Foundries Accelerate Auto Efforts

Push toward more electronics in cars turns what used to be a marginal business into a profitable one.


Foundries are ramping up their efforts in automotive chip production in preparation for a surge in semiconductors used in assisted and autonomous driving.

All of the major foundry vendors are scrambling to assemble the pieces and expand their process portfolios for automotive customers. The foundries are seeing a growing demand from automotive IC customers amid the push toward advanced driver assist systems (ADAS), electric/hybrid vehicles and traditional cars with more connectivity features. Automotive also is attractive for foundries because many devices don’t require leading-edge processes, meaning a large number of vendors can participate.

Automotive isn’t a new market for foundry vendors. In fact, many of them have participated in the sector for years. Fabless design houses with automotive chips outsource their production to the foundries. Even automotive IDMs, which have their own fabs, outsource some chip production to the foundries.

Until recently, though, automotive wasn’t a top priority for most foundries. “In the past, automotive was not considered a big money maker for the foundries,” said Jim Feldhan, president of Semico Research. “It took too long to qualify a process. There weren’t a lot of automotive customers, and it wasn’t a high-volume market when compared to computing or communications.”

The attitude has recently changed, however. “The recent demand for automotive electronics, such as ADAS, AI, sensor hubs and connectivity, has changed the foundry perspective,” Feldhan said. “The automotive market is wide open (for foundries now). TSMC has qualified their automotive process and others are doing the same. The automotive supplier landscape has also been transformed. Because of and the ‘big brain’ requirements for autonomous driving, there is now a scramble for new designs from companies such as Intel, Nvidia and Qualcomm. Of course, Nvidia and Qualcomm are major foundry customers.”

Both automotive device makers and foundries are benefiting from a surge in demand for various devices in cars, such as analog, memory, MCUs, sensors and others. In fact, the average semiconductor content per car has grown from $62 in 1990, to $312 in 2013, and to $350 today, according to figures compiled from TI, IHS and others. By 2022, the figure is projected to reach $460, according to IHS. But even today, the chip content for hybrids is $600, while luxury models are hovering around $1,000, according to McKinsey.

In total, the automotive semiconductor market is expected to reach $41.7 billion in 2017, up 11% over 2016, according to Semico. In 2016, the automotive IC market grew 8.1% over 2015, according to Semico. The figures include power discretes, sensors and optoelectronics.

Still, the automotive IC market represents a small percentage—roughly 10%—of the overall IC market. It pales in comparison to the smartphone chip market. “Keep in mind that 100 million vehicles are sold each year compared to 2 billion cellphones,” Feldhan said.

Automotive represents a small but growing business for foundries. Some foundries have seen their automotive businesses grow from zero a decade ago to about 5% to 10% of their overall sales today. Others have a larger percentage mix in automotive.

But foundries face some challenges in automotive. The requirements are rigid, while the product qualification process remains arduous and expensive. And competition is fierce.

Inside the car
Generally, the industry divides a car into five main domains—body, connectivity, fusion/safety, infotainment, and power train. Body involves the basic body controls, such as the door access, lighting and windows. Connectivity consists of the cellular, WiFi and related capabilities.

Fig. 1: Semiconductors used pervasively in automobiles. Source: UMC

Fusion/safety consists of cameras, LiDAR and radar. Infotainment involves driver information and entertainment. And the powertrain domain includes engine control and the transmission. Some fold the chassis, such as the brakes and steering, into the powertrain domain.

Fig. 2: Diversified silicon technologies required. Source: UMC

Each car also has dozens of specialized embedded computers called electronic control units (ECUs), which control the various domains in a vehicle. The ECUs are all connected via a network in what some call a distributed architecture.

Many vehicles will continue to incorporate a distributed architecture, although some higher-end models are undergoing some radical changes. “We are really at an inflection point now, where the amount of content that’s coming into the vehicle is no longer sustainable by the current architectures,” said Glen De Vos, chief technology officer at automotive technology giant Delphi. “We have to make a change there in order to enable that content will continue to grow and grow in a cost-effective way.”

