Internally, Intel has initiated “emergency correction” of its "platform blueprint’ and "own process production capacity plan for the next 1 year. It is widely rumored that CEO Pat Gelsinger plans to visit Taiwan for the third time in August to meet with TSMC Chairman Mark Liu and CEO C.C. Wei to discuss these revised plans. Intel’s 14th-gen Meteor Lake originally planned to enter mass production at the end of 2022 for launch in 1 H’2023, has been delayed to the end of 2023. Since Meteor Lake’s GPU tile is outsourced to TSMC, this delay will disrupt TSMC’s 3nm plans. According to industry sources, Intel will pay a heavy price if Meteor Lake suffers a delay. Intel and TSMC had already signed an outsourcing agreement elevating Intel to “heavenly king” customer status. Earlier, there were rumors that TSMC, upon confirming receipt of Intel’s huge order, and in order to segregate from Apple, proceeded to convert what was originally planned as an R&D Center and mini-line (P8–P9, at Baoshan Fab2 gigafab expansion phase) into a second 3nm production site. Contractually, TSMC would have been on schedule to produce the 3nm GPU tile. However, if Intel’s own Intel 4 compute tile is not ready for production due to “market conditions” and process technical problems, and Intel wishes TSMC to also delay production, then Intel will have to absorb all losses incurred. But, there are also rumors that Intel is being forced to come up with another plan, which is to proceed with production of the 3nm GPU tile as originally planned, and re-target/outsource the compute tile to TSMC 5nm or even 3nm. This may cause Intel to “lose face” but will allow Intel to “temporarily catch its breath” and save costs.
So… if Intel resorted to this plan that would soak up a lot of TSMC capacity – making MANY companies unhappy.
I’m wondering what TSMC really would want in return – money? An ability to “outsource back” some manufacturing to Intel’s foundry business on processes that actually work? Maybe (just wild speculation) an existing Intel fab in the US that they could convert faster than they could build out net new buildings here? What else?
There are so many conflicting rumors about meteor lake right now, I’ll wait a few weeks and see what Intel says at earnings. We do know there are 3 different meteor lake dies. It is likely Intel will first launch with the smallest 40mm die that is only 2 P-cores, and 6 E-cores. It would likely go into some sort of handheld device. The real notebook chip is closer to 80mm and will likely come second, and then the desktop chip is over 100mm and would come third. I suspect the “slip” rumors have heard the schedule for one of the larger die products.
Alan
I lean towards the last. TSMC wants to establish a corporate identity outside of Taiwan so the company will survive if China invades Taiwan. Taking over an existing Intel fab and upgrading it would speed up that process while increasing their overall capacity so long as Taiwan remains independent. Promise Intel first crack at those wafers and it’s a win win.
Taking over an existing Intel fab and upgrading it
That doesn’t work. The newest Intel fab in the US is fab 42, which is about 10 years old. It is capable of being outfitted with the older EUV machines, but is not high NA EUV capable. TSMC is ahead of Intel in building new fab shells that are capable of EUV lithography. Intel gets their first production N4 factory in early 2023, and second one in early 2024. Intel is severely constrained for leading edge capacity.
An interesting artifact of this is the largest tile on the meteor lake product is an old 14nm Intel SOC. The next largest is the TSMC GPU, and finally the Intel 4 CPU is tiny with only the I/O die being smaller. It is a bit of a mystery what is on that SOC tile.
Intel does have some limited Intel N4 capacity in the Oregon development fab.
–Alan
That doesn’t work. The newest Intel fab in the US is fab 42, which is about 10 years old. It is capable of being outfitted with the older EUV machines, but is not high NA EUV capable. TSMC is ahead of Intel in building new fab shells that are capable of EUV lithography. Intel gets their first production N4 factory in early 2023, and second one in early 2024. Intel is severely constrained for leading edge capacity.
Thanks for that, that’s the kind of insight I was hoping for. TSMC building new shells in the US, though?
I saw a piece not too long ago in which a senior Chinese government official (economist?) declared that TSMC establishing six fabs outside Taiwan was considered a “red line” that could trigger Chinese military action… that’s why I wondered about converting existing fabs, to get there faster.
TSMC currently has one EUV factory under construction in Arizona. As mentioned, the last Intel factory was 10 years ago and was supposed to be Intels first EUV factory. Today, it is the last DUV factory producing about half of the Intel 7 output.
Intel currently has 6 new EUV factories under construction. The first in Ireland comes online early next year, then a new one in Israel early 2024. They have two under construction in Phoenix which should be late 2024 to early 2025, and then the 2 new Silicon heartland factories in Ohio that just broke ground.
The old DUV factories do not have the necessary infrastructure to run EUV machines. ASML changed the specification again for the high NA EUV machines, so many of the existing TSMC EUV factories can not be upgraded for them either.
Alan
All of the southwest USA is in the worst drought in a millennium. All the major reservoirs on the Colorado river are at critically low water levels, affecting the big hydroelectric generators as well. I’m sure there are business reasons for choosing Arizona, but putting three fabs into an area that already imports most of its water seems like a terrible idea.
Do newer fabs no longer require large amounts of clean water as part of the processing?
They still take a fair amount of water, but also do a lot of recycling to preserve what they have. I know this has been a concern, but compared to other industries, especially agriculture, it is a small amount of water.
Alan
Do newer fabs no longer require large amounts of clean water as part of the processing?
Define “large amounts”. A typical semiconductor manufacturing facility uses two to four million gallons of ultra-pure water per day. (1) The city of Phoenix uses 2 billion gallons per day.
But farming uses way much more. The typical 120 acre circle pivot irrigation field (2) uses one million gallons of water per day. So four irrigated fields equals one fab. You can look out your plane window and see thousands of irrigated fields. Go look at the satellite view of the desert in the southwest. It’s filled with mile after mile of irrigated crops.