"However, all of these leading edge process nodes are likely to be on 12-inch wafers."
Not "likely", they are on 12-inch wafers. There's no production 8-inch Fab in existence that can do 45nm and smaller, mainly because nobody bothered fitting 8-inch fabs with immersion lithography or multi-patterning. The most advanced 8-inch fab belongs to Samsung with 65/70nm nodes, everyone else stopped somewhere between 180nm-90nm.
It's more a comment towards development work, rather than production work. For example, if I recall correctly from my visit to Fab8, GF used 8-inch test workflow for 14nm several years before scale-up and production.
Very VERY fast construction - to get a huge fab from bare ground to full production in less than 2 years (and on a new process to boot) is impressive.
(What a pity that government infrastructure projects can not manage the same speed - look at Crossrail in the UK or the Boston Big Dig in the US for really bad examples.)
It was not from bare Ground. At the time of official ground breaking Dacin had already put a lot of work into that site. But even considering that you are right, a very fast construction.
Once-off/unique projects are always more expensive. Building a new fab is not that: it's repeating and iterating on an existing process.
A better apples-apples comparison would be knowing ahead of time that you were going to build a third crossrail tunnel, and starting work on that just after finishing the first two. Or having an organisation that has a long term plan and funding to do nothing but tunnel expansions.
It still won't be China-speed but that's the cost of having property rights & input on government decisions, but I suspect you'll find that those government infrastructure projects wouldn't look so bad if they weren't always stop-start once-off projects.
If Fab18 will eventually do around 90,000 wafers a month, and it's likely on the same scale as 7nm is currently, of which N5 is ~25%, then TSMC are probably making between 20,000 and 25,000 wafers a month on N5 (averaged over the year).
But they only started half-way through the year, so perhaps Fab18 is 50% active already, at 40-50K wafers a month.
Apple sell 200m iPhones a year. Let's call that 20m a month because some chips will go into iPads and AppleTVs. Assuming A14 is 100mm^2, then each wafer produced 500 working chips. 20m iPhones a month, that's 40000 wafers every month. Then there is A14X, and then MacAppleSilicon, and Huawei had a run earlier on before the ban, and maybe AMD/NVIDIA/Intel/Mediatek/etc also need a bit at the cherry.
Why? The whole point of Apple's SoC and CPU philosophy is "transistors are cheap"! That's why they aren't scared to create massive branch predictors, 128kB L1 caches, 8MB L2 caches, etc etc.
They've hovered the size of A chips at 80 to 100 mm^2 since they started, and in the process they've kept growing both their core and uncore transistor count so that they stay much the same size by area. I fully expect this to continue.
Compare eg A11 (10nm, 87.7mm^2) to A12 (7nm, 83.3mm^2).
11% is ~100,000 wpm (for an annual total of 1.2m 5nm starts). This just 'sounds' right (given that we're usually working with roundish numbers). The question is more how fast they plan to bring additional capacity online (TSMC doesn't seem to typically transition leading edge fabs from node to node like Intel does, they mostly retain capacity for trailing edge customers). If they open another fab/module, they can ramp pretty quickly.
Intel's on ~300,000 leading edge starts per month, so this gives you an example of the scale problem AMD faces when trying to replace Intel in the datacenter.
https://customerfeedback.onl/ Share your visit experience at www.Shoecarnival.com/Feedback & grab the Chance to Win a $200 Shoe Carnival Gift Card by entering into Shoe Carnival Sweepstake.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
13 Comments
Back to Article
dotjaz - Tuesday, August 25, 2020 - link
"However, all of these leading edge process nodes are likely to be on 12-inch wafers."Not "likely", they are on 12-inch wafers. There's no production 8-inch Fab in existence that can do 45nm and smaller, mainly because nobody bothered fitting 8-inch fabs with immersion lithography or multi-patterning. The most advanced 8-inch fab belongs to Samsung with 65/70nm nodes, everyone else stopped somewhere between 180nm-90nm.