OEMs can no longer just cram more electronics into in cars. “As you continue down that line, this is where you basically run out of headroom in the current architectures. We can’t simply add more content, more ECUs and more wiring every time we add another function,” De Vos said at a recent event. “The vehicle cost structure can’t afford it. The OEMs can’t afford it. When you think about how you budget a vehicle, you only have so much to spend on all that content.”

To address the issues for some models, Delphi has developed a domain centralization architecture. For this, electronic content is aggregated in fewer multi-domain controllers as a means to reduce cost and weight.

Fig. 3: Vehicle computing evolution. Source: Delphi

Audi, for one, is using this concept in its recently announced A8 luxury car. The A8 also incorporates some autonomous driving capabilities. It can automatically take charge of driving in slow-moving traffic at up to 37.3 mph on freeways.

Fig. 4: Delphi’s Smart Architecture. Source: Delphi

The A8 has “Level 3,” or “limited self-driving” capabilities. In the ADAS world, “Level 1” involves the automation of one or more control functions in a car, while “Level 2” is the automation of two or more functions. Tesla is currently at “Level 2.” “Level 4” has high self-driving capabilities, while “Level 5” is fully autonomous, steering-wheel optional.

Full self-driving technology may not reach the mainstream for a decade or longer. Even if it never reaches the masses, ADAS is fueling the development of new devices on several fronts. According to Semico, technology drivers include adaptive cruise control, auto parking, and collision avoidance, lane departure warning and blind-spot detection.

Outsourcing trends
For years, Bosch, NXP, On Semiconductor, Renesas, STMicroelectronics, TI and other IDMs with fabs have dominated the automotive landscape.

Generally, automotive devices are manufactured in both 200mm and 300mm fabs. “In automotive, you have ADAS, telematics and infotainment. We have the miniaturization of things like LiDAR. That’s going onto a single chip. This is all happening, for the most part, on 200mm,” said Mike Rosa, director of marketing for Applied Global Services at Applied Materials.

Up until a decade or so ago, X-Fab Silicon Foundries was one of the few foundries that served the automotive sector. Other foundries did some business with automotive customers, while some stayed on the sidelines.

Over the years, though, there were two main events that changed the landscape for foundries in automotive. First, many IDMs went to a “fab lite’’ or fabless model. Second, the chip content in cars began to grow.

Starting in the 2000s, the cost of building new fabs and developing leading-edge processes became too expensive for many IDMs. Many stopped building leading-edge fabs. In general, they kept their proprietary processes in-house and outsourced some production to the foundries.

Automotive is just one of the product areas outsourced to foundries. “Historically, in automotive, a lot of the focus has been at IDMs. Foundries had some of that business,” said Walter Ng, vice president of business management at UMC. “But with the movement towards fab lite on the IDM side, a lot more of that is coming out.”

Foundries are also seeing a different product mix being outsourced from the IDMs. “Traditionally, you’ve seen some of the less critical automotive products being manufactured at foundries,” added Wenchi Ting, associate vice president at UMC. “Infotainment is an example. Display drivers are another example.”

To one degree or another, IDMs also outsource analog chips, mixed-signal ICs and sensors. “Traditionally, for the more critical applications like the power train or the chassis control components, the IDMs tend to make these components themselves,” Ting said. “Going into the future, this may change. We’re seeing chips being planned for future engine control, which requires tremendous memory bandwidth. The chips need embedded flash memories together with state-of-the-art logic processes.”

In fact, IDMs already are outsourcing some of the critical applications to foundries. For example, ADAS requires advanced microcontrollers (MCUs), and many IDMs don’t have the logic processes to make them. “They are using 28nm and 40nm for making the processors for ADAS, for example,” Ting said. “That’s a sector where most of the automotive IDMs cannot serve. They simply don’t have the in-house capabilities.”

Another area involves power discretes, which are used in hybrids and electric vehicles. IDMs still make discretes, although they are turning to foundries for other reasons. “People are running out of capacity for power discretes,” he said. “Those products are coming to foundries.”

Besides the outsourcing trend, foundries have witnessed another major event in more recent times. “We started to see a real inflection point over the last couple of years,” said Mark Granger, vice present of automotive at GlobalFoundries. “You can start to see the amount of semiconductor content in a vehicle starting to grow. With the addition of ADAS, there is certainly a surge over the last couple of years.”

