With GlobalFoundries having announced that it will cancel its 7nm deployments and cease development of leading-edge FinFETs, the number of companies with leading-edge fabs has fallen to three. But of the three, Samsung is a bit of a special case — while it has foundry customers and has competed for Apple and Qualcomm business in the past, it also builds its own silicon to support its mobile phone efforts. It’s a bit of a hybrid — half Intel, half TSMC, but with a fraction the sales of either.
According to EETimes, GlobalFoundries is leaving the leading edge because the economics simply don’t make sense. GlobalFoundries was facing a $2B – $4B additional expenditure to ramp its wafer production to the 40,000 – 50,000 wafer starts per month that are required to earn back a return on investment on the node. According to Caulfield, most of GlobalFoundries’ customers have no plans for 7nm chips. Demand for 14/16nm across the industry was half of what overall demand for 28nm, and 7nm could be half again as low as 14nm.
“When we look out to 2022, two-thirds of the foundry market will be in nodes at 12nm and above, so it’s not like we are conceding a big part of this market,” Caulfield told EETimes.
None of these statements are surprising. If you’re a computing enthusiast you live, in a very real sense, in a sort-of semiconductor bubble. The chips that we collectively care about — high-performance processors built into smartphones and PCs, including GPUs — are only a fraction of the total semiconductor industry. They constitute the highest-performing and highest-value products in the market on a per-unit basis, but in terms of unit shipments, they’re absolutely dwarfed by the combined volume of microcontrollers and small SoCs for various embedded and low-power devices. You can probably name most of the companies that buy and build these type of chips off the top of your head: Apple, AMD, Intel, Qualcomm, Samsung, MediaTek (to an extent), HiSilicon, Nvidia, Xilinx. Some of these companies, like Samsung and Intel, have their own fabs and will use their own product for ramping next-generation nodes.
As the cost of ramping next-generation nodes increases, it means those costs must be born by either the foundry, its customers, or both. And this may be where the bloom began coming off the GlobalFoundries rose and the GF – AMD partnership. Back in 2016, AMD signed a five-year WSA agreement with GF that was supposed to cover both the 14nm and 7nm nodes, with specific allowance for AMD to seek manufacturing flexibility if required, but with a clear eye towards guaranteeing that AMD would have a ramp on 7nm. As part of this negotion, AMD disclosed that it expected to buy $650M of wafers from GlobalFoundries in 2016. It later amended that amount in its form 10-K to note that it actually spent $1.1B on wafers from GlobalFoundries in 2017 and $700M in 2016.
But — and here’s the really big “but” — according to EETimes, AMD was also GlobalFoundries sole 7nm customer. That’s critically important when you consider that in 2017, much of AMD’s total business was running out of GF. GlobalFoundries builds Polaris, it builds Vega, it builds AMD’s entire Bulldozer family (which the company is still shipping), and it builds Ryzen. Apart from the consoles, GF had the entirety of AMD’s business, minus whatever 28nm GPUs the company might have shipped from TSMC in 2017. But the days when a single customer can drive a leading-edge node are over, and some of you may remember a few years back when AMD announced it would no longer pay for custom nodes. When you’ve got one customer, you’re effectively building a custom node no matter what.
This shift marks a change in the fundamentals of foundry dynamics. In the old days, you could deploy a new node for cutting-edge customers, knowing that you’d make money off it for decades to come. TSMC, for example, earned 27% of its revenue in 2017 from nodes at 65nm or above, and 65nm is more than a decade old. More than half of TSMC’s revenue still comes from the 28nm node and above. But this model assumes — in fact, requires — that customers regularly proceed to newer nodes from older ones. But only certain types of silicon benefit from FinFETs in the first place, which means the benefits of die shrinks are themselves shrinking every generation across the entire industry.
The Opposite of What AMD Wanted
The original goal of spinning off GlobalFoundries was to challenge TSMC at the top of the foundry heap and compete for leading-edge customers. The specific point of this strategy was to amortize the cost of moving to leading-edge nodes among many customers rather than attempting to carry it entirely within AMD. Paying the cost of a 7nm die shrink isn’t as bad as paying the cost of an entire foundry, but with the cost of each new node rising, the $2-$4B in additional investment required for GF to make 7nm work dwarfs what AMD would pay GF for wafers in the near term.
It’s clear that, for whatever reason, GlobalFoundries wasn’t able to convince potential customers to take a chance on its own 7nm as opposed to TSMC’s. After years of struggling to compete at the leading edge and having to cancel its own efforts to do so at 14nm, the company’s owners want to see it turn a profit by focusing on building parts for customers that are profitable now as opposed to chasing after Samsung, Intel, and TSMC in a futile attempt to earn customers and market share. It’s an understandable decision, but an ominous one. As the number of customers for new nodes continues to decline, single customer wins become overwhelmingly important. Samsung can keep its fabs fed with its own product sales, but the trend line here is impossible to miss:
We used to have 19 firms competing at the leading edge. STMicro announced it would lean on foundries for production after 14nm and never put that node into production. Currently, there are five firms offering 14nm, four of which offered it as a leading-edge node: Samsung, TSMC, Intel, GlobalFoundries (leading-edge, with GF using Samsung’s IP) and then UMC, which began offering 14nm as a new capability this summer. We may still see secondary foundries deploy on nodes like 14nm once they are no longer leading edge, but even this is uncertain. The cost structures are squeezing companies out of the market. Whether solutions like FDX can provide alternatives remains to be seen.
The goal of the GlobalFoundries spinoff was to create new customer opportunities and give AMD a partner in a more reliable position to deliver regularly foundry technology improvements. That dream largely failed to materialize. In all honesty, a departure from GF is probably best for both companies — even an incredibly aggressive Epyc and Ryzen sales ramp would not have supported AMD’s 7nm needs within the necessary time frames. Meanwhile, GF wants to focus on the market segments where it might actually have a chance of making a mark, and it’s hard to fault that. But the continuing decline in total customer base, alongside steep cost increases is an ominious preview of a future in which we may have just one company building on the leading edge — or possibly no companies at all.
GlobalFoundries to Soon Pilot 7nm and EUV Production
GlobalFoundries is pushing ahead with a plan to pilot EUV and 7nm production. We recently got a chance to sit down with the company and talk about its plans for the future.
AMD Needs More 7nm Capacity Than GlobalFoundries Can Provide
AMD is sourcing 7nm from TSMC and GlobalFoundries because it expects to have more need for 7nm volume than GF can supply.
AMD Will Fab Its 7nm ‘Rome’ Epyc CPUs at TSMC, Not GlobalFoundries
AMD is building its 7nm Epyc CPUs, codenamed "Rome," at TSMC — not GlobalFoundries.
GlobalFoundries Radically Restructures, Kills 7nm, Spins off ASIC Design Team
GlobalFoundries has announced a massive restructuring, including complete cancellation of its 7nm node and its departure from the leading edge.