One chilly Tuesday evening last December, dozens of physicists and engineers who dream up tomorrow’s transistors met in San Francisco to ponder the far future. Would today’s state-of-the-art switch—a three-dimensional transistor dubbed the FinFET—be able to carry chips “to the finish,” a distant, possibly unreachable horizon where transistors are made up of just a handful of atoms? Or would we need a new technology to get us there?
This may all sound like the tech world’s version of arguing over how many angels can dance on the head of a pin, but it actually has enormous real-world implications. The semiconductor industry pulled in revenues of US $300 billion in 2012. After decades of fulfilling Gordon Moore’s prophesy of steadily doubling transistor densities (these days every 18 to 24 months), the industry is now delivering integrated circuits with transistors that are made using what chipmakers call a 20- or 22-nanometer manufacturing process. An IC fabricated with this process, such as a microprocessor or a dynamic RAM (DRAM) chip, can have billions of transistors.
Nevertheless, there on the cutting edge, the business is troubled. Each new generation of ultradense chips demands a new manufacturing process of mind-boggling industrial and technological complexity. The struggle has become so pitched that researchers are now often at a loss for words to describe the metrics of their progress.
At the December meeting, for example, Chenming Hu, the coinventor of the FinFET, began by mapping out the near future. Soon, he said, we’ll start to see 14-nm and 16-nm chips emerge (the first, which are expected to come from Intel, are slated to go into production early next year). Then he added a caveat whose casual tone belied its startling implications: “Nobody knows anymore what 16 nm means or what 14 nm means.”
It’s actually become a fairly common refrain among industry experts. The practice of attaching measurements to chip generations has “been hijacked by marketers to an enormous extent,” one chip-design expert told me. “A lot of it’s really smoke and mirrors,” says analyst Dan Hutcheson of VLSI Research in Santa Clara, Calif. It’s “spin,” he says, that’s designed to hide widening technological gaps between chip companies.
The nanometer figures that Hu discussed are called nodes, and they are, for want of a better term, the mile markers of Moore’s Law. Each node marks a new generation of chip-manufacturing technology. And the progression of node names over the years reflects the steady progress that both logic and memory chips have made: The smaller the number, the smaller the transistors and the more closely they are packed together, producing chips that are denser and thus less costly on a per-transistor basis.
But the relationship between node names and chip dimensions is far from straightforward. Nowadays, a particular node name does not reflect the size of any particular chip feature, as it once did. And in the past year, the use of node names has become even more confusing, as chip foundries prepare to roll out 14-nm and 16-nm chips, custom-made for smartphone makers and other customers, that will be no denser than the previous 20-nm generation. That might be just a temporary hiccup, a one-time-only pause in chip-density improvement. But it’s emblematic of the perplexing state of the field.
Moore’s Law, when reflected through the steady march of node names, might seem easy and inexorable. But today a plague of intense manufacturing and design problems is forcing compromises that are sometimes sobering. And some analysts suggest that regardless of what we call the next generation of chips, the transition from old to new no longer provides nearly the kind of payoff—in cost or performance—that it used to.
“What do you mean by 14 nm?” When I asked An Steegen that question at an industry conference in July, she smiled and let out a wry, knowing laugh. “Ah…what’s in a name?” asked Steegen, senior vice president for process technology development at Imec, the Belgian research center. “Actually, not that much any more.”
It’s a state of affairs that has been nearly two decades in the making. Once upon a time, the node name told you practically everything you needed to know about a chip’s underlying technology. If you trained your microscope on microprocessors made by a handful of different companies using a 0.35-micrometer process, you’d find that their products were all remarkably similar.
In the mid-1990s, when such chips were the state of the art, 0.35 µm was an accurate measure of the finest features that could be drawn on the chip. This determined dimensions such as the length of the transistor gate, the electrode responsible for switching the device on and off. Because gate length is directly linked to switching speed, you’d have a pretty good sense of the performance boost you’d get by switching from an older-generation chip to a 0.35-µm processor. The term “0.35-µm node” actually meant something.
