Fullscreen HTML5Med Res (??MB)HTML4

More power? No – less! Techno Talk From battery-powered AI to nuclear-fuelled data centres, the race for connectivity and intelligence is becoming a race for energy efficiency. If there’s an elephant in the room or a fly in the soup (I never metaphor I didn’t like), it’s that it’s becoming challenging to write a technology-­related column these days without AI popping up in one form or another to bid us a cheery “hello”. Since I don’t like challenges (it’s not in my vocabulary), I’m not even going to try to hold back the tide. Don’t forget the battery! How much power things consume is important, especially in remote and portable devices, where the internet ‘rubber’ meets the real-world ‘road’. Things get a little hairy when we talk about AI-enabled, battery-powered wearables like smartwatches and augmented reality (AR) glasses. It’s only getting worse with features like conversational user interfaces that let users interact with devices via spoken natural-language commands. This all reminds me of an old joke about a somewhat absent-minded professor type character staggering down a busy city street, lugging an enormous and clearly very heavy suitcase. A trio of well-heeled young financiers observe him approaching, and their leader decides to have a bit of fun. As the old fellow draws level, the leader calls out, “Hey granddad, do you have the time?” With much grunting and wheezing, the man lowered his suitcase and pulled back his jacket sleeve to reveal a spaceage contraption strapped to his wrist. Upon a spoken command, a high-resolution, full-colour 3D hologram of the Earth springs into existence, complete with real-time clouds and other atmospheric effects, slowly rotating and displaying the time across every time zone. The snotty young men stop laughing and gawk, mouths open in wide-eyed astonishment. “I must have that,” says the leader at last. “I’ll write you a million-­ pound cheque here and now!” The professor pauses for a moment, then shrugs. “Well, that would certainly fund my other experiments… all right.” Cheque and watch duly exchanged, the three young men set off down the street. “Just a moment!” calls the professor after them, gesturing at the suitcase. 66 “Don’t forget the battery!” How low can you go? The reason the wristwatch example springs to mind is that I was just chatting with the folks at Ambiq (www. ambiq.com). Their claim to fame is ultra-­low-power processors that enable edge AI in billions of battery-powered devices—everything from wearables and smart glasses to smart cameras and industrial sensors. The lower the supply voltage, the lower the power consumption. Most state-of-the-art application processors in devices like smartphones and wearables typically operate in the 0.7–1.2V range. Ambiq dares to be different with its subthreshold power-optimised technology (SPOT). Ambiq’s forthcoming Atomiq device will boast an Arm Cortex-M55 CPU and an Arm Ethos-U85 NPU core dedicated to AI acceleration, both running with core voltages of just 0.3V. As far as I’m concerned, this is skirting the boundary between engineering and black magic! What makes this particularly interesting is the class of applications Atomiq unlocks: always-active devices that don’t need to be tethered to the cloud or constantly recharged. Think conversational AR glasses that can respond to spoken commands all day, health-monitoring wearables that never sleep, battery-powered cameras performing real-time object recognition, or industrial sensors quietly applying AI at the edge without anyone having to worry about where the nearest power outlet is. Max the Magnificent Feeling connected Not long ago, apart from telephones that were firmly tethered to the wall, things largely operated in splendid isolation. Early embedded systems did their jobs quietly and efficiently, but they did them alone, with no expectation of communicating with anything else. A thermostat controlled the heating, a smoke detector monitored the house; that was pretty much the end of the story. Today, we increasingly expect everything to be connected all the time. Using technologies like Wi-Fi, Bluetooth, Thread and Zigbee, devices are expected to discover each other, communicate and collaborate seamlessly. Once everything is connected, the next question becomes: what do we do with all that data? This is where companies like Synaptics (www.synaptics. com) come into play. Synaptics explores ways to make devices more intuitive. That involves embedding intelligence into everyday interfaces and quietly reimagining how humans engage with machines and data. Synaptics describes the SYN765x as a Wi-Fi 7 AI-native connected microcontroller, a device designed from the ground up to combine high-performance wireless connectivity with on-device AI processing. That makes it perfect for smart hubs, security cameras with local object recognition, or even industrial nodes making decisions without waiting for the cloud. From local links to global reach Connectivity doesn’t stop at Bluetooth and Wi-Fi. Today’s smartphones Practical Electronics | June | 2026 seamlessly hop between networks, typically preferring Wi-Fi when it’s available and falling back to the mobile network when it’s not. However, there are still plenty of places where terrestrial coverage is patchy at