TSMC announces plans for $16 billion foundry to push the envelope on 5nm, 3nm process nodes

By Joel Hruska

• By Joel Hruska on December 12, 2016 at 7:30 am

We’ve talked several times this year about TSMC’s plans to aggressively take back the market share it has lost to Samsung at 14nm, but the company isn’t just planning to ramp next-generation process nodes as quickly as possible. TSMC announced it will build a $15.7 billion foundry facility focused on the upcoming 5nm and 3nm process nodes. Given how long it takes to build next-generation fabs, that’s not surprising — it takes time to line up permits and permissions to build on a site, and several years of construction before any fab opens for business. Even once a facility is running some volume, it can take 9-12 months to bring the entire fabrication plant fully online.

“We’re asking the government to help us find a plot that is large enough and has convenient access so we can build an advanced chip plant to manufacture 5-nanometer and 3nm chips,” TSMC spokesperson Elizabeth Sun told the Nikkei Asian Review.

As foundry node shrinks have become more difficult, fewer and fewer companies have attempted them. The list below is from 2011, but it illustrates the point well. As of that year, there were 19 companies with production capacity at 130nm and just four companies at 20/22nm (GlobalFoundries eventually joined this group, bringing the total to five).

Image by iSuppli

There are currently four foundries offering 14/16nm technology as well — Samsung, TSMC, GlobalFoundries, and Intel. As Nikkei Asian Review points out, pure-play foundries like TSMC are now dependent on a tiny number of customers to drive the bulk of their revenue, with ~16% of TSMC’s revenue coming from Apple and 16% from Qualcomm. In the old days, graphics cards from AMD and Nvidia helped drive the early adopter cycle, but that hasn’t been true for several years. Mobile devices, not GPUs, are driving foundries to newer process nodes, and that’s one reason why we’ve seen Teams Red and Green skipping the interim nodes that are tuned for mobile products and their yearly cadences but don’t deliver much in the way of improvements for high-powered silicon. As costs continue to rise with each successive foundry generation, it’s going to be harder and harder to find companies willing to invest in cutting-edge technology — so much so, that some companies are actively looking for ways to enhance older nodes as a way to entice firms to adopt them.

TSMC currently plans to introduce 7nm technology by Q1 2018, while Samsung has said it will ship 7nm by the end of that year. Intel, meanwhile, will ship 10nm technology by the back half of next year. As always, keep two things in mind: Foundry plans and introductions can and do change, and you can’t compare process nodes between companies simply by the number in front of the “nm.” Generally speaking, Intel’s process nodes have smaller feature sizes than their counterparts at TSMC/GloFo/Samsung. Intel is still talking up its plans to push into the client foundry business, at least with a handful of customers, but there’s basically no link between the name attached to a node and any specific feature size within the SoC.

As of this writing, the 7nm planned by Samsung and TSMC is expected to be generally comparable to what Intel is calling 10nm, though that could change as more data on specific feature sizes becomes available. TSMC has already begun risk production on 10nm, and Samsung has multiple 10nm launches planned for 2017 as well. Qualcomm is partnering with Samsung at the 10nm node, though they may have switched back to TSMC at 7nm to take advantage of the earlier launch window.

Presumably this facility will be outfitted with EUV from the start, since we’ve previously heard that TSMC wants to introduce next-generation lithography at that node. Of course, it’s anyone’s guess as to whether EUV will be practically ready to roll or not. We’ve seen positive signs this year, but extreme ultraviolet lithography is running more then a decade behind its original timelines for deployment.