Quote Originally Posted by zerazax View Post
I don't know if its necessarily Nvidia's design that as limitations, but that ATI's was designed from the start to be super-scalable.

They said that R600 would be the foundation for 3 generations of video cards, and while a lot of people said that R600 was a failed architecture, R700 definitely vindicated the design. R600's failures were more likely due to the fab process and leakage which killed any chance of higher clocked cores or the rumored original specs of 480SP's, rather than 320.

That being said, it is true that G92 and GT200 have all been heavily based on the G80 (G92 basically just a shrink) and Nvidia did hit a wall earlier on the scaling of its design
Dont think thats the problem. X2900XT was 80nm and something like 740Mhz. The mid-end X2600XT was a blazing 800Mhz.. a milestone nVidia has yet to conquer 3 years later.

AMD moved to 55nm early on, at same time as nVidia launched 8800GT on 65nm. Likewise, AMD first to 40nm. So, you'd think AMD would be the one with yield problems, right?

Lets compare last 3 gen product launches.
X2900XT/PRO - full chip, 600-740 clocks
HD3850/HD3870 - full chip, 668-825 clocks
HD4850/HD4870 - full chip, 625-850 clocks
HD5xxx... only one where not selling all full chips at launch.

nVidia? crippled chips gallore - 8800GTS, 8800GT/9800GT, GTX260.. even their mid-range where you'd think yields wouldn't be an issue.

Clearly, nVidia designs are either more susceptable to defects or tight schedules are pressuring them to cut corners... litterally.

Quote Originally Posted by saaya View Post
more importantly, note how fast clocks decline on nvidias 40nm gpus with added complexity! over a 100mhz drop for the most complex 40nm part so far, and its only a cut down G92... so a tweaked G92 in 40nm can only clock to 550mhz, but fermi which is 5+ times more complex will reach 650+mhz? (clock derived from nvidias flops numbers mentioned at the super computing event)
die shrink - > lower clocks??... this happened for each and all of AMD's Athlon64s. Recall those top of the line FX were 130nm. Couple years later the 90nm were offered at highest clockspeed while 65nm for mid/low end.

And then ofcourse Phenom1 but that was more of a power budget issue.

And power budget is clearly no prob for a 32sp 40nm nVidia chip... so how does AMD make 850Mhz 2B tran and nVidia only 600Mhz and shaders running barely as high as original 90nm GTX....?
1. They do it on purpose, so Fermi looks good compared to GT220 /GT240
2. Fermi delay will be announced and they'll ship a 40nm G92 in the meantime
3. Design not scalable. Could be a prob if even crippled cut-down Fermi can only run (extrapolating)... 400Mhz.
4. Simply inexperienced and incompetent engineers. Can't be the "process" since 4770 was getting great clocks early on before things were ironed out.
5. Management.

Hate to repeat it so many times, but nVidia Fermi is way late, their DX10.1 cards are a joke, huge die GT200s are sinking profits, and dont even have Intel/AMD chipset business to fall back on. If they dont get PERFECT Fermi out with whole lineup down to bottom, it could be not just NV30, but more like 3Dfx time