Tool chains are the bottleneck
Edge AI needs software, not just powerful hardware
In an interview with Markt & Technik, Rich Simoncic, Microchip Technology's COO, explains that edge AI tends to fail more due to a lack of software integration than due to hardware issues. He says that clear use cases, automated tool chains and simple APIs are crucial, rather than high TOPS.
|
embedded world Conference 2026 »Learning from the Octopus: Nature's Blueprint for Intelligence Everywhere« |
|---|
|
This is the title of Rich Simoncic's keynote speech at this year's Embedded World Conference. Simoncic is the COO of Microchip. He believes that the octopus is a prime example of the power of distributed intelligence. This is because most of the octopus's neurons are located in its arms and skin, enabling them to perceive and act independently. This principle also applies to the embedded and edge worlds, where billions of networked devices must operate efficiently and securely with low energy consumption. In his presentation, Simoncic outlines the shift from central processing units to intelligent, distributed systems. He also addresses technical challenges such as software complexity, AI integration, development tools, standards and costs. His key message is that the future of AI lies not in ever larger ‘brains’, but in adaptive, secure and distributed systems modelled on nature. The presentation will take place on 10 March 2026 at 10:15 a.m. in the NCC East Convention Centre at the Nuremberg Exhibition Centre. This keynote speech is free to attend for all trade fair visitors and conference participants. |
Markt & Technik: Everyone is talking about AI at the edge. In which markets is AI at the edge currently used most frequently, and what are the biggest advantages of this approach at the edge?
Rich Simoncic: It's true, everyone is talking about AI at the edge, but I would say that in many cases it's simply a matter of trying it out. I think AI at the edge is not about some fancy device, but about the possibilities it offers. With AI at the edge, a model is trained on a large, powerful system and then transferred to a small device, which is then able to make decisions at the edge thanks to AI.
An example: one of perhaps 1,000 diffusion pumps in a semiconductor factory. These diffusion pumps are the main cause of equipment failure in a semiconductor factory. If these pumps were monitored for vibrations, heat and noise, it would be possible to detect when a pump is starting to cause problems and replace it before it actually fails. And that's just one very simple example of where AI at the edge really brings benefits.
Another example: large industrial fans. These also often break down. And if there were a way to detect when they are about to fail, they could be replaced during regular maintenance before they bring the entire factory to a standstill.
These are probably the biggest areas of application we are currently seeing: predictive maintenance of motors to detect when they are about to fail. This is particularly interesting when you consider how many motors are used in critical applications worldwide.
The second largest area for AI at the edge is extending battery life. We recently collaborated with a manufacturer of portable tools. We collected data from their handheld tools – specifically a cordless drill – and developed a model based on this data. We then transferred this AI model to a microcontroller that was already installed in the drill. This meant that no additional hardware was required. What is more important is the ability to collect data, train a model and transfer this model to the device in order to perform inferences there.
After we did this, the manufacturer was able to extend the battery life of these devices by 15 per cent.
However, we also realized the following: more than half of the applications we see actually only require relatively simple models. And these are then simply transferred to the component in the edge.
This is where the analogy with the octopus comes in. It has tentacles, each with its own suction cups, which represent a certain form of simple intelligence for deciding what to do. If you transfer this approach to technical systems, say a robot, then there is the central GPU, but also fingers, grippers, etc., which must be able to quickly decide whether to let go because something is hot or cold, or whether they are gripping something too tightly, or whether something is in the way. They must be able to decide quickly. And it is precisely this type of intelligence that is needed at the edge.
And what does that mean?
That AI is often misjudged and we started off doing this too. When we started thinking about AI at the edge, we thought we needed accelerators, NPUs and all those complex things. But it turned out that most of our customers just need help.
With what?
Collecting data, creating models, and then porting those models into their application. And of course, there is room for both: simple and complex solutions. Sure, everyone wants those ‘sexy’ devices, but that's not really the point. Many developers simply need help creating a model that will help them solve a specific problem. And that's exactly what the edge is all about, and that's what Microchip focuses on.
How does Microchip support developers? Do you help them train the models and then port them to a component?
Yes, we help them collect the data, train the models, and then we return the model so they can load it onto their component.
You said that predictive maintenance is currently the largest area of application for AI at the edge. Will this picture change in the coming years?