As a result, foundry vendors continue to bolster their efforts in automotive. Each vendor has a different strategy. Many take processes from other markets and offer them for automotive customers. Many have also developed processes tuned for automotive.

There is one constant, however. The requirements are more stringent in automotive, compared to other markets. “You want to strive towards zero dppm, or defective parts per million,” Granger said. “You work continuously to improve both on the foundry side as well as the device manufacturers’ side to get to the highest levels of reliability and the lowest level of failures in customers’ hands.”

Both automotive device makers and foundries must adhere to various quality standards, such as the AEC-Q100. This standard involves the failure mechanism stress test for chips. There are other requirements as well. “There is a reason for (these requirements),” said Rajeev Rajan, vice president of IoT at GlobalFoundries. “It’s functional safety. Whatever goes into manufacturing and product development needs to tie back to the regulatory aspects for safety as well as the traceability of parts if anything goes defective.”

These requirements create some challenges. For example, foundries typically develop a process technology and qualify it with a relatively normal sampling size.

In automotive, though, foundries must conduct more and rigorous inspection, test and other screening steps. “You can’t have parts fail because it impacts safety,” said Robert Cappel, senior director of marketing at KLA-Tencor. “So you’re seeing a much different level of quality and yield. There also is a focus on latent reliability defects. A part may pass a test but fail over time within a car. The requirements are changing.”

This becomes particularly important as carmakers begin pushing logic—the brains behind autonomous and assisted driving—into the most advanced process nodes. The goal is to use the latest technology to achieve performance improvements, but how these devices will fare over time in an automobile’s harsh environment isn’t clear.

“Reliability is critical,” said Tom Quan, a director at TSMC. “Right now we’re at 16nm FFC, moving down to 7nm, which will be the autonomous driving platform. There’s a 150˚ C junction, IP, and an SoC has to be validated. It also requires IP that is ISO 26262 compliant, which is more like a checklist, and the IP has to be certified by the manufacturer.”

Along with this, customers are looking for the potential failure rates for devices. It falls back on the foundries to do more inspection, testing and simulation using a larger sampling size, all of which takes time and adds cost to the process.

MCU process trends
Device makers, meanwhile, are outsourcing a slew of products to the foundries, including MCUs. MCUs perform the central processing functions in a system. More than 100 MCUs are used in a vehicle in various places.

For instance, MCUs are used in the body control domain. “Those use microcontrollers that range from 8- and 16-bit with a certain amount of integrated nonvolatile memory combined with high-voltage analog content to interface with the battery,” said Ron Martino, vice president of application processors at NXP.

Other domains require high-end MCUs. “As time has gone on, different requirements have pushed into 32-bit and now 64-bit compute (functions),” Martino said.

Last year, for example, NXP unveiled the i.MX 8 family of 64-bit applications processors. The devices are designed to enhance the automotive dashboard graphics, such as the instrument clusters, infotainment visuals, heads-up displays and rear-seat screens. Based on a 28nm FD-SOI process from Samsung, NXP’s devices include up to six 64-bit ARM v8-A cores, DSPs and memory. Many of NXP’s MCUs use bulk CMOS, although FD-SOI makes sense for the i.MX 8 and related chips.

“(FD-SOI) has a much larger dynamic range in terms of the ability to tune the device,” Martino said. “That allows us to get a much wider range of power/performance optimization on a single platform.”

Not all are pushing FD-SOI, however. Many say bulk CMOS is suitable for most applications. Both camps, however, are seeing other integration trends for MCUs. “For automotive, we are getting requests to put embedded flash together with BCD,” UMC’s Ting said. “It’s an MCU with BCD and embedded flash. This is replacing existing solutions, which require an external standalone memory to record certain parameters for the car.”

In automotive, bipolar-CMOS-DMOS (BCD) is a specialty processes used for motor-control applications, such as mirror positioning and seat adjustment. BCD combines the advantages of bipolar for analog, CMOS for digital, and DMOS for power and high-voltage.

MCUs are also integrating NOR or EEPROM. The mainstream market for embedded NOR flash is 40nm and above, although the industry is moving toward 28nm.

