Die shrink question

[Megazero]

This could be a dumb question but I will ask anyway. I know that the GP104 chip of the GTX 1070 and 1080 is the first of the 16nm chip to come out, and that shrinking from 28nm to 16nm make the transistor smaller hence more transistor can be put on the same size chip, less power consumption and all that jazz. But... why do the chip size stay the same after all these time? for example if we stick with 16nm but increase the size of the chip by double it would yield the same performance gain as keeping the chip size with say a 8-10nm transistor right? So why do it stay the same?

[don_svetlio]

Cuase Nvidia want to sell the big chip for 1500$

[Celmor]

There are no advantage from just going with the die shrink and keeping the transistor number the same and therefore making the chip smaller other than so more chips fit on a wafer.

The increase in transistor numbers is what increases your ROPs, CUDA cores and whatnot, the die shrink is needed because if otherwise you would just increase the transistor number without a die shrink and therefore making the chipp "bigger" (in 2D only of-course) the pathways get longer, signals take longer, and multiple-bit wide buses get issues because not all of the bits arrive at the same time - to circumvent this effect on motherboards you see some paths on motherboards making "circles"/not taking the shortest path (see image), so all bits of the corresponding bus arrive at the sime time at the chip, not to mention that the chips get more expensive.

 

A problem with die shrinks that has to be dealt with every time is that the thermal density is getting higher. If you stay with the same transistor number but do a die shrink and therefore making the chip smaller and make no other adjustments, the chip will output the same heat but on a smaller surface area and therefore heat sinks have a harder time of displacing the heat.

I can recommend the The Tech Report Podcast with Scott Wasson and David Kanter if you wanna learn more about this sort of stuff, for example this episode analyzing the Skylake architecture.

[Naeaes]

There are electrical issues like doubling the distance from one side to the opposite doubles the resistance. 

 

But the real reason, I think, is the yield. The wafer of silicon that the chips are cut out of can only be mafe so big. So if you double the area of one chip, you halve the number of chips you get.

Simply double the price per chip? 

Then you'd double the value of each chip you have to throw out st the end when they turn out to be faulty.

Basically they're just optimizing the amount of money they can get out of the amount of silicon wafer they have.