In a big shift to their manufacturing operations – and a big political win domestically – TSMC has announced that the company will be building a new, high-end fab in Arizona. The facility, set to come online in 2024, will utilize TSMC’s soon-to-be-deployed 5nm process, with the ability to handle 20,000 wafers a month. And with a final price tag on the facility expected to be $12 billion, this would make it one of the most expensive fabs ever built in the United States.

Operating over a dozen fabs across the globe, TSMC is responsible for a significant share of global logic chip production, particularly with leading-edge and near-leading-edge processes. The company has become perhaps the biggest winner amidst the gradual winnowing of fabs over the past two decades, as manufacturer after manufacturer has dropped out, consolidating orders among the remaining fabs. And with GlobalFoundries dropping out of the race for cutting-edge manufacturing nodes, TSMC is one of only three companies globally that's developing leading-edge process nodes – and one of the only two that’s a pure-play foundry.

This success has become both a boon and a liability for TSMC. Along with Korean rival Samsung, the two companies have seen massive growth in revenues and profits as they have become the last fabs standing. As a result, TSMC serves customers both locally and globally, particularly the United States and China, the two of which are not enjoying the best of relations right now. This leaves TSMC trapped in the middle of matters – both figuratively and literally – as China needs TSMC to produce leading-edge chips, and the United States is now increasingly reliant on TSMC as well following GlobalFoundries’ retreat.

As a result, the Taiwan Semiconductor Manufacturing Company is going to do something it’s never done before, building a near-leading-edge fab in the US, outside of its home base of Taiwan. The new facility, set to be constructed in Arizona, will use the company’s 5nm process, which is currently TSMC’s most advanced manufacturing process. And while this will no longer be the case by the time it comes online in 2024, when 3nm processes are likely to be available, it would still make the Arizona facility among the most advanced fabs in the world, and by far the most advanced contract fab in the United States.

The Arizona facility would be joining TSMC’s other US fab, which is located in Camas, Washington. It, like TSMC’s other non-Taiwanese-fabs, is based around older technologies, with the Camas fab in particular focusing on building flash products using relatively large process nodes (350nm to 160nm). As a result, the Arizona fab represents a significant shift for TSMC; it’s not the first US fab for the company, but it’s the first time TSMC has built such an advanced fab in another nation.

All told, the Arizona fab is set to be a medium-sized facility – a “megafab” in TSMC parlance – despite its use of an advanced manufacturing node. The 20,000 wafers per month throughput of the fab is well below TSMC’s largest “gigafabs” in Taiwan, which can move more than 100,000 wafers per month. As a result while the fab will add to TSMC’s 5nm capacity, it won’t become a massive part of that capacity. Though with an expected price tag of $12 billion, it will still be a very expensive facility to build.

According to TSMC, the primary impetus for building the fab – and especially to build it in the United States instead of Taiwan – is specifically to have high-end production capacity within the United States. With GlobalFoundries dropping out of the race for leading-edge nodes, the US government and other sensitive fabless chip designers are in want of another leading-edge facility within the US to build their chips. Given their location, TSMC’s Taiwanese fabs are seen as security risk, and the US would prefer to be self-reliant rather than relying on a foreign partner – a concern that’s been magnified by the current coronavirus pandemic and the supply chain issues that has created.

The end result is that while the fab is not explicitly reserved for military and other sensitive chips, it’s hard to imagine a scenario where the big commodity chip manufacturers such as AMD and Qualcomm don’t continue to use TSMC’s larger and more efficient fabs in Taiwan. In other words, while the Zen processor you’re playing World of Tanks on isn’t likely to come from Arizona, the chip that goes into an actual tank just might. Unaspiringly then, the planned Arizona fab comes with a lot of interest and support from both Arizona and the US Government, for both strategic and business reasons.

Interestingly, while the desire for more domestic fabs within the US has been well known, the US’s two major domestic fabs find themselves on the outside looking in with this deal. Intel has long tried to lure customers to its own contract manufacturing business, including making pitches to the US government, with little luck. Meanwhile GlobalFoundries is currently the US government’s preferred fab for military hardware – inheriting those fabs and the resulting business from IBM – but whom faces the prospect of slowly losing business as government customers switch to newer manufacturing processes. So in some respects it’s a bit surprising to see outsider TSMC land such a deal, though there are sizable political concerns involved as well.