But around that same time, the link between performance and node name began to break down. In pursuit of ever-higher clock speeds, chipmakers expanded their tool kit. They continued to use lithography to pattern circuit components and wires on the chip, as they always had. But they also began etching away the ends of the transistor gate to make the devices shorter, and thus faster.
After a while, “there was no one design rule that people could point to and say, ‘That defines the node name,’” says Mark Bohr, a senior fellow at Intel. The company’s 0.13-µm chips, which debuted in 2001, had transistor gates that were actually just 70 nm long. Nevertheless, Intel called them 0.13-µm chips because they were the next in line. For want of a better system, the industry more or less stuck to the historical node-naming convention. Although the trend in the measurements of transistors was changing, manufacturers continued to pack the devices closer and closer together, assigning each successive chip generation a number about 70 percent that of the previous one. (A 30 percent reduction in both the x and y dimensions corresponds to a 50 percent reduction in the area occupied by a transistor, and therefore the potential to double transistor density on the chip.)
The naming trend continued as transistors got even more complex. After years of aggressive gate trimming, simple transistor scaling reached a limit in the early 2000s: Making a transistor smaller no longer meant it would be faster or less power hungry. So Intel, followed by others, introduced new technologies to help boost transistor performance. They started with strain engineering, adding impurities to silicon to alter the crystal, which had the effect of boosting speed without changing the physical dimensions of the transistor. They added new insulating and gate materials. And two years ago, they rejiggered the transistor structure to create the more efficient FinFET, with a current-carrying channel that juts out of the plane of the chip.
Through all this, node name numbers continued to drift ever downward, and the density of transistors continued to double from generation to generation. But the names no longer match the size of any specific chip dimension. “The minimum dimensions are getting smaller,” Bohr says. “But I’m the first to admit that I can’t point to the one dimension that’s 32 nm or 22 nm or 14 nm. Some dimensions are smaller than the stated node name, and others are larger.”
The switch to FinFETs has made the situation even more complex. Bohr points out, for example, that Intel’s 22-nm chips, the current state of the art, have FinFET transistors with gates that are 35 nm long but fins that are just 8 nm wide.
That is, of course, the view from a chip manufacturer’s side. For his part, Paolo Gargini, the chairman of the International Technology Roadmap for Semiconductors, says the node is and always has been defined by the proximity of wires on the first metal layer on the back of the chip, a dimension that was reflected well in DRAM and, later, flash memory, but not in logic.
Regardless of definition, numbers in node names have continued to decline. Along with them, the distance between transistor gates and that between the closest copper wires on the back of the chip have also decreased. Both of those features help define how dense a chip can be and thus how many more you can produce on a single silicon wafer to drive down costs.
But the difficulty inherent in printing ever-finer features has now taken its toll. “When we got to around 28 nm, we were actually pushing the limits of the lithographic tools,” says Subramani Kengeri, vice president of advanced technology architecture at GlobalFoundries, the world’s second-biggest chipmaking foundry after Taiwan Semiconductor Manufacturing Co.
To deal with this, Kengeri and his colleagues were forced to adopt a lithographic technique called double patterning. It lets technicians pattern smaller features by splitting a single patterning step into two, relying on a slight offset between the two steps.
Intel used the technique to form transistors on its 22-nm chips, but it stuck to single patterning to make the densest metal layer. Pushing the technique to its limits, the company made wires with a pitch of 80 nm, which encompasses the width of one wire and the space to the next. By adopting double patterning, GlobalFoundries and others could push the pitch down to about 64 nm for their 20-nm chips. But that move came with a significant trade-off: Double-patterned chips take longer to make, adding significantly to the cost.
Carrying this technique over from the 20-nm node to 14 nm would mean that chipmakers would have to double-pattern even more layers of the chip. So last year, Kengeri and his colleagues announced a chip industry first: They would put a stop to the shrink. GlobalFoundries’ line of 14-nm chips, which are slated to begin production in 2014, may be the foundry world’s first FinFET transistors. But the company will build the new chips with the same wiring density used in its 20-nm chips. “The first-generation FinFET is basically reusing all of that and plugging a FinFET into that framework,” Kengeri says. “It’s really a 20-nm FinFET, in a way.” Nevertheless, the company refers to these as 14-nm chips because they offer roughly a generation’s-worth jump in performance and energy efficiency over its 20-nm chips.
