Your browser is not supported. Please update it.

25 June 2015

Integration and 14/16nm, more than design, are driving IoT chips

As the mobile device chipmakers eye the internet of things (IoT) for growth, lowering power consumption moves from a key concern to an obsession. In that environment, successful migration to new process technologies is as critical as the latest chip design, and the silicon vendors, and their foundries, are making the leap to the processes which will enable a new wave of ultra-low power connected devices.

The current leap is to 14/16-nanometer geometries with FinFET, a 3D system of stacking transistors which allows for greater density, reducing size and power. Over the horizon, 10nm beckons. All the foundries, and the major manufacturers like Intel, have recently been outlining their roadmaps, and at the Freescale Technology Forum this week, the announcement of multicore IoT processors fabricated in 16nm was as important as the designs themselves.

Not all IoT devices are about tiny geometries. Stuart Forbes, Freescale’s digital networking manager, said in an interview that the firm has no current plans for 10nm, and that it will be several years before it can make microcontrollers in 16nm because it is extremely difficult to implement flash and analog elements in this technology. IoT chips will not follow the single-track route of mobile processors, ever-downwards in geometry terms, but will have far more diverse requirements. For instance, TSMC and other foundries have seen a rejuvenation of their 55nm businesses thanks to the IoT – the 55nm to 40nm node is well trusted and reliable, and is very suitable and cost-effective for slow data rate, low power applications such as smart home lighting and metering, where battery life of up to one year is desirable.

But the low level processes are the ones which grab headlines, as Freescale found at its Forum.

The company’s next generation of multicore processors will use a 16nm FinFET process, it announced, with two of them – the QorIQ LS1048A and LS1088A, being targeted at edge networking gateways for the IoT (an important application also hotly pursued by Intel). Freescale promised to offer the 16nm chips from mid-2016 – they will run on eight 64-bit ARM Cortex-A53 cores running at 1.5 GHz. The 16nm process will provide twice the performance of the current 28nm generation in a similar power envelope (or the same performance at half the power), while the chips will improve their power consumption further by harnessing enhancements to the ARM architecture.

The new processors will also support Freescale’s StarCore digital signal processing (DSP) chips and will include hardware acceleration at Layer 1 for LTE-Advanced capabilities. Packet and security acceleration are also important to the new 16nm platform, said the company.

“The world’s networks are moving and changing faster than ever before, driven by the convergence of extreme virtualization, software-centric network topologies, continued expansion of the IoT, and growing demand for increased, flexible intelligence at the network’s edge,” said Tom Deitrich, general manager of Freescale’s digital networking group, in a statement.

While the QorIQ processors target IoT infrastructure and larger devices, Freescale is arguably a bigger presence at the tiny end of the market, with its significant businesses in microcontrollers for the automotive, industrial and telecoms sectors. This positions it well to move into the middle ground between microcontrollers (MCUs) and microprocessors, where ARM, Intel and many chip vendors see considerable potential to generate new product categories and revenues. These emerging chips are targeted at devices – such as some wearables, connected car systems and sensor nodes – which require more data processing capabilities than an MCU allows, but with power consumption close to MCU levels. Some are addressing it with souped-up MCUs, as is the ARM Cortex-M7 platform, while others are shrinking their processors and adapting them for the massive scale and price sensitivity of the IoT market.

Freescale’s contribution to this new category involves a whole new label- instead of SoC, it dubs its new design an SCM (single chip module). This is designed to provide the functionality of an SoC with lower power and size, and with a high level of pre-integration, so that customers can get products to market up to 25% more quickly. The latter is an important consideration in the IoT market, where there will be a huge variety of firms making devices and sensors, many of them lacking significant inhouse design resource, and needing an almost readymade base platform in order to achieve the necessary volume economics.

Freescale says its initial SCM is based on dual ARM Cortex-A9 cores (the A9, like many older technologies, is gaining a second lease of life from the IoT. Having been introduced for smartphones and digital TV in 2008, it is still a workhorse for ARM in general purpose, low power applications without hefty processing demands).