albertmamama - Tuesday, August 25, 2020 - link
Maybe Dr. Cutress meant no 18-inch wafers?Yojimbo - Tuesday, August 25, 2020 - link
Where are there any 18 inch wafer fabs and who is making the equipment that goes into it?Ian Cutress - Wednesday, August 26, 2020 - link
It's more a comment towards development work, rather than production work. For example, if I recall correctly from my visit to Fab8, GF used 8-inch test workflow for 14nm several years before scale-up and production.Duncan Macdonald - Wednesday, August 26, 2020 - link
Very VERY fast construction - to get a huge fab from bare ground to full production in less than 2 years (and on a new process to boot) is impressive.(What a pity that government infrastructure projects can not manage the same speed - look at Crossrail in the UK or the Boston Big Dig in the US for really bad examples.)
smalM - Wednesday, September 2, 2020 - link
It was not from bare Ground.At the time of official ground breaking Dacin had already put a lot of work into that site.
But even considering that you are right, a very fast construction.
voiceofunreason - Monday, September 28, 2020 - link
Once-off/unique projects are always more expensive. Building a new fab is not that: it's repeating and iterating on an existing process.A better apples-apples comparison would be knowing ahead of time that you were going to build a third crossrail tunnel, and starting work on that just after finishing the first two. Or having an organisation that has a long term plan and funding to do nothing but tunnel expansions.
It still won't be China-speed but that's the cost of having property rights & input on government decisions, but I suspect you'll find that those government infrastructure projects wouldn't look so bad if they weren't always stop-start once-off projects.
psychobriggsy - Wednesday, August 26, 2020 - link
If Fab18 will eventually do around 90,000 wafers a month, and it's likely on the same scale as 7nm is currently, of which N5 is ~25%, then TSMC are probably making between 20,000 and 25,000 wafers a month on N5 (averaged over the year).But they only started half-way through the year, so perhaps Fab18 is 50% active already, at 40-50K wafers a month.
psychobriggsy - Wednesday, August 26, 2020 - link
But who is using this?Apple sell 200m iPhones a year. Let's call that 20m a month because some chips will go into iPads and AppleTVs. Assuming A14 is 100mm^2, then each wafer produced 500 working chips. 20m iPhones a month, that's 40000 wafers every month. Then there is A14X, and then MacAppleSilicon, and Huawei had a run earlier on before the ban, and maybe AMD/NVIDIA/Intel/Mediatek/etc also need a bit at the cherry.
psychobriggsy - Wednesday, August 26, 2020 - link
TBH A14 is probably 70mm^2 at most because of 5nm density, even if it has more cores and gpu and ai and cache...name99 - Wednesday, August 26, 2020 - link
Why? The whole point of Apple's SoC and CPU philosophy is "transistors are cheap"!That's why they aren't scared to create massive branch predictors, 128kB L1 caches, 8MB L2 caches, etc etc.
They've hovered the size of A chips at 80 to 100 mm^2 since they started, and in the process they've kept growing both their core and uncore transistor count so that they stay much the same size by area. I fully expect this to continue.
Compare eg A11 (10nm, 87.7mm^2) to A12 (7nm, 83.3mm^2).
Sahrin - Wednesday, August 26, 2020 - link
11% is ~100,000 wpm (for an annual total of 1.2m 5nm starts). This just 'sounds' right (given that we're usually working with roundish numbers). The question is more how fast they plan to bring additional capacity online (TSMC doesn't seem to typically transition leading edge fabs from node to node like Intel does, they mostly retain capacity for trailing edge customers). If they open another fab/module, they can ramp pretty quickly.Intel's on ~300,000 leading edge starts per month, so this gives you an example of the scale problem AMD faces when trying to replace Intel in the datacenter.
Fabyous12 - Saturday, September 26, 2020 - link
https://customerfeedback.onl/Share your visit experience at www.Shoecarnival.com/Feedback & grab the Chance to Win a $200 Shoe Carnival Gift Card by entering into Shoe Carnival Sweepstake.