3 ms·

But for the cost, couldn't (for instance) a 16nm fab produce 90nm chips? If not, how complex would it be to "port" an existing 90nm chip, to produce a 16nm rev

But for the cost, couldn't (for instance) a 16nm fab produce 90nm chips? If not, how complex would it be to "port" an existing 90nm chip, to produce a 16nm revision? Impossible, or quite easy but not cost effective? From the POV of the car companies, could such revisions be used directly, or would they need to be qualified in the same fashion as new parts?

coryrc 5y ago
No and about as expensive as having designed it in the first place. Given design costs are the largest portion of per-unit costs it's not a good investment.
- pedrocr 5y ago
  Could you explain why not? Naively, having a much more dense process should allow automatic conversion. Going from 90nm to 16nm is 10x the density, that's a lot of margin for an automatic tool to use. Why doesn't that work?
  - coryrc 5y ago
    It's more than just a size change and there are features besides transistors. Say you have a 100 fF capacitor; is that because that's the right value based on an external constraint (say, interacting with a crystal) or because it's matching the inductance of a long internal path? Because you adjust them differently based on circumstance. And your transistors with a specific load must still supply the same current as before so they can't shrink as much as ones for internal logic. Also because the materials have changed the dialectic constant probably has and now the relative sizes of components need to adjust. And circuits are designed to minimize switching loss based on the old switching time and now they'll be wrong. Oh, and the breakdown voltage of the new process can't handle the voltage many of these old circuits use IIRC.
VLM 5y ago
If you're bored you can go to mycmp.fr and check their process catalog and compare a typical 55nm run to a 160nm run. Its not quite as standardized as ordering PCBs over the internet. Different metallization (you usually don't get to choose Al or Cu) will have different resistances. Generally the smaller processes will be faster so a design with no race conditions or metastability problems on 160 might not run reliably or at all on 55. Some processes are analog oriented so they'll guarantee up to 60 volts or more, if you're trying to design power devices, other processes are logic oriented and they might have a standard voltage of 2.5, 1.1 etc. You can design something that'll run on a 55nm process and something that'll give similar performance on a 160nm process BUT they'll be different designs. I think the closest analogy would be changing processes is like changing manufacturing material. You can make a car piston out of steel or aluminum but you can almost never just swap materials in an existing assembly line.

coryrc 5y ago

No and about as expensive as having designed it in the first place. Given design costs are the largest portion of per-unit costs it's not a good investment.

pedrocr 5y ago
Could you explain why not? Naively, having a much more dense process should allow automatic conversion. Going from 90nm to 16nm is 10x the density, that's a lot of margin for an automatic tool to use. Why doesn't that work?
- coryrc 5y ago
  It's more than just a size change and there are features besides transistors. Say you have a 100 fF capacitor; is that because that's the right value based on an external constraint (say, interacting with a crystal) or because it's matching the inductance of a long internal path? Because you adjust them differently based on circumstance. And your transistors with a specific load must still supply the same current as before so they can't shrink as much as ones for internal logic. Also because the materials have changed the dialectic constant probably has and now the relative sizes of components need to adjust. And circuits are designed to minimize switching loss based on the old switching time and now they'll be wrong. Oh, and the breakdown voltage of the new process can't handle the voltage many of these old circuits use IIRC.

VLM 5y ago

If you're bored you can go to mycmp.fr and check their process catalog and compare a typical 55nm run to a 160nm run. Its not quite as standardized as ordering PCBs over the internet. Different metallization (you usually don't get to choose Al or Cu) will have different resistances. Generally the smaller processes will be faster so a design with no race conditions or metastability problems on 160 might not run reliably or at all on 55. Some processes are analog oriented so they'll guarantee up to 60 volts or more, if you're trying to design power devices, other processes are logic oriented and they might have a standard voltage of 2.5, 1.1 etc. You can design something that'll run on a 55nm process and something that'll give similar performance on a 160nm process BUT they'll be different designs. I think the closest analogy would be changing processes is like changing manufacturing material. You can make a car piston out of steel or aluminum but you can almost never just swap materials in an existing assembly line.