3 ms·

Perhaps (re)bootstrapping the old nodes industry is exactly what's needed? So traditionally, the chip industry worked such that that the high-end products move

by jabl 5y ago

Perhaps (re)bootstrapping the old nodes industry is exactly what's needed? So traditionally, the chip industry worked such that that the high-end products moved first to a new process node, and less demanding equipment was fabbed on equipment a couple of generations old before that equipment was thrown on the scrap heap. Rinse and repeat, everybody was moving towards smaller and smaller nodes, just that some could trail the leaders by a few years. But it seems these days the world is different, in that the cost per transistor has actually been going up for the leading edge processes, and power consumption has remained roughly flat. So for anything that doesn't need the latest gen process, there's really no upside to moving. Depending on who you ask, it seems the cheapest cost per transistor is somewhere in the 20-40nm ballpark. If so, it would imply there would be a long-time demand for these process nodes, and there is a need for an industry to supply tooling for these nodes, once the existing crop of second-hand equipment (left over when the leading edge users moved on to smaller nodes) is in need of replacement.

KptMarchewa 5y ago
>cost per transistor has actually been going up for the leading edge processes Isn't this just because of overall chip shortage? We might see very different scenario if TSMC, Samsung and Intel overinvest now. Also, you could say that cost per transistor increased for older processes as well... >power consumption has remained roughly flat Any source for this? Anything I've saw contradict that.
- jabl 5y ago
  > Isn't this just because of overall chip shortage? AFAIU, no. The basic problem, again AFAIU and I'm not an expert on this topic so take with a grain of salt, is that while the marginal cost of each transistor might still be less than at the previous generation, the R&D and capital costs have grown so immense that once you amortize those costs over the stuff produced for the lifetime of that process generation you end up with a higher per transistor cost compared to the previous generation. > Any source for this? Anything I've saw contradict that. Just search the web for articles about the breakdown of Dennard scaling, plenty of stuff to read on the topic. Even wikipedia mentions it https://en.wikipedia.org/wiki/Dennard_scaling https://en.wikipedia.org/wiki/Dennard_scaling
- opwieurposiu 5y ago
  The cost per transistor is going up because the latest nodes require too many process steps. An older node might require 90 steps, the lastest nodes require 900 steps. Every step requires a "tool" and every step has a chance of yield loss. The transistor is getting harder to make faster then it is getting smaller. The new EUV steppers with soft x-ray wavelength can do in one shot what took 4 shots with 193nm ArF tools. And for each shot you have to apply, bake, develop inspect and strip the resist. This was supposed to help but the EUV steppers are so expensive it is kind of a wash price wise.
sidewndr46 5y ago
> Depending on who you ask, it seems the cheapest cost per transistor is somewhere in the 20-40nm ballpark. While correct, this is only true because of all the old infrastructure sitting out there that has been fully depreciated but still works fine. It used to be the latest and greatest thing and paid for itself back then. Nowadays it is just sitting on the asset sheet of a bunch of companies who are all going "if it still makes money, keep running it!". The paradox of this all is what you identified: "if so, it would imply there would be a long-time demand for these process nodes". No one is going to invest in developing old nodes because the ROI isn't there. The demand is, but there probably isn't much cost differential between building a nearly leading edge fab and a 60nm fab. If you could sell a turnkey 60nm fab right now that was delivered and setup quickly the demand would be insane. But no one can because fabs have never really been turnkey. When you factor in the cost to develop all of this, the capital you'd have tied up, etc. the ROI suddenly fails to compete with interest bearing bonds I'd guess.

by jabl 5y ago

KptMarchewa 5y ago

>cost per transistor has actually been going up for the leading edge processes Isn't this just because of overall chip shortage? We might see very different scenario if TSMC, Samsung and Intel overinvest now. Also, you could say that cost per transistor increased for older processes as well... >power consumption has remained roughly flat Any source for this? Anything I've saw contradict that.

jabl 5y ago
> Isn't this just because of overall chip shortage? AFAIU, no. The basic problem, again AFAIU and I'm not an expert on this topic so take with a grain of salt, is that while the marginal cost of each transistor might still be less than at the previous generation, the R&D and capital costs have grown so immense that once you amortize those costs over the stuff produced for the lifetime of that process generation you end up with a higher per transistor cost compared to the previous generation. > Any source for this? Anything I've saw contradict that. Just search the web for articles about the breakdown of Dennard scaling, plenty of stuff to read on the topic. Even wikipedia mentions it https://en.wikipedia.org/wiki/Dennard_scaling https://en.wikipedia.org/wiki/Dennard_scaling
opwieurposiu 5y ago
The cost per transistor is going up because the latest nodes require too many process steps. An older node might require 90 steps, the lastest nodes require 900 steps. Every step requires a "tool" and every step has a chance of yield loss. The transistor is getting harder to make faster then it is getting smaller. The new EUV steppers with soft x-ray wavelength can do in one shot what took 4 shots with 193nm ArF tools. And for each shot you have to apply, bake, develop inspect and strip the resist. This was supposed to help but the EUV steppers are so expensive it is kind of a wash price wise.

sidewndr46 5y ago

> Depending on who you ask, it seems the cheapest cost per transistor is somewhere in the 20-40nm ballpark. While correct, this is only true because of all the old infrastructure sitting out there that has been fully depreciated but still works fine. It used to be the latest and greatest thing and paid for itself back then. Nowadays it is just sitting on the asset sheet of a bunch of companies who are all going "if it still makes money, keep running it!". The paradox of this all is what you identified: "if so, it would imply there would be a long-time demand for these process nodes". No one is going to invest in developing old nodes because the ROI isn't there. The demand is, but there probably isn't much cost differential between building a nearly leading edge fab and a 60nm fab. If you could sell a turnkey 60nm fab right now that was delivered and setup quickly the demand would be insane. But no one can because fabs have never really been turnkey. When you factor in the cost to develop all of this, the capital you'd have tied up, etc. the ROI suddenly fails to compete with interest bearing bonds I'd guess.