NOR flash suffers from data loss over time, however. So, the automotive industry is taking a harder look at the next-generation memory types, such as MRAM and ReRAM. These technologies have the non-volatility features of flash with unlimited endurance.

“The emerging memories, such as MRAM and ReRAM, are very attractive,” Ting said. “They have characteristics that the traditional memories can’t match. The problem is that they are only in small-scale production at most today. The industry needs to accumulate more experience and flush out all of the potential problems. That will take some time.”

Driverless cars
Clearly, meanwhile, autonomous driving is receiving a lot of attention, if not hype. But the technology and the associated chips are in its infancy.

For example, Audi’s A8 car incorporates a central driver assistance controller, dubbed the zFAS, which enables Level 2/3 autonomous driving. Developed by Delphi, the zFAS board incorporates some older-generation chips. The latest chips weren’t available when Delphi began the project.

The board, for example, includes Mobileye’s EyeQ3 processor and Nvidia’s Tegra K1, according to Audi. Nvidia’s Tegra K1 is a 28nm GPU. Based on a 40nm process, the EyeQ3 enables visual recognition and detection. Mobileye, which was recently acquired by Intel, is also developing chips based on 28nm FD-SOI and finFETs at 10nm and below.

Today’s autonomous-enabled cars also incorporate 6-8 cameras, 6-8 LiDAR devices and 6-8 radar units, according to Delphi’s De Vos. With those devices, a vehicle is collecting 50 terabytes of data every hour, but that’s not nearly enough for full self-driving technology.

Fig. 5: ADAS Centralized fusion/control. Source: Delphi

“Even with all that data, the perception system is still the limiting factor in automated driving,” De Vos said. “We are not at that point where our sensor systems are even close.”

For Level 5, a vehicle may require the ability to collect 150 terabytes of data an hour, meaning the industry requires some new breakthroughs to enable the next wave of both traditional and self-driving cars.

So what’s next? “ADAS and autonomous require a lot of sensors,” GlobalFoundries’ Granger said. “You also have to process all this data. There are two trends in the industry. One is to make the sensors smart. So there is processing in each one of those sensors.”

Then, the industry is developing more powerful MCU/MPUs as a means to handle the data. “When you start to get into Level 3, 4 and 5, you need advanced nodes to be able to process all of the sensor data in one holistic view to enable the car to make the best decision,” Granger said.

There is also work taking place in LiDAR and radar. Radar detects objects, but it can’t discern one object from another. Using laser-based light, LiDAR measures the distance to a target with more precision. In some cases, LiDAR devices are based on gallium nitride (GaN), a III-V technology.

For radar, a device may incorporate an MCU, DSPs as well as RF blocks based on silicon-germanium (SiGe) at 77-GHz. Seeking to reduce the cost of radar, GlobalFoundries and its partners are developing a CMOS-based mmWave radar chip. Using 22nm FD-SOI, the device would include the MCU as well as short- and long-range radar. It also eliminates the need for SiGe.

“It allows you to do fast processing, so that people can implement radar that will attempt to improve upon the previous capabilities and to challenge LiDAR,” he said. “LiDAR is a very interesting technology, but it has also traditionally been expensive.”

Autonomous driving, of course, requires a software component. For example, startup AImotive recently rolled out a neural networking technology for the automotive industry. “All of these sensors are great,” GlobalFoundries’ Rajan said. “But a lot of stuff they do is triangulate data from different input points. You need something in the backend, either inside the car or in the cloud, that has the ability to process this data, crunch it away and give meaningful and actionable decisions as well as focused responses back to the car or anybody in control from a mission critical standpoint.”

Clearly, automotive is bringing new opportunities for foundries. Automotive may take more time to win customers, but it’s a more steady and predictable business, at least for now.

Related Stories
Electric Vehicles Set The Pace
Rethinking Car Design
Connecting The Car
Automotive’s Unsung Technology
Autonomous Cars Drive New Software
Self-Driving Cars Rattle Supply Chain
The LiDAR Gold Rush
Rethinking Verification For Cars
Test More Complex For Cars, IoT

Leave a Reply

(Note: This name will be displayed publicly)