Originally Posted by Majin SSJ Eric
I agree but of course if yields aren't as good this time or if the performance of Maxwell isn't as impressive as Kepler's was they may be forced to release a large-die flagship like they did with Fermi. Judging by the 750Ti I absolutely don't see that happening though.
Say what now? The 750ti is a 60W TDP card
designed primarily for mobile - indeed GM107 is now appearing in various ultrabooks. If that performance is scaled with CUDA cores for a 250W TDP part, it will be beastly regardless of 28nm or 20nm. As far as what happens, I have no idea. Keep in mind, though, that efficiency of low end parts (which is what Maxwell and kepler were designed for first and foremost) has a direct effect on the scaling of high end parts. I don't know if Maxwell will be feasible for 28nm, but if it remains 28nm it will still far outperform GK110. The only question is whether 28nm has the required transistor density for a big die Maxwell. That I don't know. It could require 20nm.
The GK107 was the first kepler chip to debut and it appeared in ultrabooks well before GK104 hit the streets. That part was of a higher TDP and performed WAY LOWER than the GM107, and if you extrapolate how well GK110 performs based on GK107, and then extrapolate a 60W TDP GM107 scaled with more CUDA cores for a 250-300W TDP GPU.....the Maxwell will obviously perform way better once scaled upwards. GM107 is double the performance per watt of the GK107. That same trend should apply to high end maxwell GPUs - a 250W Maxwell would perform well above an outgoing kepler once Maxwell is scaled up with thousands of CUDA cores. Again, the GM107 is a 60W TDP part. The 60W TDP Kepler performs at *half* the speed of GM107.Edited by xoleras - 3/9/14 at 4:47am