Pretty much nothing on the chip will actually be 5nm. Some distance between things may be 5nm. Anyway...
You can’t really improve things much by getting more of the stuff on the chip to those smaller sizes. The size of a transistor (by size we mean the ration of width to length of its gate) is determined by what it has to drive. If one transistor outputs to three receivers, you need to size it up enough to handle that load. Once you’ve size it up, then you need to size up whoever drives THAT, etc. Practically speaking, you can’t do much about it unless you go to radical new architectures that don’t execute the kinds of instruction sets we are talking about.
5nm is not as small as we can go, so we will just keep getting smaller. Other tricks will be used along the way, like improved semiconductors (SOI, strained semiconductors, maybe III-V semiconductors, etc.), better low-K dielectrics to reduce wire load, diagonal wiring layers, new and improved gate structures to provide better leakage characteristics, etc. etc.
I knew I'd find you here.
Spot on about transistor and gate sizes.
I'd also like to add that the dominating factor is not transistor size it's wire and signal integrity.
Smaller wires have less current capacity and even if feature sizes shrink, the metal does not scale the same way.
For longer distances you may need fatter wires to reduce inductance and carry the current.
You may also need additional spacing due to signal issues. (aggressors)
When I'm doing development, I don't worry a lot about the gates. I now worry about the wires between those gates and flops. I look at the timing reports to see if I have heavily loaded nets that are going to be slow, or gates where the delay is much higher than expected.
Even if you make transistors faster, metal doesn't get faster in a shrink. Actually, metal might be slower. Yeah, I know the presumption is that with things closer in the shrink, we can ignore that metal is thinner/slower with lower current capacity.
I haven't see chips get smaller. I'm actually seeing more and more chips (not mobile) reach the reticle limit.
This isn't a free lunch. That's all I'm saying.
There are lots of challenges down in the single digits.
I'd like to see the performance of process monitors (ProcMon) and ring oscillators. Some skew lots with sample structures would be nice too.
[doublepost=1554763607][/doublepost]
I was designing chips when they said quarter micron was the practical floor. And 90nm. And etc. etc.
It’ll be fine.
I was there too; .6, .35 and .25 micron.
I remember them saying we would hit a brick wall a .1 micron.
I remember them saying 90nm, then 48nm and on and on.....
I've seen optical processes using the capillary properties of water along with the refraction properties to give better performance on the lenses to focus in the lithography. We now have EUV being developed for 2020 and beyond.
There are barriers and we keep jumping over them in different ways.
I guess I'm really showing how long I've been around this stuff and how much of a true geek I really am.