|
Sirius87 said:
This figure is a few years old, but it might still be accurate to some point. And it quite nicely shows that Nvidia basically expects the same thing as you. But they also still expect the lowering of production costs per transistor by around 10% per year. And smaller nodes also allow higher clock frequencies, which would further increase their computational power per dollar. I won't expect big jumps but cost savings of around 20% per year still seem likely. My proposed console on page 3 of this thread was based on the assumption of a reduction of manufacturing costs for the chip components by this 20% per year and the availability of the 16nm node in the middle of 2017, reducing the requirements for cooling of the system. |
As it says in that slide ... Transitioning to 20nm or 14nm gives you barely any cost saving and 10% is not tiny but it's marginal plus those are their best case projections judging from the dotted lines. The guys that you actually want to hear from are the semiconductor foundries themselves such as TSMC or Samsung and both of them project no cost savings. As far as increasing clocks, I remain skeptical about that since the gate dielectrics have already reached their theoretical limits. Cost savings of 20% each year sounds insane for the most part if there's absolutely nothing to drive those costs down such as EUV ...
FYI TSMC's 16nm is essentially a redone 20nm process with FinFETs.