The SCM will contain the key elements of an SoC, but with more elements integrated on top of the processor cores, notably gigabytes of DRAM memory, a power management chip and a flash memory chip. This leads to a 50% reduction in board space, says Freescale. The first model, the i.MX 6D, is just 14 x 17 x 1.7 millimeters, including a gigabyte of Micron DDR2 memory on top. It integrates total of 109 components in the package, all interconnected using a proprietary technology, without which a 6-inch printed circuit board would be needed to mount all these parts, Freescale claimed.

The SCM is not so much about new design, but about bringing together lots of elements which are manufactured using different processes, and integrating them into a single package. It is a smart move by Freescale, since it cannot be emulated by many competitors, and certainly not by the new breed of IoT SoC specialists. When the firm’s acquisition by NXP is complete, there may be even more opportunities to include more components in the packages. Integration will be more important to most IoT chips than cutting edge design, reinforcing trends which are already affecting the handset mass market, and so the ability to make and combine everything from processors to sensors, plus logic and interconnects, is a significant advantage. Jim McGregor, founder of semiconductor analyst firm Tirias Research, expects Freescale to tap more of its activities soon, adding RF modules for WiFi and BLE into future i.MX 6 models.

The power consumption will be the major challenge – as everywhere in the IoT. The i.MX 6D can only target mains-powered IoT and graphics devices, since it consumes 3-5 Watts. But the family will be extended shortly into “portables, wearables, autonomous sensing nodes, and automotive applications”, system solution VP Nancy Fares told EETimes. At its Forum, Freescale did demonstrate samples of the next generation chip for mobile IoT gadgets.

“The current i.MX 6D is only suitable for AC powered devices like household sensor hubs, or portables with big batteries like e-readers, IoT drones or larger wearables like virtual reality glasses,” McGregor said. “But potentially, Freescale could challenge Apple in integration by putting everything in the same package for low power devices, like smartwatches and fitness devices.”

The SCM i.MX 6D is slated for delivery in August along with its development system and evaluation board preloaded with either Linux or Android from Arrow Electronics.

Another approach to the hinterland between MCUs and processors is to offer a heterogeneous platform combining both. Freescale has achieved this too, with its i.MX 7 app processor. This runs on a mixture of high speed MPUs and low power MCUs, either of which can be turned on and off depending on the task being performed, to deliver the best performance:power ratio. This approach has been taken in mobile processors – by TI in the days of OMAP, for instance – but not in the embedded world. The i.MX 7 contains a single- or dual-core ARM Cortex-A7 and a single 288MHz Cortex-M4 microcontroller. Other on-chip domains can also be switched off and on as required. The current generation is 32-bit, but a 64-bit version is expected to be unveiled later this year. The i.MX 7, which is being evaluated for up to 10 designs including ereaders, wearables, and smart home or industrial control devices, said Freescale.

Meanwhile, another important trend in IoT chip integration is the convergence of the microcontroller with the connectivity. Freescale announced a new family of microcontrollers within its Kinetis range, with integrated BLE and/or 802.15.4 connectivity. These are targeted at home automation and industrial control devices. “We’re having a convergence between our microcontrollers and connectivity devices. In the IoT market this is an absolute requirement,” Emmanuel Sambuis, the company’s director of MCUs, told EETimes. “We will release more MCUs with connectivity interfaces. This is similar to what happened in the past where devices began to have a USB interface.” Of course, as in other areas of IoT chips, there are trade-offs to tight integration. Some packaging types can impair RF performance, for instance.

Freescale’s KW40Z runs on an ARM Cortex-M0+ processor running at 48MHz with an integrated 2.4 GHz multi-protocol radio that can access one BLE and one 802.15.4 network. The KW30 is a BLE-only device and the KW20 just runs 802.15.4. The single-mode BLE chip is particularly targeted at chips for wearables, while Sambuis believes 802.15.4 (which underpins both ZigBee and Thread) will dominate the smart home and industrial automation because of its mesh capabilities and ultra-low power levels. Where connections to both protocols are required, the multimode MCU runs concurrently with two modems and one analog front end.

