3 ms·

And that's even assuming the new fabs will produce the chips currently out of stock. The current exceptional shortages are mostly in legacy nodes, whereas newl

by crote 4y ago

And that's even assuming the new fabs will produce the chips currently out of stock. The current exceptional shortages are mostly in legacy nodes, whereas newly constructed fabs will mainly be producing on cutting-edge nodes. Building new fabs for legacy nodes simply isn't economically viable.

Cthulhu_ 4y ago
Wouldn't that mean the old fabs or their equipment will go down the line to make older chips? I figured that's how it goes with these things.
- magicalhippo 4y ago
  The way I've understood it, the expensive part of producing a chip is making the mask set, and the mask set is intimately related to the process producing the chip. That is, you typically can't take a mask set used at one foundry and start producing at another foundry without further ado. In some cases one can do it[1], but I imagine it requires a fair bit of money to port the mask sets and qualify the new parts. A specific process at a foundry might be sufficiently unique that there's few if any alternative foundries, I believe this was the case for AKM's factory which burned down[2]. If an old part still sells it's often high-volume "jelly bean" stuff, in which case it makes economic sense to let the old machines run as long as they can as they're written off entirely, but not to invest a lot in upgrades. If it's low-volume niche stuff, the price per unit is high, but that means any upgrade cost would drive up prices a lot due to low volume. Thus any customers still hanging on to some old design requiring that part might consider such a price hike a good time to redesign their circuit instead. [1]: https://www.latticesemi.com/-/media/LatticeSemi/Documents/ProductChangeNotification/PCN12/PCN01A12NotificationofIntenttoUtilizeanAlternateQualifiedMaskSetandFoundryfortheispPACPOWRPowerManagerIIProducts.ashx?document_id=44281 https://www.latticesemi.com/-/media/LatticeSemi/Documents/Pr... [2]: https://www.diyaudio.com/community/threads/akm-factory-fire-will-have-a-terrible-impact-on-the-audio-indusrty.362358/#post-6395374 https://www.diyaudio.com/community/threads/akm-factory-fire-...
- crote 4y ago
  Moving to a different node (or even a different fab of the same node) requires significant engineering effort. Meanwhile, we are talking about chips which cost less than a dollar. Heck, some chips even cost way less. For example, Diodes Incorporated's 74LVC1G06FZ4-7 currently sells for $0.016 in quantities of 10.000. It's an absolutely trivial component logic-wise, so there is nothing to gain by shrinking the die. In fact, due to the current capabilities and ESD tolerance shrinking it would make it worse. Chips like that are never going to move to a different node - they will be produced until the fab breaks down or becomes unprofitable, at which point the chip is discontinued.
phpisthebest 4y ago
This is false, most of the TSMC and other fabs under construction right now will not be 5nm or lower nodes. They are legacy nodes
- riskable 4y ago
  This is mixing apples & oranges: The new fabs aren't the latest & greatest tech (e.g. 2-5nm) but they're still going to be producing 25-40nm chips using 300mm wafers which is smaller than most automotive-using chips which are 40-90nm (with the majority still being 90nm). The real problem though is that 300mm wafer size: A 40nm chip that was made for 200mm wafers still needs to be redesigned for 300mm wafers. It seems like that wouldn't be necessary--you'd think it would just mean you could cram more chips into a bigger space--but it's not so simple. It's not as difficult as designing for a completely new process (e.g. 5nm) but it's still a non-trivial change.
- crote 4y ago
  We are seeing shortages of chips initially taped out around 2007/2008 - think the STM32 family or the ATmega32U4. They are using technology from the early 2000s, think 65-90nm. For comparison the RP2040, released in 2020, uses the 40nm node. That's technology which is literally a decade old and it is considered a modern chip! Meanwhile, TSMC has publicly stated that they are not going to build more 40nm fabs and is actively trying to move customers to 28nm or newer[0]. There isn't any profit in building 40nm fabs anymore, let alone even more legacy ones. [0]: https://www.extremetech.com/computing/337567-tsmc-tells-its-customers-to-get-off-older-nodes-move-to-28nm https://www.extremetech.com/computing/337567-tsmc-tells-its-...
  - phpisthebest 4y ago
    It is not really profitable to build FAB's in the US at all with the regulations, labor problems, and 100 other things CHIPs act and other government initiatives changes the profitability calculation since the government is paying a large part capital costs

by crote 4y ago

Cthulhu_ 4y ago

Wouldn't that mean the old fabs or their equipment will go down the line to make older chips? I figured that's how it goes with these things.

