The race to get designs done on schedule doesn’t guarantee they’ll hit market windows.
The semiconductor industry could become a victim of its own success. With so many semiconductors being consumed inside of cars, home electronics and industry, capacity shortages are beginning to surface in some areas.
Foundries set rates depending upon a complex mix of process technology, equipment depreciation, customer demand and the need to push customers from one node the next depending upon available capacity. Chipmakers, meanwhile, look at it from the opposite side, searching for a balance between end customer demands for price, performance and power, which usually matches up nicely with the needs of the foundries. But as demand for that capacity begins to spill over the limits of supply, choices need to be made on both sides that cost more money and more time, creating uncertainty for everyone.
This is global economics at its most granular and complicated level. The semiconductor supply chain is a finely tuned instrument with suppliers and customers scattered around the world, some neatly stratified by company size, some by volume of chips produced, others by vertical markets. But there are so many variables and unique features in the semiconductor industry that it’s virtually impossible to comprehend, and therefore manage, how all the pieces interact. Sometimes the pieces get out of alignment, and occasionally that shows up on the manufacturing side.
While there are many places to manufacture different kinds of chips, from specialty and general-purpose foundries to captive fabs, there is a limit to just how many wafers can be produced each day. At advanced nodes, this push forward is being constrained by the cost of outfitting new state-of-the-art fabs. At older nodes, it’s usually constrained by the inability to find any equipment, whether new or used. So as demand increases, capacity remains relatively stable and demand pushes in from all sides.
A good way to look at manufacturing is from a vertical market perspective. Large mobile electronics companies such as Apple and Qualcomm consume huge portions of available capacity at or close to the leading-edge process nodes. A shift by either of those players from one foundry to another could seriously impact capacity at a new foundry. Likewise, any blip in a market they serve—a shift from a 24-month refresh cycles to 30 months, for example, could have serious ripples in opening up capacity. On the flip side, an 18-month refresh cycle could cause a severe capacity shortage, probably in multiple foundries.
Those two companies are so large that much of the industry either works with them or around them at advanced 300mm fabs. So much of the market demand has been concentrated in those two companies, in fact, that a hiccup in one is likely to be felt everywhere and by everybody.
Automotive and consumer Internet of Things chipmakers are beginning to pose, at least collectively, a similar type of force at 200mm foundries using older technology nodes.
“We’re seeing a capacity crunch already,” said Asim Salim, vice president of manufacturing operations at Open-Silicon. “That includes embedded flash, OCP and mixed signal, as well. If you’re doing sensing, you’re talking to analog. Some of this is available on 180nm. It’s clearly not sexy in terms of CapEx but we’re seeing emerging demand, and a lot of the foundries have been caught off guard. There is a need for more and more 200mm wafers because IoT applications are high-volume applications. A lot of the capacity that’s there is already allocated.”
The IoT is a wild card when it comes to capacity because no one knows yet what will be popular, how well individual devices such as a full-function smart watch will sell, and how many other markets and technologies a successful product category will pull along with it.
“If the product is not available after it’s announced, that’s not good for the company,” said Salim. “But in some cases capacity is already allocated. If you forecast more and commit to continued demand and CapEx at a niche node, but you get it wrong, then you are just ‘capacity sitting.’ The reality is that companies will have to make a strong case to fabs for more capacity. If they hear from multiple sources that there is a need, they will listen.”
Just to show how fast this can turn around, used equipment vendors said in October that the second half of the year would not be very strong.
“Now, at the end of the year, they’re saying it’s really good,” said Joanne Itow, director of manufacturing at Semico Research. “There are a lot of companies in need of 200mm equipment and it’s hard to find the right match. The tools are not in stock or they’re looking for the right tool.”
Some new 200mm equipment is still being sold, as well. But how much and to whom isn’t easy to follow because some of these fabs are owned by foundries, some by chipmakers. There also are some new 200mm fabs being built, according to Christian Dieseldorff, an analyst in SEMI’s industry research and statistics group. He said about 60% to 75% of the equipment is used.
But while any increase in capacity sounds good for older nodes (Fig. 1), it looks dramatically insufficient when compared with recent IoT projections (Fig. 2, below).
“We count 109 200mm fabs with a capacity of about 3,700,000 wafers per month,” said Dieseldorff. “We count 97 300mm fabs with almost 4,000,000 wafers per month by the end of this year.”