Kengeri hopes that by putting a one-generation pause on shrinking chips and focusing on introducing 3-D transistors, GlobalFoundries will catch up with Intel, which is already shipping 3-D devices in its 22-nm chips. GlobalFoundries’ 14-nm chips aren’t any denser than—and therefore cost just about as much as—the previous generation, but they’re still a big improvement, Kengeri says. “Our point—and our customers agree—is that as long as they see that value, they don’t care what the technology is called or what is inside.”
“It is quite a controversial move,” says William Arnold, chief scientist at ASML, the world’s largest maker of semiconductor-fabrication equipment. “The customers of the foundries, the people who are making cellphone parts, are very skeptical of not being able to get a shrink along with a performance improvement. They’re pretty vocal about saying that they’re not happy about that.”
The foundries’ latest move aside, chips are still more or less doubling in density from node to node, says Andrew Kahng, a professor at the University of California, San Diego, and an expert on high-performance chip design. But for Kahng, the steady progression of node names masks deeper problems. There is a difference, he says, between “available density” (how closely you can pack circuits and wires on a chip) and “realizable density” (what you can actually put into a competitive commercial product).
The sheer density and power levels on a state-of-the-art chip have forced designers to compensate by adding error-correction circuitry, redundancy, read- and write-boosting circuitry for failing static RAM cells, circuits to track and adapt to performance variations, and complicated memory hierarchies to handle multicore architectures. The problem, Kahng says, is that “all of those extra circuits add area.” His group has been scouring company specs and deconstructing images of chips for years, and they’ve come to an unsettling conclusion: When you factor those circuits in, chips are no longer twice as dense from generation to generation. In fact, Kahng’s analysis suggests, the density improvement over the past three generations, from 2007 on, has been closer to 1.6 than 2. This smaller density benefit means costlier chips, and it also has an impact on performance because signals must be driven over longer distances. The shortfall is consistent enough, Kahng says, that it could be considered its own law.
This might be a recoverable loss. So far, Kahng says, the chip industry has made it a priority to keep up the pace of Moore’s Law, ensuring that manufacturers can continue to build and release new product families while using a new process every 18 to 24 months. This means there hasn’t been time to explore a number of design tricks that could be used to cut down on power or boost performance. “When you’re on that kind of schedule, you don’t have time to optimize things,” he says. As the value of the simple shrink decreases, he says, chipmakers should then be able to revisit their designs and find chip-improving approaches they may have missed or else left on the cutting-room floor.
When will the scaling stop? Today’s patterning technology, which relies on 193-nm laser light, is becoming an ever more costly challenge, and its natural successor, shorter-wavelength extreme ultraviolet lithography, has been long delayed.
Kahng says chipmakers may face a more immediate struggle with wiring in just a few years as they attempt to push chip density down past the 10-nm generation. Each copper wire requires a sheath containing barrier material to prevent the metal from leaching into surrounding material, as well as insulation to prevent it from interacting with neighboring wires. To perform effectively, this sheath must be fairly thick. This thickness limits how closely wires can be pushed together and forces the copper wires to shrink instead, dramatically driving up the resistance and delays and drastically lowering performance. Although researchers are exploring alternative materials, it’s unclear, Kahng says, whether they will be ready in time to keep up with Moore’s Law’s steady pace.
Many people in the industry, who have watched showstopper after showstopper crop up only to be bypassed by a new development, are reluctant to put a hard date on Moore’s Law’s demise. “Every generation, there are people who will say we’re coming to the end of the shrink,” says ASML’s Arnold, and in “every generation various improvements do come about. I haven’t seen the end of the road map.”
But for those keeping track of the road, those mile markers are starting to get pretty blurry.
A version of this article originally appeared in print as “The End of the Shrink.”