It takes about $10 billion or even $15 billion to build a brand new state-of-the-art chip manufacturing plant, and only a few companies in the world can amass the kind of expertise and capital to do it. Globalfoundries admitted yesterday it would no longer invest in a bleeding-edge 7-nanometer fabrication plant, or fab. Advanced Micro Devices said at the same time that it would switch from Globalfoundries to TSMC to build its latest high-end processors.
Globalfoundries will instead focus on making investments in older 14-nanometer and 12-nanometer factories (a nanometer is a billionth of a meter, and the number refers to the distance between circuits), which require less upfront investment and carry a far smaller burden of technology development. That means that just a few companies — TSMC, Intel, and Samsung — are still in the race to build 7-nanometer chip factories.
“Globalfoundries was falling further and further behind, so AMD really had no choice but to switch,” said Linley Gwennap, analyst at the Linley Group. “Also, AMD has been using TSMC for years to build its graphics chips, so they already have a strong relationship.”
This is all happening against the backdrop of changes in Moore’s law, the prediction made in 1965 by Intel chairman emeritus Gordon Moore that the number of transistors on a chip will double every year. The chip industry has been on the metronome of Moore’s law for decades, and that progress has delivered computing power in a smartphone that once took a whole room of computers. Intel executives say that if every industry had moved forward on the same line as the chip industry has, we would be able to travel to the sun on a single gallon of gas, feed the world’s population on a square kilometer of land, and travel at 300 times the speed of light. If you feel like technology has been moving along at an accelerated pace, it’s not your imagination. It’s because of Moore’s law.
But Gordon Moore himself has said in the past that he believed that we would hit an economic barrier to Moore’s law before we hit a technological one. The technological barrier may be at hand as well, as even chip manufacturing giant Intel had to delay mass production of its 10-nanometer chips (due to a difference in naming conventions, Intel’s 10-nanometer chips are the equivalent of 7-nanometer chips from its rivals) until 2019. The task of making chips at smaller and smaller dimensions is getting difficult as engineers run into the laws of physics and the behavior of individual atoms in less than razor-thin manufacturing layers.
At the dawn of chip factories in the 1950s and 1960s, almost every Silicon Valley chip maker created its own chips in factories. They built wafer fabs in Silicon Valley, and then offshored the assembly of the wafers into individual chips to places where labor costs were lower. The rising manufacturing costs went hand-in-hand with greater throughput, so each chip factory could make more revenue. But amassing the capital for a new factory became huge.
That trend led to the rise of foundries, or contract manufacturers who took the designs of “fabless” companies and built them in factories. TSMC was founded in 1987. At first, it saw rising competition as demand for foundries rose, but now many of its rivals have dropped out due to the rising capital costs. And that’s where we are today. It’s perhaps no wonder that TSMC’s market value exceeded Intel’s for the first time in March 2017. Today, Intel is valued at $223 billion, and TSMC is $211 billion.
Dan Hutcheson, analyst at VLSI Research, said that Globalfoundries’ decision to skip 7-nanometer will allow the company to focus on fully depleted silicon-on-insulator, where it has a unique advantage for chips that require features such as low power, radio frequency, and things needed for other chip markets.
“It simply didn’t make sense for them to continue playing in the me-TSMC-too poker game, where the house always has the advantage,” Hutcheson said in an email. “AMD has a long history of working with TSMC that dates back to their entry, via acquisition, into GPUs.”
He aadded, “As for cost to build a fab, it’s now at $15 billion heading to a cool $20 billion. The problem isn’t capital as much as it is finding the people who know what they’re doing at the leading edge.”
For AMD, the switch to TSMC required some foresight, as the company had to go to the trouble to make sure that its processor designs could be manufactured in TSMC’s factories as easily as they could be built in Globalfoundries’ factories.
“Since I consider TSMC 7-nanometer to be in the same class as Intel 10-nanometer (and thus better than Intel 14-nanometer), AMD will actually have superior transistors than Intel until at least late next year. Who would have thought?” said Gwennap.
For AMD, there is some risk in going to a single source at TSMC to make its chips. TSMC will have to scramble to ensure enough capacity to build AMD’s chips and serve its existing customers as well. But the switch will likely simplify AMD’s engineering work, as TSMC already makes AMD’s Radeon graphics chips.
“AMD had to make a call to keep pace with bleeding-edge process nodes and TSMC had the lead. I’d guess Globalfoundries just couldn’t scale big enough to pay for the 7-nanometer development costs without AMD as a customer — possibly even with AMD as a customer,” said Kevin Krewell, analyst at Tirias Research.
In other news, the head of AMD’s computing and graphics division, Jim Anderson, took a job as CEO at Lattice Semiconductor on Monday. But Gwennap didn’t see the latest departure as a huge problem for AMD, which has a strong talent bench.