TSMC and even SMIC are closing process gap with Intel:

The major foundries are racing to stabilize their 14/16nm processes and secure deals with vital customers like Apple and Qualcomm. In this generation, they will have to contend with Intel, whose CEO Brian Krzanich is determined to develop a foundry business, leveraging the company’s vast investment and expertise in process technology. They will also have to contend with the rising stars of China, making this a crowded field vying to make the chips of 2015 and beyond.

Earlier this year, it seemed that market leader TSMC was being squeezed in the low geometries, by Intel’s entry into the field with its 14nm FinFET process, and Samsung’s licensing deal with GlobalFoundries for its own 14nm process. But the Taiwanese giant bounced back, promising 10nm manufacturing by the end of 2016, which would be in the same timeframe as Intel’s 10nm introduction.

TSMC has also announced plans for a compact, low power version of its 16nm FinFET process (16FF+), which will be in production in mid-2016 and promises to deliver 10% better performance and 50% lower power consumption than competing silicon, for lower end handsets and wearables.

The firm indicated the vital importance of the low level processes when it said recently that its 10nm nodes would account for 55% of its revenue by 2020. That was a stunning admission, given that no single node has ever accounted for more than 40% of TSMC’s revenue (its most successful, 28nm, peaked at 37% last year), and it shows how far the company’s future hopes are pinned on the IoT. Liu said in April that 10nm would be the technology of choice for low power chips for connected devices, and that it would be a node with wide applicability, relevant to current customers who make communications, graphics and applications processors, and also to new IoT specialists. In the last quarter of 2014, 28nm was TSMC’s leading source of wafer revenue, according to its own data, delivering 30% of the total, followed by 20nm with 21%, 150nm+ with 17%, 40nm with 13% and 65/55nm on 11%.

By focusing so heavily on 10nm, and accelerating its progress so it can launch in the same timeframe as Intel, TSMC is distracting attention from its pressures in 14/16nm. Here, analysts believe Intel has achieved higher transistor density, and that it will take another year for TSMC to complete a full shrink from 20nm (which will involve its most aggressive shrink ever, reducing transistor area by at least 70%). However, Intel has pressures of its own. It is accustomed to buying a lead of at least a year over other chip manufacturers, thanks to its $2.5bn annual R&D spend. But Samsung got to 14nm FinFET only six months after its US rival, and TSMC may narrow the gap further at the 10nm stage. That would affect both the competitive positioning of Intel’s own chips, and its future foundry business.

And other foundries are coming up behind. UMC is the world’s second largest, but has been overshadowed in the mobile market. Now it has skipped the 20nm node altogether in order to catch up with its rivals in 14nm and be a credible player for mobile and embedded silicon again. It recently validated an ARM Cortex-A core on its 14nm FinFET process and says that will tape out later this year.

Even more concerning to Intel, Samsung and TSMC will be Qualcomm’s deepening alliance with China’s top foundry SMIC, which could help that firm enter the global top tier in future. Qualcomm has been seeking second sources for chips since problems at TSMC hit its Snapdragon supplies in 2013, and by working with SMIC, it not only adds an eager customer to its list, but helps appease Chinese government hostility. It has been investing in the company to help it get its fabs on a par with more established foundries, and the two companies, plus Huawei and the Belgian IMEC research institute, have announced a joint venture to develop 14nm process technology.

This will not be operational in SMIC fabs until 2020, by which time TSMC and the others will have 10nm well underway, but this nevertheless reduces the timelag between tier one manufacturers and Chinese foundries from two generations to just over one, and those timescales are sure to shrink further, especially with the help of funds and expertise from the likes of Qualcomm. Ironically, that influx of knowhow will also come partly from Intel, which is also working closely with Chinese firms, although it also hopes to offer fab services in the country. Rockchip said recently that Intel’s process technology was a key reason it formed a close SoC alliance with the US giant.

“SMIC has been a slow follower, an underdog of UMC that is an underdog to TSMC, and this puts them back in the front line,” said Malcolm Penn of market analyst firm Future Horizons. “It’s impossible to chase this foundry market, you have to intercept it and this sounds like an intercept strategy to me.”

“We believe that this collaboration will consolidate the IC domain, increase its resources and capabilities, and thereby improve the overall level of China’s IC industry,” said Steve Chu, a VP at Huawei in a statement.