3 ms·

I always thought chips were dominated by fixed capital costs of fabs. If they were dominated by variable material cost of wafers, as the article seems to imply

by csense 5y ago

I always thought chips were dominated by fixed capital costs of fabs. If they were dominated by variable material cost of wafers, as the article seems to imply, it wouldn't make sense that we see 90nm microcontrollers that sell for $1 and a high-end 16nm PC CPU that sell for $1000. So the question is, what's the reason that 90nm microcontroller sells for $1? I'm trying, and failing, to figure out an economic model that explains the market dynamics we actually observe. If building a new 90nm fab costs $billions, almost as much as building a new 16nm fab, why does the 90nm microcontroller sell for 0.1% of the price of the 16nm Xeon? If building a new 90nm fab costs 0.1% as much as building a new 16nm fab, why can't existing chip companies, some startup or GM themselves spend $10's of millions building a fab that can unblock $100's of millions of product, and alleviate the shortage?

sbierwagen 5y ago
Chips made on modern processes are dominated by capital costs. Old chips are made on old foundries, which are fully depreciated and therefore have no capital costs. If you made a 90nm foundry today, then it probably couldn't sell those microcontrollers for less than $100 each, just like a modern CPU. >why can't existing chip companies, some startup or GM themselves spend $10's of millions building a fab that can unblock $100's of millions of product, and alleviate the shortage? They can't do it fast enough. Standing up a new foundry takes years under the best circumstances, and today you couldn't do it all, since all the tooling is sold out and deeply backordered. If GM could snap their fingers today and magic a cleanroom into existence, they'd still be waiting a hell of a long time to put tools in it. Secondly on the price question, microcontrollers just have way fewer gates than a desktop CPU. A dozen registers, a thousand bytes of RAM, a few kilobytes of flash. This makes them physically smaller, which lets you put more on a wafer, and makes the unit price cheaper. The total die area of the ATmega8 is just 7.9 square millimeters... at the 500nm node! https://zeptobars.com/en/read/atmel-atmega8 https://zeptobars.com/en/read/atmel-atmega8 This gets you 8,100 dies from a single 300mm wafer. (If there are any 300mm foundries at 500nm, which there probably aren't) Thirdly, what makes you think anyone could build a 90nm foundry at all? Take a look at a list of foundries: https://en.wikipedia.org/wiki/List_of_semiconductor_fabrication_plants https://en.wikipedia.org/wiki/List_of_semiconductor_fabricat... I don't see anything making 90nm after 2014. Which makes sense. Fabs don't make all their tools in-house, that's done by vendors. As an example of one tool, by one vendor, the NXE:3400B EUV stepper by ASML. Unit cost, $175 million: https://www.tomshardware.com/news/tsmc-euv-tools-order https://www.tomshardware.com/news/tsmc-euv-tools-order These are ultra-bespoke, super-low-volume machine tools. ASML makes a couple dozen or a hundred steppers of a given model, then shuts down production and starts upgrading to the next node. Is it even possible to buy a new 90nm stepper today? I'm sure they have all the documentation and could roll back to the previous generation easily enough, if they had to. (Which they don't! They are maxed out just supplying current-gen tools) But given all the voodoo in semiconductor lithography, people retiring etc, I bet that rolling back two decades would pretty much require starting from scratch.

by csense 5y ago

sbierwagen 5y ago

Chips made on modern processes are dominated by capital costs. Old chips are made on old foundries, which are fully depreciated and therefore have no capital costs. If you made a 90nm foundry today, then it probably couldn't sell those microcontrollers for less than $100 each, just like a modern CPU. >why can't existing chip companies, some startup or GM themselves spend $10's of millions building a fab that can unblock $100's of millions of product, and alleviate the shortage? They can't do it fast enough. Standing up a new foundry takes years under the best circumstances, and today you couldn't do it all, since all the tooling is sold out and deeply backordered. If GM could snap their fingers today and magic a cleanroom into existence, they'd still be waiting a hell of a long time to put tools in it. Secondly on the price question, microcontrollers just have way fewer gates than a desktop CPU. A dozen registers, a thousand bytes of RAM, a few kilobytes of flash. This makes them physically smaller, which lets you put more on a wafer, and makes the unit price cheaper. The total die area of the ATmega8 is just 7.9 square millimeters... at the 500nm node! https://zeptobars.com/en/read/atmel-atmega8 https://zeptobars.com/en/read/atmel-atmega8 This gets you 8,100 dies from a single 300mm wafer. (If there are any 300mm foundries at 500nm, which there probably aren't) Thirdly, what makes you think anyone could build a 90nm foundry at all? Take a look at a list of foundries: https://en.wikipedia.org/wiki/List_of_semiconductor_fabrication_plants https://en.wikipedia.org/wiki/List_of_semiconductor_fabricat... I don't see anything making 90nm after 2014. Which makes sense. Fabs don't make all their tools in-house, that's done by vendors. As an example of one tool, by one vendor, the NXE:3400B EUV stepper by ASML. Unit cost, $175 million: https://www.tomshardware.com/news/tsmc-euv-tools-order https://www.tomshardware.com/news/tsmc-euv-tools-order These are ultra-bespoke, super-low-volume machine tools. ASML makes a couple dozen or a hundred steppers of a given model, then shuts down production and starts upgrading to the next node. Is it even possible to buy a new 90nm stepper today? I'm sure they have all the documentation and could roll back to the previous generation easily enough, if they had to. (Which they don't! They are maxed out just supplying current-gen tools) But given all the voodoo in semiconductor lithography, people retiring etc, I bet that rolling back two decades would pretty much require starting from scratch.