Power is but one factor that limits fast clocks?ninelven said:
I would say that not only is power a limitation, but also clock scaling. Not all architectures scale perfectly (or even at all) with increased clocks.
The quickest example I can think of is the GF104, which had negligible increase in games (although I'd presume shader based and other applications were helped more) despite a generous increase in clocking.
There can also be speed paths in some part of the chip that will not work reliably at those speeds, or would need impractical cooling solutions and damaging voltage bumps to reach it.
It would seem to be a valid design tradeoff to improve per-clock capability or manufacturability in return for the the inability to scale to a clock no practical implementation would reach anyway.
It would seem to be a valid design tradeoff to improve per-clock capability or manufacturability in return for the the inability to scale to a clock no practical implementation would reach anyway.
It would also allow for better yields would it not? Overclockability isn't a surefire method and scaling isn't 1:1 nor is it parallel. Unless you can get parity for overclocking on every overclockable part.
What's also apparent in this case is that not all programs scale well with overclocks/higher frequencies at all. From a manufacturing persepctive, it's much more economical if not logical to make a wafer where it all can reach a certain clock rather than aim for the fastest clock for one die.
Final Destiny
Newcomer
Anyone recall hearing about a quad-GTX680?
Saw it mentioned here:
http://vr-zone.com/articles/post-co...deon-quad-gpu-attempts-for-kepler-/16239.html
Saw it mentioned here:
http://vr-zone.com/articles/post-co...deon-quad-gpu-attempts-for-kepler-/16239.html
what do you think guys? http://www.kitguru.net/components/g...-gtx570-overstock-delays-gtx660-until-august/
That's probably not the main reasoning.
GK106 has been MIA for the better part of the quarter and none of the leaks seem to have even mentioned it. That's why the 670M is a Fermi and the 660M is a GK108 on an unhealthy dose of 'roids
The 660Ti might be using a cut-off GK104 if further cost-cutting permits (geez, look at the 670 already), but anything lower than that is probably not a feasible endeavour now at 28nm pricing.
Plus, the 570 has been selling so well for a supposedly highend card that it's hard to believe that there's overstock (it's supposed to have wound down in production after the 7870 release)
UniversalTruth
Veteran
There is no such thing as "overstock". Only wrong pricing.
Price reduction is the ultimate creator of demand. It is so efficient that no feature set can be even close in efficiency compared to fair pricing.
Price reduction is the ultimate creator of demand. It is so efficient that no feature set can be even close in efficiency compared to fair pricing.
That's probably not the main reasoning.
GK106 has been MIA for the better part of the quarter and none of the leaks seem to have even mentioned it. That's why the 670M is a Fermi and the 660M is a GK108 on an unhealthy dose of 'roids
The 660Ti might be using a cut-off GK104 if further cost-cutting permits (geez, look at the 670 already), but anything lower than that is probably not a feasible endeavour now at 28nm pricing.
Plus, the 570 has been selling so well for a supposedly highend card that it's hard to believe that there's overstock (it's supposed to have wound down in production after the 7870 release)
ok, interesting points of view, thank you
One thing are numbers on paper, another one is real power consumption. Many reviews show 15-20W difference between GTX 680 a HD 7970.
1. yields / manufacturing capacity
2. GTX 680 already runs ~150 MHz higher than HD 7970
But it also has a higher latency for the basic arithmetic operations. That probably means a longer pipeline which can be clocked higher all other things equal. Although, Kepler still lacks result forwarding which means register access is part of the RAW latency (in AMD's GPUs it is obviously not), but that does not explain all of the difference (4 versus ~10 cycles, without register access I would guess 4 vs. 6 as the number of pipeline stages, but can't know for sure of course).
UniversalTruth
Veteran
A1xLLcqAgt0qc2RyMz0y
Veteran
WTH?
He is just saying that GTX680 yield is bad and there is hard evidence all over the world.
The only thing hard here is what you have for AMD.
It comes in at 20 watts under a 7970. And if you want to talk factory overclocked all I need to say is 7970 Ghz Edition.
Back at you.
With you as a spokesman for AMD I can see why Apple left AMD.Shame on fruiti.co. Wrong decision, should have gone with AMD chips.
http://micgadget.com/21980/apple-is-close-to-finally-updating-the-mac-pro/
No one sane would use inferior AMD products after being burnt like that. And the fact Kepler swept up almost all of the Ivy Bridge notebook design wins is quite telling.
No one sane would use inferior AMD products after being burnt like that. And the fact Kepler swept up almost all of the Ivy Bridge notebook design wins is quite telling.
Bouncing Zabaglione Bros.
Legend
Apple have always been a company that plays its suppliers off against each other. IIRC, they moved to AMD after Nvidia's disastrous bumpgate period. There was lots of talk then about how AMD swept up all the design wins and how Apple was never going to back to Nvidia.... until the next time...
Replacing bump material should be a lot easier than rewriting graphics driver for new production batch.
But not for the batches already built into the products (not only) Apples's customers had already in their hands when the problem surfaced.
Similar threads
