Page 50 - EE Times Europe Magazine | April2019

50 EE|Times EUROPE — Boards & Solutions Insert

MICROCONTROLLERS
Adapting the Microcontroller for

AI in the Endpoint

By Sally Ward-Foxton

hat do you get when you cross AI capital investment
with the IoT? The artificial intel- is increasing, as are
ligence of things (AIoT) is the startup and M&A
Wsimple answer, but you also get activity, he noted.
a huge new application area for microcontrol- Today, the TinyML
lers, enabled by advances in neural network Committee believes
techniques that mean machine learning is that the tech has
no longer limited to the world of supercom- been validated and
puters. These days, smartphone application that initial prod- Used in tandem, Arm’s Cortex-M55 and Ethos-U55 have enough
processors can (and do) perform AI infer- ucts using machine processing power for applications such as gesture recognition,
ence for image processing, recommendation learning in micro- biometrics, and speech recognition. (Image: Arm)
engines, and other complex features. controllers should
Bringing this kind of capability to the hit the market in two
humble microcontroller represents a huge to three years. “Killer apps” are thought to be common: Arm. The embedded-processor–
opportunity. Imagine a hearing aid that can three to five years away. core giant dominates the microcontroller
use AI to filter background noise from con- A big part of the tech validation came last market with its Cortex-M series. The company
versations, smart-home appliances that can spring when Google demonstrated a version recently announced the brand new
recognize the user’s face and switch to their of its TensorFlow framework for microcon- Cortex-M55 core, which is designed specif-
personalized settings, and AI-enabled sensor trollers for the first time. TensorFlow Lite for ically for machine-learning applications,
nodes that can run for years on the tiniest of Microcontrollers is designed to run on devices especially when used in combination with
batteries. Processing the data at the endpoint with only kilobytes of memory (the core Arm’s Ethos-U55 AI accelerator. Both
offers latency, security, and privacy advan- runtime fits in 16 KB on an Arm Cortex-M3; are designed for resource-constrained
tages that can’t be ignored. with enough operators to run a speech key- environments.
However, achieving meaningful machine word-detection model, it takes up a total of But how can startups and smaller compa-
learning with microcontroller-level devices 22 KB). It supports inference but not training. nies seek to compete with the big players in
is not an easy task. Memory, a key criterion this market?
for AI calculations, is often severely limited, BIG PLAYERS “Not by building Arm-based SoCs, because
for example. But data science is advancing The big microcontroller makers, of course, [the dominant players] do that really well,”
quickly to reduce model size, and device and are watching developments in the TinyML laughed XMOS CEO Mark Lippett. “The only
IP vendors are responding by developing tools community with interest. As research enables way to compete against those guys is by hav-
and incorporating features tailored for the neural network models to get smaller, the ing an architectural edge … [that means] the
demands of modern machine learning. opportunities get bigger. intrinsic capabilities of the Xcore in terms of
Most have some kind of support for performance, but also the flexibility.”
TINYML TAKES OFF machine-learning applications. For example, XMOS’s Xcore.ai, its newly released cross-
As a sign of this sector’s rapid growth, the STMicroelectronics has an extension pack, over processor for voice interfaces, will not
TinyML Summit, a new industry event held STM32Cube.AI, that enables mapping and compete directly with microcontrollers, but
in February in Silicon Valley, is going from running neural networks on its STM32 family the sentiment still holds true. Any company
strength to strength. The first summit, held of Arm Cortex-M–based microcontrollers. making an Arm-based SoC to compete with
last year, had 11 sponsoring companies; this Renesas Electronics’ e-AI development the big guys better have something pretty
year’s event had 27, and slots sold out much environment allows AI inference to be imple- special in its secret sauce.
earlier, according to the organizers. Atten- mented on microcontrollers. It effectively
dance at TinyML’s global monthly meet-ups translates the model into a form that is usable SCALING VOLTAGE AND FREQUENCY
for designers has grown dramatically, orga- in the company’s e2 studio, compatible with Startup Eta Compute released its much-
nizers said. C/C++ projects. anticipated ultra-low-power device during the
“We see a new world with trillions of NXP Semiconductors said it has customers TinyML show. The ECM3532 can be used for
intelligent devices enabled by TinyML using its lower-end Kinetis and LPC MCUs for machine learning in always-on image-
technologies that sense, analyze, and autono- machine-learning applications. The company processing and sensor-fusion applications
mously act together to create a healthier and is embracing AI with hardware and software with a power budget of 100 µW. The chip uses
more sustainable environment for all,” said solutions, albeit primarily oriented around its an Arm Cortex-M3 core plus an NXP DSP core
Qualcomm Senior Director Evgeni Gousev, bigger application processors and crossover — either or both of which can be used for ML
co-chair of the TinyML Committee, in his processors (between application processors workloads. The company’s secret sauce has
opening remarks at last month’s conference. and microcontrollers). several ingredients, but the way it scales both
Gousev attributed this growth to the devel- clock frequency and voltage on a continuous
opment of more energy-efficient hardware STRONG ARM-ED basis, for both cores, is key. The approach
and algorithms, combined with more mature Most of the established companies in the saves a lot of power, particularly because it’s
software tools. Corporate and venture- microcontroller space have one thing in achieved without a phase-locked loop (PLL).

APRIL 2020 | www.eetimes.eu