16 SMX (shader engines with 192 Shader-ALUs)
384-Bit wide GDDR5 memory bus
6-Pin & 8-Pin power connectors for Tesla K20-cards
3x higher double precision