best: remote regions, oceans, mountains and the occasional stubborn ‘not spot’ in the middle of a city. This is where low Earth orbit (LEO) satellite constellations enter the fray, extending connectivity far beyond the reach of traditional infrastructure. It won’t be long before access to these non-terrestrial networks (NTNs) becomes commonplace. Future devices are likely to treat connectivity as a hierarchy: try Wi-Fi first, then the mobile network, and finally satellite if all else fails. The result is a world in which we (and our devices) are effectively never out of touch, with a continuous thread of communication stretching from the living room to the most remote corners of the planet. At present, terrestrial 5G networks and satellite-based systems are still somewhat different beasts, each with its own standards, constraints and challenges. CEVA (www.ceva-ip.com) recently announced its PentaG-NTN platform, which is designed to enable seamless integration between mobile and satellite communications. By providing the building blocks for multi-mode devices that can operate across both terrestrial and non-terrestrial networks, companies like CEVA are helping to turn the vision of truly ubiquitous connectivity into reality. Powering the cloud Power consumption is becoming a major concern in the cloud and in data centres, where the numbers can get eyewateringly large very quickly. Training AI is the process of building the model in the first place—feeding it vast quantities of data and adjusting millions or billions of parameters until it behaves as desired. This is Practical Electronics | June | 2026 computationally intensive and can consume enormous amounts of energy, often requiring thousands of GPUs running for days or even weeks. After the model has been trained, while the system is actually used—recognising images, understanding speech, generating text and so on—the energy usage is far lower each time. However, this usage is multiplied by the number of inferences made, which can be in the millions or even billions. So while training represents a large, one-off (or at least occasional) energy investment, making inferences is where the real long-term power consumption lies. Every time someone asks a question, issues a voice command, or triggers an AI-enabled device, more power is used. And that adds up. The scale of this demand is already reshaping the infrastructure that supports it. Some companies are even exploring small modular reactors (SMRs)—essentially compact nuclear power stations—to keep their AI data centres fed. When nuclear-powered data centres start to sound like a sensible option, you know things have escalated rather dramatically. How much? I designed my first application-­specific integrated circuit (ASIC) back in the dark ages we used to call 1980. This was at the five-micron node, meaning the smallest structures we could create on the surface of the silicon die (chip) were 5 millionths of a meter across. Today, people routinely create chips at the five-nanometre node (and below), which means transistors are around 1000 times smaller and consume correspondingly less power. Well, sort of… The feature sizes aren’t quite as small as 5nm might suggest, but they are still very tiny. ASICs in the 1980s contained only a few thousand transistors (if you were lucky). By comparison, a high-end AI processor today can contain tens to hundreds of billions of transistors. So, while each transistor may be far more efficient than its ancestors, we now use vastly more of them. Modern AI accelerators can each draw 500-1000 watts. Pack eight of these into a server and you’re looking at several kilowatts per box. Fill a rack with such systems, and you can easily reach 20–50 kilowatts per rack. Multiply that by tens of thousands of racks in a hyperscale data centre, and you begin to see why power consumption has become such a pressing concern. When less is more Until recently, I would have said that a custom ASIC offers the highest possible performance while consuming the least possible power. That’s still true as far as it goes, but there’s a wrinkle. A custom ASIC accelerator, like a GPU, is designed to handle a wide variety of workloads, including AI training and inference, as well as other tasks. Another way to look at this is that the GPU has been engineered to be as general-purpose as possible, but this translates into being a ‘jack of all trades and master of none’. As a result, performing inference using a GPU can consume significantly more power than one might hope. This is where ElastixAI (www.elastix. ai) comes in. Their approach is based on hardware tailored specifically for AI inference workloads. That means leveraging field-programmable gate arrays (FPGAs) configured to implement the precise dataflows required by large language models (LLMs) and other AIs. By mapping the computation more directly onto the hardware and avoiding much of the overhead associated with more general-purpose architectures, these FPGA-based solutions can deliver significantly higher performance per watt than GPUs. This is especially true for techniques such as mixture-of-experts (MoE) models, in which only a subset of the network is active at any given time. In a world where every watt counts, that’s the sort of idea that tends to attract one’s attention. PE 67