Embedded World 2019 vs 2026: What Changed in Edge Hardware, Software, and Real Products
What stood out to me at my first Embedded World in March 2026 was not just the scale of the event, but how perfectly it captured the shift across embedded systems over the last seven years. Thinking back to similar embedded exhibitions I attended around 2019, bringing intelligence to the edge still felt defined by hard trade-offs. If you wanted meaningful compute on-device, you usually had to compromise on power, latency, model size, or deployment complexity. In 2026, that balance felt completely different. The industry is no longer asking whether intelligence can move to the edge; organizers and exhibitors alike are demonstrating that it has already happened.1 3
Hardware Built for the Edge, Not Just Adapted to It
The clearest signal of this shift is in the silicon. In 2019, edge AI often felt shoehorned into constrained systems. Today, heterogeneous computing is the mainstream default, with CPUs routinely paired with dedicated NPU, DSP, or GPU accelerators specifically for on-device AI.
Here is how that evolution looks across the hardware spectrum:
- The High-Performance Edge (Qualcomm): The conversation has moved far beyond basic computer vision. Qualcomm’s Dragonwing IQ8 family perfectly illustrates this shift by using “LLM-native” metrics. They market the platform as delivering up to 40 dense TOPS and proudly publish token-per-second benchmarks for running a 13-billion parameter Llama 2 model entirely on-device 4.
- The Ultra-Low-Power Edge (STMicroelectronics): Energy budgets are no longer the ultimate blocker for smart workloads. ST’s new STM32U3 series utilizes near-threshold design to drastically reduce dynamic power consumption, bringing sophisticated capabilities to battery-reliant devices 5.
- New Computing Paradigms (Innatera): Neuromorphic computing has graduated from research to reality. Innatera’s spiking-neural approach focuses on always-on sensing, achieving upto 500x lower energy and 100x shorter latency than conventional processors.Their radar-based human presence detection runs at sub-milliwatt power while maintaining over 99% accuracy 6.
Software Optimisation Tooling Has Become a Core Part of Edge AI
Software tooling has matured right alongside the silicon. Optimization is no longer a late-stage engineering hurdle; it is a dedicated product category.
- Incredibuild: tackles bottlenecks at the developer level by distributing compilation and custom build steps across available network resources, shrinking the time it takes to iterate and ship software 8.
- Nota AI’s NetsPresso: treats model deployment as a hardware-aware workflow 9. Their pipeline handles compression and benchmarking to bring models to virtually any device, explicitly supporting Large Language Models (LLMs) and Vision-Language Models (VLMs) 10 11.
- ENERZAi: takes optimization a step further with its “extreme quantization” approach 15. Using its Optimium engine, the company can generate 1.58-bit kernels that reduce AI model memory usage by over 77% and boost processing speed by nearly 2.5x. This is especially notable in the broader context of recent research interest in 1-bit LLMs 16, where ultra-low-bit approaches are being explored as a practical way to run high-performance voice and language models entirely on-device, even on constrained targets lacking dedicated AI chips.
Real Solutions, Real Context
What made this year’s show so convincing was the presence of credible and deployable systems in action:
A great example is the Arduino VENTUNO Q. By combining a Dragonwing IQ8 processor with an STM32H5 microcontroller, it offers a clean, developer-friendly platform that merges AI compute with deterministic control for robotics 12. Meanwhile, ST demonstrated how a machine learning algorithm can be powered entirely by ambient light. Using their new STM32U3 equipped with an HSP hardware accelerator and printed organic photovoltaic modules, they successfully ran a person-detection camera setup completely battery-free 14.
The Bottom Line
For me, Embedded World 2026 proved that the generative AI wave of recent years didn’t just introduce new models—it forced real, tangible engineering progress across the entire embedded stack.
Ultimately, what I observed in Nuremberg reinforces the exact trends I have been tracking over the past few years: rapid embedded hardware evolution, the rise of purpose-built AI optimization tools, and the deep penetration of these solutions into everyday physical products.
References
- [1] Embedded World 2019 breaks new records in exhibitor numbers
https://audioxpress.com/news/embedded-world-2019-breaks-new-records-in-exhibitor-numbers - [2] Embedded Award 2026, Products Honoured
https://www.embedded-world.de/en/press/press-releases/2026/03/embedded-award-2026-products-honoured - [3] Embedded World 2026 Closing Report
https://www.embedded-world.de/en/press/press-releases/2026/03/embedded-world-2026-closing-report - [4] Qualcomm Dragonwing IQ8 Series
https://www.qualcomm.com/internet-of-things/products/iq8-series - [5] STM32 Ultra Low Power Microcontrollers (MCUs)
https://www.st.com/en/microcontrollers-microprocessors/stm32-ultra-low-power-mcus.html - [6] Innatera Brochure
https://innatera.com/storage/app/media/Resources/2025/Innatera-Brochure.pdf - [8] Incredibuild
https://www.incredibuild.com/ - [9] Nota AI, Embedded World 2026
https://www.nota.ai/ew2026 - [10] NetsPresso LLM
https://netspresso.ai/llm - [11] Nota AI, Deploying an Efficient Vision Language Model on Mobile Devices
https://www.nota.ai/community/deploying-an-efficient-vision-language-model-on-mobile-devices - [12] Arduino VENTUNO Q
https://www.arduino.cc/product-ventuno-q - [13] ENERZAi
https://enerzai.career.greetinghr.com/ko/home - [14] HSP: The new hardware accelerator that transforms an ultra-low-power STM32U3 into an AI machine
https://www.edge-ai-vision.com/2026/03/hsp-the-new-hardware-accelerator-that-transforms-an-ultra-low-power-stm32u3-into-an-ai-machine/ - [15] AI Lightweighting Competition: Enerzai’s Breakthrough on the Global Stage with 1.58-Bit ‘Extreme Quantization’
https://enerzai.com/resources/newsroom/ai-lightweighting-competition-enerzai-s-breakthrough-on-the-global-stage-with-1.58-bit-extreme-quantization - [16] The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits
https://arxiv.org/pdf/2402.17764