But it’s the older nodes, in particular, that are creating the greatest worries with the Internet of Things. “The IoT has put a spotlight on system solutions,” said Michael Buehler-Garcia, senior director of marketing for Calibre Design Solutions at Mentor Graphics. “The question is how many chips are 0.35 (microns) and 180nm. All of the sensors are at older nodes.”
Single source vs. multi-source
Another problem that enters into the capacity issue is the uniqueness of the process technology. It used to be a relatively simple matter to build a chip at one process node with standard IP and tools, and then if capacity closed up at one foundry, move it to another foundry at the same process node. That changed at 40nm, and it has become even more pronounced at every node below that.
To complicate things further, even older nodes are no longer interchangeable. There are so many new flavors of older nodes to accommodate low power, multiple power domains and more advanced IP that those nodes require substantial redesigns.
“We’ll see additional capacity at 28nm with SMIC and UMC, but with all the movement around gate first, gate last, high k/metal gate and low-cost poly SiON processes, you can’t do a second source of your design,” said Kevin Kranen, director of strategic alliances at Synopsys. “It now requires dual designs and dual IP sourcing.”
This may be good news for IP suppliers and tools vendors initially, but it can create significant disruption in the supply chain if anything goes wrong. While the application processors found in smart phones get a lot of attention, there are many other components inside a smart phone that are not built at the latest process node.
“Capacity has tightened for all but the largest players,” said Greg Waters, president and CEO of Integrated Device Technology. “There are simply fewer places to get semiconductors manufactured.”
He’s not alone in seeing that. Across the industry, the general consensus is that capacity everywhere is growing scarce, although there seems to be at least some availability at 28nm and 20nm.
“Capacity at advanced nodes feels like a boom time,” said Mike Gianfagna, vice president of marketing at eSilicon. “Silicon demand and supply already has caused some challenges for us. It’s nothing like in 2000, but the semiconductor world is better managed now than it was then. Looking ahead, a lot will depend on how well nodes diversify and how much competition there is between TSMC, GlobalFoundries and UMC.”
Gianfagna said that today, most of the volume is at 45nm and 90nm, but the big push will be at 28nm. If that node really takes off as expected, the lag time between creating designs and shipping silicon will grow. But he said it might also grow at older nodes using processes that are tweaked for lower power and new protocols. “It’s older technology, but it’s a new design style.”
All of this is bad enough with high-volume chip runs, but what about IoT devices where no one is quite sure what will do well in the market and what won’t do well? There are a couple of answers to that question. The first involves different approaches to volume production.
“One solution to this is the multi-project wafer, where you pre-book fab capacity and make a number of chips in parallel, then you go to the one that will do well later,” said Drew Wingard, chief technology officer at Sonics. “The jury is still out as to whether the superchip concept will work for IoT applications because they may not be able to tolerate the extra power and area. If it is viable, then what we have been doing for the past 15 years will be applicable where you blow fuses and put up firewalls. What’s changed, though, is that in the SoC space the choices we were making used similar interfaces. In the IoT it’s very different—you have ZigBee, Bluetooth, WiFi and others. If you turn off WiFi, that’s a lot of overhead.”
Wingard said that with the IoT, the real tradeoffs are between the cost of communication and computation. How those functions are partitioned has a very real effect on where the capacity gets used on the manufacturing side because the process technologies will be different.
The second approach involves lawyers. “If you need to get 10,000 wafers a month and you only can get 8,000, that’s not really a problem,” said one industry insider, who spoke on condition that he not be named. “But if you’re a startup and you’re only doing 20 wafers a quarter, that’s different. The big foundries are scared to death of killing the small guys because they’re afraid they’re going to get sued.”
A couple tactics that foundries can use to make sure that all of their capacity is utilized, though, are to require legal commitments for a certain number of wafers over a set time period. That safeguards the foundry when capacity is extremely limited. A second option is to offer some capacity at 28nm, for example, and provide either additional capacity at 20nm or require chipmakers to pay a premium.
This isn’t a new problem, either. In every boom period, there has always been constrained capacity. “Access to silicon is always cyclical, said Charles Janac, chairman and CEO of Arteris. “You can go to new fabs or, with automotive you can keep technology on older fabs.”
No matter how this plays out, though, the likely outcome is that chipmakers will be doing more scrambling to get their chips manufactured. “The key is to look for options and not only to go to a niche 200mm factory or a large foundry,” said Semico’s Itow. “Right now there are other options available.”