3 ms·
> Note that no, the state of the art fabs cannot be re-outfitted to make legacy chips. The old equipment they were developed on is generally no longer available
by tzs 5y ago
> Note that no, the state of the art fabs cannot be re-outfitted to make legacy chips. The old equipment they were developed on is generally no longer available and the profit from making those inexpensive chips would not justify the capital spend on new equipment.
Interesting. I had assumed that the difference between state of the art fabs and older fabs was that state of the art fabs can make smaller things, and perhaps more geometrically complicated things, but that they don't have to. If they wanted to make a design that only uses bigger things and simpler geometry they could.
Where does this assumption break down?
- 11thEarlOfMar 5y agoState of the art machines are used to make state of the art high-value chips. FPGAs, GPUs, CPUs, RAM, SRAM As those chips run through their life-cycle, their value drops. After 15 years, the price the chips can sell for will drop 90% or more. If there is a spike in demand at that point, there is no ROI on purchasing brand new equipment. The demand spike is seen as temporary. Hence, the chip makers are telling the buyers to design out the old chips and upgrade to new. Many companies can do this. But some are constrained by industry regulation (aerospace) or industry practice (semiconductors). I've mentioned before the irony that companies that make semiconductor equipment are struggling to ship more tools because they can't get enough (of certain chips) to complete building them.
- nl 5y agoWhile this is mostly correct, it is worth noting that FPGAs are mostly 28nm and 20nm processes, and have just started shipping on 16nm[1]. That doesn't require state of the art. The FPGA market is comparatively small though so doesn't affect the market dynamics much. [1] https://www.xilinx.com/about/generation-ahead-16nm.html https://www.xilinx.com/about/generation-ahead-16nm.html
- belltaco 5y ago>Interesting. I had assumed that the difference between state of the art fabs and older fabs was that state of the art fabs can make smaller things, and perhaps more geometrically complicated things, but that they don't have to. If they wanted to make a design that only uses bigger things and simpler geometry they could. >Where does this assumption break down? Why would they want to? There is heavy and overbooked demand for the 'smaller and complicated things' that make more money, and add the extra cost of the design to make the 'bigger and smaller' things, and it's a no brainer.
- tzs 5y agoThey wouldn't want to. What I'm curious about is if they can without retooling or purchasing equipment. Say due to chip shortages affecting the US military the government invoked the Defense Production Act to require some company with only state of the art fabs to make an older chip. From what I've read, the way a layer on a chip is made is a coating is deposited on the surface that is sensitive to light at certain wavelengths. They then use light of wavelengths that the coating is sensitive to in order to transfer a pattern from a photomask to the surface, making some change to the coating at those places where the light is not blocked by the photomask. Then they can treat the surface to remove the coating at those places were it was changed by the light. They then can deposit thing onto or remove things from those parts of the surface not still protected by the coating. Then they can use something that can remove the coating. The result is a layer on the chip with things deposited or removed right where they wanted them to, as determined by the pattern on the photomask. Repeat for more layers. The major improvements from generation to generation are the ability to use shorter and shorter wavelengths of light for the photolithography which allows for transferring patterns that have smaller features, and maybe the ability to make thinner layers so you can make chips with more layers. I don't see anything in the photolithography process that requires using small feature size. If for some reason you wanted to build an older chip that only used much larger features, I'd expect that it would just be a matter of using the photomasks for that older chip, and setting the layer thickness to what the old design called for.
- showerst 5y agoI think the deal breaker is "without retooling". For one thing there's a lot of steps involved in photomask -> repeat. If your masks are all built for a smaller wafer size, or a different focal length setup or physical process diameter, then they literally won't fit in the die, or won't focus correctly. You'll run into similar issues all up and down the chain. If your doping and prep all assumes a smaller wafer or different geometry, you're out of luck either physically because things don't fit, or because your chemistry needs to be redone. Your various deposition, bake and cure steps will necessarily also assume a process geometry, because some of the processes diffuse out from the center, or at different angles relative to some source that's bombarding them. It's sort of akin to saying "well if we can build a vw bug they can build a model-T". Generations of equipment changes make things incompatible.
- RicoElectrico 5y agoCMOS devices scale in all 3 dimensions, including gate oxide. It is quite unlike, say, PCB fabrication, which is way more flexible with regards to layer stack-up.