Yes. Take energy consumption, for example. I am convinced that this area of application will use AI to save energy, as in the example of the cordless drill mentioned earlier. But when it comes to energy saving, the possibilities for AI at the edge are extremely diverse.
This could include, for example, thermostats for the home or hot water heating. AI recognises user behaviour and controls when and how much heating is used accordingly, so that only the energy that is really necessary is consumed.
Similar to the cordless drill, there are many applications where 5, 10, 15 or even 20 percent energy savings can be achieved. In summary, I would say that sustainability and energy consumption will be the next area to drive AI at the edge.
There is certainly enough demand for AI at the edge – but what does implementation look like, and what do you think are the biggest hurdles for developers?
I would say that the biggest problem for developers is that they don't know how to implement AI at the edge. There is simply no good API yet, no proper application front end that developers can access.
The idea is to feed in the data, perform inference in the cloud, create the model there, and then transfer this model to the component. But the average embedded control engineer – that is, the engineers we meet in almost every company – simply does not have the expertise to do this.
Do they lack expertise in using the tools or AI know-how?
The tools are simply not ready yet. That's why engineers have to go to experts who can help them create these models. A team at Microchip is working on developing tools that are as simple as possible for embedded developers.
Our goal is to make it possible for anyone to create these models – by simplifying the whole process to the point where all you need is a simple API to load data and use a cloud service to create the model. And then port the model back to your own device. Until this becomes easy, we won't see widespread use in the industry. At the moment, it's all still quite complicated – and that's exactly why it's not as widespread as it could be.
So the underlying hardware isn't the real problem? Many hardware manufacturers like to point to the number of TOPS, but that's not what AI at the edge is all about?
No, it's not about TOPS – even if many people don't like to hear that. Of course, the more TOPS, the higher the ASP, but for many applications, I don't even need an accelerator.
If there is an accelerator on the chip, then computing power of 50 GOPS, or even just 5 GOPS, is sufficient for many applications. Even a very simple, extremely small accelerator for distributed intelligence can already do much of what we need today – in fact, a large part of it.
If, on the other hand, I put a 200 TOPS or even 1000 TOPS accelerator on a chip, then we're talking about an intelligent edge device for around £15. And no one is going to install a £15 controller in a wall thermostat just to make it a little bit ‘smarter’.
That would be too expensive...
Exactly, if the task is to recognize a face or five to ten keywords, then we don't need a 100 TOPS accelerator; 20 to 50 GOPS are sufficient for that. If a light switch is turned on and off by voice or a television is switched over, then that should happen at the edge; no one wants that decision to be made in the cloud.
Perhaps it would be good if you could define AI at the edge more precisely?
For me, AI at the edge begins where an analogue signal needs to be perceived or measured, be it temperature, noise, vibrations – anything from the environment. That is the starting point. A decision then has to be made based on this signal.
After that, the data can perhaps be passed on to a hub that makes more complex decisions. And this hub can in turn be connected to the cloud, where significantly more computing power is available.
That means I basically see three levels: the cloud with large GPUs, then a hub with NPUs – perhaps 200 or 1000 TOPS – and at the very edge, a very small system that simply measures and reacts.
Many providers emphasise that they support Python, ONNX and other common tools/frameworks, including model zoos, which actually seems quite mature from a software perspective. They criticise the lack of a simple API. So what is really crucial for developers?
Python is important because most AI developers today work with high-level languages. But the real problem is that there are many individual tools – for training, edge AI and deployment – that don't work well together.
And while model zoos help a little, they don't solve the core problem: every application and every dataset is different. Even if there is a suitable model, it has to be readjusted with your own data.
That sounds like a lot of effort…
Exactly. Today, there is no automated environment that shows how data is collected, fed into the model and adapted. That's why edge AI is not yet widespread.
And Microchip is working on that?
Yes. Our goal is to create a software platform that greatly simplifies and automates this process. We hope to provide initial solutions within the next year – deliberately focusing on software, not just hardware.
Unusual for a hardware company?
Good hardware is useless without good software. That's why we are investing heavily in software and integrating solutions from partnerships and acquisitions to bring customers to market faster.
Last question: Will learning directly at the edge be possible in the future?
Basically, yes. But that requires more computing power. If models are to continuously adapt at the edge, the TOPS discussion becomes relevant again. It's hard to say exactly when that will happen, but technology is evolving faster than ever.
