Silicon Labs Tackles IoT Challenges
Providing silicon to the IoT market requires a maniacal focus on power consumption at 180nm and 90nm design rules. Silicon Labs is succeeding in combining low-power wireless, MCUs, and flash in its SoC products.
The semiconductor industry is looking to the Internet of Things (IoT) to drive new demand as traditional sectors such as PCs flatten out or decline. To understand what it takes to create successful IoT solutions, a close look at Silicon Labs in Austin, Texas, illustrates how the mix of cost, low-power wireless, and other factors makes the IoT market as different as a sushi bar is from a barbecue joint.
Founded in 1996 by three veteran mixed-signal designers, Silicon Labs started out with the goal of applying standard foundry CMOS to non-traditional applications, including RF. That RF CMOS foundation, according to the company’s executives, prepared Silicon Labs for the IoT market, where wireless data transmission is a key part of the IoT equation.
After industrywide slow-growth years of the financial downturn, the company’s board of directors promoted Tyson Tuttle, who had started at Silicon Labs as a wireless design engineer, to the CEO position in March 2012, charging him with finding faster-growing markets. Tuttle focused the fabless 1,100-person company on IoT solutions, where the company’s portfolio of microcontrollers (MCUs), RF and sensor solutions would form a foundation. About a third of the company’s revenues now come from IoT markets; in the first calendar quarter of 2015, revenue from that sector was about $61 million, up 26% year-on-year.
"To be able to integrate the wireless function together with the microcontroller, you have to be in the same process technology," Tuttle said in a recent industry presentation. "Then you must have the flash memory integration, and you have to figure out how to make all of the wireless connectivity work in the presence of all that digital processing going on." In the IoT marketplace, he added, power consumption is as important as cost, because many of the end-node products are expected to run on coin cell batteries for years at a time.
AN IoT ACQUISITION STRATEGY
During his tenure as CEO, Tuttle has led three key IoT-related acquisitions: Ember (Boston), an early pioneer of ZigBee wireless solutions; Energy Micro (Oslo, Norway), which developed a series of low-power ARM-based 32-bit microcontrollers; and most recently Bluegiga Technologies (Espoo, Finland), a provider of Bluetooth and Wi-Fi wireless solutions for wearables and other applications.
Because IoT solutions are used in thousands of different applications, with tens of thousands of customers, Tuttle noted that "you have to build these SoCs in ways that are very general purpose." Some of the biggest markets, he said, are smart metering for energy, water, gas, and heat; a range of wearables; home automation and security; diverse industrial and factory automation applications, and RFID and digital shelf labels for retail customers.
POWER CONSUMPTION IS KEY
Sandeep Kumar, senior vice president of worldwide operations at Silicon Labs, said, "The key challenge is that these end-user node devices for the IoT are wireless. Whether they are smart energy meters or health monitors or security systems, they will all deliver their data into the network and cloud through low-power wireless. And, a motion sensor or a door sensor has to last 5 to 10 years on a cell battery."
Kumar started his career at Bell Laboratories in New Jersey. He left Agere Systems to join Silicon Labs in 2006 to oversee the company’s manufacturing strategy. A key part of this strategy was an initiative to drive design-for-cost for Silicon Labs’ IC products. This methodology involves a variety of manufacturing inputs and variables, ranging from process technology to test coverage, to determine the most cost-effective manufacturing solutions.
"We did a very detailed analysis of what would happen if we moved [from 180nm] to 110, 90, and 55 nanometers, through our design-forcost process. The chip design team is a critical part of this process. To achieve our performance and cost goals, we concluded that the sweet spots for our IoT portfolio were 180nm and 90nm. We also continue to look at smaller geometries. We need embedded memory for many of our IoT devices, and yet nothing is available at 28nm or 40nm. We have concluded that with the features and performance we need in our SoCs, 55nm was not the most cost-effective process node for our IoT manufacturing needs," Kumar said.
Sandeep Kumar, senior vice president of worldwide operations at Silicon Labs
Many of Silicon Labs’ components are made on 200mm wafers at 180nm design rules, and that 180nm-based portfolio is growing. Other IoT-centric SoC devices, where larger blocks of flash and RAM memory are needed or faster processing is required in a smaller die size, are being targeted to 90nm foundry processes on 300mm wafers. (Other Silicon Labs products, such as digital TV tuner ICs, are made on 55nm technologies.)
At a SEMI event held in Austin in mid-May, Kumar said a growing concern is fab availability. "All the 180nm fabs are full, which is kind of counterintuitive. That has never happened before, to my knowledge. All of the foundries are on allocation at 180nm because of the demand for sensors and other trailing-edge devices."
Daniel Cooley, vice president and general manager for MCU and wireless products at Silicon Labs, said the boom in smart-card chips is one application sopping up 180nm capacity. "There is a shortage of 180nm capacity worldwide. Every time there is excess capacity at TSMC we buy it, because we don’t know when we will get excess again. We call them every week, asking ‘Did anybody cancel? Did anybody push out?’"