magicalhippo 4y ago
The way I've understood it, the expensive part of producing a chip is making the mask set, and the mask set is intimately related to the process producing the chip. That is, you typically can't take a mask set used at one foundry and start producing at another foundry without further ado. In some cases one can do it[1], but I imagine it requires a fair bit of money to port the mask sets and qualify the new parts. A specific process at a foundry might be sufficiently unique that there's few if any alternative foundries, I believe this was the case for AKM's factory which burned down[2]. If an old part still sells it's often high-volume "jelly bean" stuff, in which case it makes economic sense to let the old machines run as long as they can as they're written off entirely, but not to invest a lot in upgrades. If it's low-volume niche stuff, the price per unit is high, but that means any upgrade cost would drive up prices a lot due to low volume. Thus any customers still hanging on to some old design requiring that part might consider such a price hike a good time to redesign their circuit instead. [1]: https://www.latticesemi.com/-/media/LatticeSemi/Documents/ProductChangeNotification/PCN12/PCN01A12NotificationofIntenttoUtilizeanAlternateQualifiedMaskSetandFoundryfortheispPACPOWRPowerManagerIIProducts.ashx?document_id=44281 https://www.latticesemi.com/-/media/LatticeSemi/Documents/Pr... [2]: https://www.diyaudio.com/community/threads/akm-factory-fire-will-have-a-terrible-impact-on-the-audio-indusrty.362358/#post-6395374 https://www.diyaudio.com/community/threads/akm-factory-fire-...
crote 4y ago
Moving to a different node (or even a different fab of the same node) requires significant engineering effort. Meanwhile, we are talking about chips which cost less than a dollar. Heck, some chips even cost way less. For example, Diodes Incorporated's 74LVC1G06FZ4-7 currently sells for $0.016 in quantities of 10.000. It's an absolutely trivial component logic-wise, so there is nothing to gain by shrinking the die. In fact, due to the current capabilities and ESD tolerance shrinking it would make it worse. Chips like that are never going to move to a different node - they will be produced until the fab breaks down or becomes unprofitable, at which point the chip is discontinued.

phpisthebest 4y ago

This is false, most of the TSMC and other fabs under construction right now will not be 5nm or lower nodes. They are legacy nodes

riskable 4y ago
This is mixing apples & oranges: The new fabs aren't the latest & greatest tech (e.g. 2-5nm) but they're still going to be producing 25-40nm chips using 300mm wafers which is smaller than most automotive-using chips which are 40-90nm (with the majority still being 90nm). The real problem though is that 300mm wafer size: A 40nm chip that was made for 200mm wafers still needs to be redesigned for 300mm wafers. It seems like that wouldn't be necessary--you'd think it would just mean you could cram more chips into a bigger space--but it's not so simple. It's not as difficult as designing for a completely new process (e.g. 5nm) but it's still a non-trivial change.
crote 4y ago
We are seeing shortages of chips initially taped out around 2007/2008 - think the STM32 family or the ATmega32U4. They are using technology from the early 2000s, think 65-90nm. For comparison the RP2040, released in 2020, uses the 40nm node. That's technology which is literally a decade old and it is considered a modern chip! Meanwhile, TSMC has publicly stated that they are not going to build more 40nm fabs and is actively trying to move customers to 28nm or newer[0]. There isn't any profit in building 40nm fabs anymore, let alone even more legacy ones. [0]: https://www.extremetech.com/computing/337567-tsmc-tells-its-customers-to-get-off-older-nodes-move-to-28nm https://www.extremetech.com/computing/337567-tsmc-tells-its-...
- phpisthebest 4y ago
  It is not really profitable to build FAB's in the US at all with the regulations, labor problems, and 100 other things CHIPs act and other government initiatives changes the profitability calculation since the government is paying a large part capital costs