At any rate, TSMC’s Arizona fab is expected to break ground in 2021, with completion scheduled for 2024. At which point it is anticipated to employ over 1600 workers as TSMC becomes the United States’ second leading-edge foundry provider.

Source: TSMC

Comments Locked

97 Comments

View All Comments

  • Kevin G - Sunday, May 17, 2020 - link

    3D but how?

    More commonly known as 2.5D, putting multiple logic and memory chips on top of an intersposer to link them all together has been done for years now in the main stream. Advancements are still taking place here, especially with reducing costs, increasing interposer size, and being able to put several high temperatures dies next to each other (thermodynamic fun).

    Full chip stacking with logic on logic can be built today but that runs into thermal issues with existing manufacturing and design techniques. Intel's Fervos is the closest to this concept but side steps the issue by only using low power dies for mobile. If expectations are tempered to this context, there is room to be optimistic as memory could also be stacked in this fashion and end up reducing overall system power further. Just don't expect anything that'd fall under the category of high performance.

    There are some various methods for growing 3D traces between circuits as well as 3D logic gates. While this works in labs, the catch is that the time to manufacture is incredibly long. Not impossible, just impractical from what I've seen thus far. Thermal issues were also expected but I have yet to see this fully explored as I have yet to hear of any research vehicle that built one large enough to be a good thermal reference.

    There is also the possibility of optical based logic gates. This technology does work in the lab but the problem of mass manufacturing these has been elusive. A 3D version of these has been demonstrated too but also has the interesting concept of how the logic gates are suspended in 3D space and not block the optics of other gates. Optics do have a future else with existing designs and packaging techniques though.
  • nandnandnand - Sunday, May 17, 2020 - link

    https://www.darpa.mil/attachments/3DSoCProposersDa...
    https://www.youtube.com/watch?v=6ir_--MgMJI

    First wafer with layers of CNFETs and RRAM on it made last year. Likely to be significant progress by the end of 2021. TSMC and others will license it or develop their own designs within the next few years.
  • Jimbo123 - Saturday, May 16, 2020 - link

    In 2023, Intel is already in 5nm, that is 1 year ahead of TSMC
  • wilsonkf - Sunday, May 17, 2020 - link

    Intel should have been in 5nm this year.
  • Wilco1 - Sunday, May 17, 2020 - link

    TSMC already has 5nm today and will be at 3nm in 2023. Intel *may* have their 7nm process in volume by then if they get everything right. But let's first wait for their 10nm desktop CPUs and servers...
  • FunBunny2 - Sunday, May 17, 2020 - link

    "In 2023, Intel is already in 5nm"

    Increasingly, Xnm means about as much as a fish story or schlong story, just in reverse: "mine is smaller than youuuuuuuurs!! Nyah Nyah".
  • FXi - Monday, May 18, 2020 - link

    Here's a history lesson on just how this ends up going.
    https://www.theverge.com/2020/4/12/21217060/foxcon...

    "Plans" for buildings are made all the time. Long term commitment to putting them to work and hiring a workforce are another matter entirely.
  • melgross - Monday, May 18, 2020 - link

    Been attempting to get TSMC to build a US plant for several years. At first, Apple wanted the plant to dedicate itself to apple only chips. TSMC declined. Samsung’’s Texas plant was essentially dedicated to Apple’s production, so it seems reasonable. While this article doesn’t mention it, I wonder if that’s the deal here. After all, this is a big plant, but not the biggest. The question is how much production a major company like Apple uses a month, on average. They are TSMC’ s biggest customer, and Apple is known to buy equipment for their manufacturing partners, and receive exclusivity as a result.
  • FunBunny2 - Monday, May 18, 2020 - link

    "Apple is known to buy equipment for their manufacturing partners, and receive exclusivity as a result."

    what was that company that had a spiffy new way to make sapphire for screens? Apple reeled them in, and then they went bankrupt. Corning turned out to be the winner.
  • Cullinaire - Tuesday, May 19, 2020 - link

    The difference is TSMC actually had a viable technology going in unlike GTAT.

Log in

Don't have an account? Sign up now