RF FUNCTIONS COME LAST AT FOUNDRIES
Foundries traditionally start out by building a new process aimed at high-performance servers or smartphone processors; only later do foundries offer non-volatile memories and support for wireless.
"The last things are the RF parameters. RF is one of the hardest things to do, because these processes were designed from the ground up to handle digital," Cooley said. The foundries have to add mask layers, or additional metal layers to provide the capacitors and inductors needed to create RF and passive components. A foundry often will add transistors with a higher threshold voltage to reduce power consumption, especially when the device is in sleep mode.
"In terms of manufacturing, there are only a few fabs out there available to fabless companies that can do all of this," Cooley said.
FINDING BUSY FABS KEEPS COST DOWN
Kumar said an important investigation for Silicon Labs is determining how much foundry capacity will be available in the future, and estimating the yields and wafer costs at those fabs. If a foundry partner has a nearly full 200mm fab and a half-full 300mm fab, "there is no question" that the 200mm fab will be cheaper, he said.
"One question we ask our suppliers is how they see their capacity in five years. It is a fine line; we don’t want to be in a node that is oversubscribed either, so you don’t get capacity. We want to be in fabs that are very busy. We don’t want to be in a fab that is only 50 percent utilized, because then we are likely to see an increase in price, or a fab shutting down," Kumar said.
Applying what he calls "intuitive science," Kumar queries his suppliers to find fabs that are likely to be 90–95% utilized in three years. Last year, two fabs were shut down in Germany and Korea; Kumar said Silicon Labs had predicted those closures because utilization rates were so low that the fabs were losing money.
Kumar said he is keeping a close eye on the limited availability of used 200mm fab equipment. "We have very confidential meetings with our foundry partners, where we share our 5- and 8-year plans. We learn from them the factual data of how they plan to invest, and how they see customers moving from 200mm into 300mm that will open up 200mm capacity."
Kumar also said TSMC "is giving us capacity" because they see Silicon Labs as a key player in the IoT space, given its fabless nature. Other big players in IoT chips have their own fabs and won’t provide the major foundries with much IoT-related business. And Silicon Labs needs partners to augment the capabilities of its engineering staff to keep up with the myriad projects the company has on its plate. "I wish we could go out and hire a hundred engineers to handle all of the projects we have on our wish list, but our financial guys have different ideas," Kumar said, laughing.
The back end also presents unique challenges. IoT chips are so small and lightweight that they can literally float in the air during the testing phase. Silicon Labs has developed test algorithms to deal with the combinations of flash, control logic, and RF circuitry on its wireless microcontrollers. It uses built-in self-test (BIST) extensively. Noting the cost limitations facing nearly all IoT end-node markets, Kumar said semiconductor companies "must plan for a commoditized market. We are selling all of this functionality for the price of a slice of pizza."
Adding flash adds about 30% to the number of mask layers. The major foundries are investigating new types of non-volatile memory, such as magnetic RAM (MRAM), ferroelectric RAM (FRAM), and resistive RAM (ReRAM) that could be embedded at less cost. In about five years, embedded flash could be a thing of the past, Cooley predicted.
The amount of SRAM data memory on the company’s IoT devices is growing. “In MCUs, there used to be a rule of thumb, an 8-to-1 ratio of flash to SRAM, with in some cases a megabyte of flash and 128k of SRAM,” Cooley said. Now, that ratio is swinging to 3-to-1, or 2-to-1, with perhaps a megabyte of flash and 512k of SRAM. The change relates to the need to save on battery power, provide the memory necessary for wireless networking stacks, and even buffer the frames of small displays for wearable devices.
IoT applications often are in sleep mode nearly all of the time, waking up only to send data to the cloud. With a large chunk of SRAM, the SoC can store data autonomously from sensors and other peripherals. Once the buffer is full, it can send data to the cloud, or trigger the MCU to wake up. "With more SRAM, there are more things you can do when the MCU is off," said Cooley.
And compared with flash, RAM-based buffers are faster to read, and require less power to write due to the large charging pumps on flash. In short, RAM is lower power, and faster to access, than flash.
A NEW ERA OF MICROCONTROLLERS
Cooley said the need for IoT devices to wirelessly deliver data to the cloud differentiates the IoT from traditional embedded markets, where MCUs were often standalone devices.
"I think you are starting to see in the market a new class of wireless microcontrollers. They are not MCUs, and not wireless chips dedicated to a single function," he said, arguing that few companies have the design expertise to combine a low-power MCU with RF capabilities in a foundry process.
Cooley also said wireless MCUs present firmware challenges to companies that have specialized either in MCUs or wireless ICs. "Only a few companies have the deep software knowledge to do wireless stacks in what we believe is a new class of devices. The software, the RF in CMOS, and the core MCU—you have to put all three together to succeed in the Internet of Things market."
Tuttle noted that other companies will have their work cut out for them to catch up with Silicon Labs in the IoT space.
"We have been building our IoT capability for five years, and we feel we have a several-year lead over what other companies have. It is a subtle point, but it is not easy to integrate these functions in the same technology, and add flash memory, and make it all work. It is difficult if you haven’t done this before."
For additional information about Silicon Labs visit www.silabs.com