Page 55 - EE Times Europe November 2021 final

EE|Times EUROPE 55

The Status of AI at the Edge? It’s Complicated

cooling fan or other thermal monitoring. These Edge AI is better suited to the neuro- memory and sends events to other NPUs over
hardware requirements are fairly substantial, morphic architecture, which processes only a mesh network, without host CPU inter-
given the limited space we have to work with. events for both CNNs and spiking neural vention. Small in size and high in density,
Then there’s the expense. They can’t be networks (SNNs). This architecture is far more NPU-based processors consume ultra-low
cost-prohibitive to design — which would be power-efficient, as only relevant information power — in micro- or milliwatts — and also
likely if we had to modify and resize all those is identified as a notable event or “spike.” alleviate some of the device space constraints.
internal components to fit — and they can’t That airport security camera shouldn’t bother Compared with edge-to-cloud approaches,
be cost-prohibitive to buy, especially in large processing hours and hours of images in AI-capable edge devices won’t need reliable
quantities for industrial or commercial use. which nothing in its field of vision is moving internet connectivity, which is in short supply
Among these complicated tasks, the most — it should cut to the chase. SNNs can also in the field and in transit. They also have
challenging might be working within an learn on the fly as they receive new informa- potential cybersecurity advantages because
extremely low power budget. Running neural tion, without the taxing retraining of CNNs. they aren’t sending data offsite.
networks consumes power resources, and If a device is expected to adapt constantly to So there you have it: difficult, but not
at the edge, power is precious. Traditional changes in its environment, SNNs in the neu- impossible. Thorny, but becoming more
convolutional neural networks (CNNs) are romorphic architecture, converted from CNNs practical all the time as AI technologies and
notoriously compute-intensive, which trans- or true SNNs, have a clear advantage. (You device hardware evolve.
lates to power-intensive. Compounding this might say that SNNs have the edge.) If you keep costs from spiraling out of con-
inefficiency is that CNNs intake the data and Another way to maximize efficiency is trol and design around power efficiency, edge
perform a repetitive, sequential training pro- offloading tasks from CPU resources onto AI just might become a real thing. ■
cess. Any time there is new data to learn, they neuromorphic processors based on cores
have to start anew and go through a massive called neural processing units (NPUs). Each Rob Telson is vice president of worldwide
retraining operation. NPU incorporates its own compute engine and sales at BrainChip Holdings Ltd.

OPINION | PHOTONICS

Silicon Photonics in the medical equipment market. But the
medical market will probably not generate
Sticks Its Head photonics revenues until 2024. Potential
applications include treatments for diseases
including diabetes.
Today, dynamized by cloud applications
Above the Parapet for home office and personal use, video
on demand, and 5G expansion, the pri-
mary silicon photonics application is still
By Eric Mounier and Alexis Debray, Yole Développement (Yole) optical communication, with the technol-
ogy integrated into 25% to 30% of optical
transceivers. Some applications, such as
Yole initially reported on silicon immunoassays (Genalyte) and fiber-optic
photonics applications in 2011. It is gyroscopes (KVH), will continue to grow,
1
interesting to compare our vision at that while LiDAR and photonic computing appli-
time with what is happening today. cations are emerging markets.
In 2011, silicon photonics was still an Consumer health applications are gain-
emerging technology, with only ing in importance with the release of
two industrial players: Luxtera and smartwatches that include an expanding com-
Kotura. At the market level, it was plement of sensors. Silicon photonics is also
obvious that datacom would be the expected to be integrated into other wear-
primary market for silicon photonics, though the medical sector had already been identified as ables, such as earphones. As in LiDAR, silicon
an interesting opportunity. photonics will enable compact and affordable
At the start of the 2010s, silicon photonics suffered from a lack of industrial infrastructure for optical modules.
design and foundry activities. When Luxtera and STMicroelectronics announced a partnership The consumer health application today
early in the decade, it was seen as a first step toward setting up a foundry service dedicated to is concentrated in the collaboration that
silicon photonics. The total market for silicon photonics at that time was valued at US$65 million Rockley Photonics started with Apple in
(mainly for datacom). 2017. Apple remains an important client of
In 2021, the industrial and market landscape for silicon photonics looks far different. While Rockley, with US$70 million of non-recurring
datacom and then telecom have long been considered the most important silicon photonics engineering commitment to date. The fitness
markets, Rockley Photonics’ announcement of silicon photonics technology’s use for consumer market is part of Apple’s strategy; its Apple
applications has changed this vision. Fitness+ service is integrated into the Apple
Rockley recently expanded the range of possible applications for its non-invasive biomarker Watch and offers various exercise apps, such
sensing into new medical technology segments. The company has signed strategic partner- as Pilates, yoga, and muscle strengthening.
ships with two of the world’s 10 largest medical equipment and device manufacturers (one of The Apple Watch can measure heart rate and
them is Medtronic); together, the two companies represent more than US$40 billion of revenue performance in real time.

www.eetimes.eu | NOVEMBER 2021