Embedded World Review

AI and software were strong themes at embedded world 2024

By Philip Ling

This year’s embedded world exhibition?and conference received over 32,000 visitors and hosted more than 1,100 exhibitors in Nuremberg, Germany. These figures surpassed last year’s and are a clear demonstration of a strong electronics industry.

As part of that industry, Avnet was present in the form of Avnet Abacus , Tria , Avnet Silica , EBV Elektronik and Witekio .

Alongside many new product introductions were two strong themes at this year’s event. The first, AI, could be felt everywhere. The second, software, remains the unseen enabler of the embedded sector. But the profile of embedded software was raised this year thanks to interest in the Cyber Resilience Act (CRA).

Let’s start with AI. It feels wrong to say AI is the next big thing because it has well and truly arrived. That was apparent from the number of AI-based demonstrations on show from semiconductor manufacturers and distributors, including Avnet. But there were also many exhibitors whose entire business model is based on AI in an embedded application. Specifically, AI at the edge.

Edge-based inferencing isn’t new, the industry has been moving in this direction for several years. Some might say edge AI has always been the goal, rather than becoming a new objective. The current interest is in moving other parts of the AI value chain to the edge, too.

Moving AI training to the edge

Generative AI, typically always carried out in the cloud, can be moved to the edge if the hardware supports it. There were few examples of how this is possible today. Cedric Vincent , a software engineering manager with Tria , offered one. His demonstration used a combination of speech-to-text and text-to-speech software to listen to questions and provide verbal answers. The responses were created using generative AI in the form of LlaMA, an open-source foundation language model, running locally. Vincent trained the model on information about Avnet Embedded. The demonstration runs on a COM-HPC module, understanding and replying to simple, spoken questions such as “How long has Avnet Embedded been trading?”

Training AI at the edge is another area of interest. The concept was covered in an expert panel that took place during the event in the exhibitors’ forum. Training at the edge is attractive for several reasons. Cloud-based training typically employs large amounts of processing power consuming large amounts of energy. Edge-based training would not be as quick, but it could consume less power. Privacy is another reason for training at the edge, as the data would not need to be sent to the cloud. Performance is the trade-off, which is why edge-based training would require new hardware solutions.

Power dissipation could be the deciding factor when it comes to edge AI. According to Gianluca Filippini , FAE and machine learning specialist at EBV Elektronik , while there’s a trend to leverage large AI models using powerful hardware, the industrial sector still needs compact, low-power solutions. Filippini’s demonstration was based on a System-on-Module (SoM) called Astrial developed by System Electronics, an EBV supplier. The SoM features an i.MX 8M Plus microcontroller from 恩智浦半导体 and a Hailo 8 neural network processor. The demo uses AI, accelerated by the SoM, to detect a person in front of the 安森美半导体 AR0821 HDR image sensor. The person then uses their hands, moving in free space and seen by the system, to control a small robot. System Electronics worked with Kalpa to develop the demo and the code is available here.

Filippini explained that OEMs want to avoid active cooling and fans wherever possible. “Using advanced cooling systems or bulky heatsinks is often a limiting factor, in terms of cost and weight,” he said. “For these reasons, the ‘sub-10W’ threshold is becoming a line that can’t be crossed for many edge AI applications.” ?

We are already pushing that limit. New, low-power solutions will be necessary to move more AI features to the edge without crossing the 10W limit.

What you need to know about cyber resilience

Software has long been a fundamental of embedded development but this year the Cyber Resilience Act took the spotlight. Participants in an expert panel expressed strong opinions on this. The advice offered wasn’t limited to technical insights. From a business point of view, OEMs may need to start charging for software updates to fund software maintenance once a product is in the market.

In some cases, the software stack may become too difficult to maintain, meaning some products may be withdrawn from the market due to the challenge of software updates. Security isn’t a bug, they said. Even if an OEM could develop software that has zero bugs, it will still be susceptible to future cyber threats.

At a minimum, panelists stated that new products coming to market and subject to the CRA should have secure boot and a way to implement secure over-the-air updates.

Avnet Silica and Witekio were talking to engineers about CRA during the event, and Avnet Silica is currently delivering a series of CRA seminars. You can register for your local seminar here: Cyber Resilience Act Seminars | Avnet Silica?

These two themes dominated this year’s conference and we can expect them to be just as important when the company behind the event takes it to China and North America for the first time later this year.

Access many of the sessions recorded at this year’s event by visiting the organizer’s landing page.