IoT and Edge AI

New Trends in Embedded Systems and Software Engineering

The hardware and software trends reshaping embedded systems: on-device AI, RISC-V, security regulation, Rust, Zephyr, modern testing and AI-assisted coding.

On this page
  1. The short version
  2. Hardware and platform trends
  3. Software engineering trends
  4. Which trends matter for you?
  5. Is AI replacing embedded systems engineers?
  6. FAQ

The short version

Embedded systems are changing on two fronts at once. The hardware is getting more capable: microcontrollers now ship with neural accelerators, RISC-V has become a credible alternative to Arm, and wireless chips keep absorbing more functions. At the same time, the engineering practice around that hardware is maturing: security is turning into a legal requirement in some markets, devices are expected to receive updates for years, and teams are adopting tools and languages that used to be rare in firmware.

The main trends, in brief:

  • AI on the device, from TinyML on ordinary microcontrollers to MCUs with built-in neural processing units (NPUs)
  • RISC-V moving from experiments into shipping products
  • More integrated chips with radios, security hardware and multiple cores on one die
  • Connectivity consolidation around standards such as Matter, Thread, Bluetooth LE and cellular IoT
  • Security by regulation, led by the EU’s Cyber Resilience Act
  • Long-lived devices that need over-the-air updates and ongoing maintenance
  • Open RTOS platforms such as Zephyr gaining ground
  • Memory-safe languages, especially Rust, being used alongside C and C++
  • Modern tooling: continuous integration, simulation and hardware-in-the-loop testing
  • AI-assisted development, with all the review it requires

None of these replaces the fundamentals. C, microcontroller peripherals, interrupts, timing and debugging with real instruments are still the core of the job.

AI moves onto the microcontroller

For years, “AI in an embedded device” usually meant sending data to the cloud. That’s shifting. TinyML (running small machine learning models on microcontrollers) is now practical for tasks like keyword spotting, gesture recognition and vibration anomaly detection, often with models of only tens of kilobytes.

The bigger change is dedicated silicon. Arm’s Ethos-U NPUs are paired with Cortex-M cores in a growing number of MCUs, and several vendors offer their own neural accelerators. These run inference far faster and at much lower energy than the CPU alone, which opens up small vision and richer audio models on battery-powered devices.

What this means in practice: skills such as model quantization, memory budgeting (weights in flash, activations in SRAM) and building sensor pipelines become part of embedded work. Our guide to embedded AI and computer vision at the edge covers the hardware tiers, and Getting Started With TensorFlow Lite for Microcontrollers walks through a first model.

RISC-V becomes a mainstream option

RISC-V is an open instruction set architecture. Anyone can design a processor that implements it without paying an ISA licence fee. It has moved well beyond research boards. Espressif’s ESP32-C series, for example, uses RISC-V cores, and RISC-V cores are common inside larger chips as management or helper processors.

Standard profiles such as RVA23 give software vendors a more stable target, and vendors use custom extensions to tune cores for signal processing or AI.

A realistic view: Arm Cortex-M still dominates general-purpose microcontrollers thanks to its tooling, middleware and certified ecosystem, and that isn’t likely to change quickly. RISC-V is best seen as a second architecture worth being comfortable with. Concepts like registers, interrupt controllers and privilege modes carry over between the two.

More capable, more integrated chips

Microcontrollers that once had kilobytes of memory and a few tens of megahertz now commonly offer hundreds of megahertz, megabytes of flash and RAM, DSP instructions and hardware crypto. Other integration trends include:

  • Wireless SoCs that combine the processor and one or more radios on a single chip, sometimes supporting several protocols
  • Heterogeneous multicore parts that mix application cores, real-time cores and accelerators in one package
  • Hardware security blocks such as TrustZone for Cortex-M, secure key storage and secure boot ROMs
  • Newer non-volatile memories like MRAM and RRAM appearing in advanced-node MCUs, often as a replacement for embedded flash

More integration lowers cost and board size, but it also makes debugging and configuration more complex. Multicore and mixed-criticality systems are a skill area of their own.

Connectivity consolidates

The wireless landscape is slowly converging:

  • Matter, an application-layer smart home standard that runs over Thread, Wi-Fi and Ethernet, aims to let devices work across the major smart home ecosystems.
  • Bluetooth LE keeps adding features, such as direction finding and, more recently, Channel Sounding for distance measurement.
  • Cellular IoT (LTE-M, NB-IoT, and increasingly 5G RedCap for mid-tier devices) covers wide-area devices without Wi-Fi.
  • LoRaWAN remains a strong option for low-data, long-range sensors.

For a practical comparison, see Wi-Fi vs Bluetooth LE vs LoRa vs Zigbee.

Security and regulation become requirements

This is arguably the trend with the most direct impact on product teams. The EU Cyber Resilience Act (CRA) sets cybersecurity requirements for products with digital elements sold in the EU. Its vulnerability and incident reporting obligations started applying in September 2026, and most of its other obligations follow in December 2027. Other regions have their own rules, such as the UK’s product security regime for consumer connectable products, which bans default passwords among other things.

In firmware terms, this pushes practices that were once optional toward standard:

  • Secure boot anchored in a hardware root of trust
  • Signed, rollback-protected firmware updates
  • No shared default passwords, and debug ports locked in production
  • A software bill of materials (SBOM) for each release, so you know which components and versions ship in a product
  • A vulnerability disclosure process and someone responsible for handling reports
  • Crypto agility: NIST published its first post-quantum cryptography standards in 2024, and devices meant to last a decade or more may need a path to adopt them

When devices control physical equipment, the stakes rise further. That’s the subject of What is cyber-physical security?

Devices are treated as long-lived software products

The old model was to ship firmware once and rarely touch it again. Connected devices are now expected to receive bug fixes and security patches for years. That changes how firmware is designed. Partition layouts, bootloaders and memory headroom all have to allow for updates, and teams need release processes that look more like web software than a one-off ROM. The mechanics are covered in Over-the-Air (OTA) Firmware Updates for IoT Devices.

Open RTOS platforms gain ground

Zephyr, an open-source RTOS hosted by the Linux Foundation, has attracted broad vendor support. Its devicetree and Kconfig configuration let one codebase target boards from many manufacturers, and it bundles networking, Bluetooth, file systems and security components. FreeRTOS remains very widely used, especially for smaller and simpler devices. At the high end, embedded Linux (often built with Yocto or Buildroot) is standard for gateways and devices with displays or cameras.

The shift is toward platform-based firmware: reusable drivers, standard build systems and layered architectures instead of a one-off codebase for every board.

Memory-safe languages: Rust next to C and C++

Memory-safety bugs such as buffer overflows and use-after-free errors are a major source of security vulnerabilities, and government security agencies have publicly encouraged a move toward memory-safe languages. In embedded work, Rust is the main candidate. It has an active embedded ecosystem (HAL crates, the Embassy async framework, RTIC), and qualified toolchains such as Ferrocene exist for some safety-standard work.

C and C++ aren’t going anywhere. They dominate existing codebases, vendor SDKs and certification history, and modern C++ use in firmware has grown steadily. A sensible pattern many teams follow is to keep proven C code and consider Rust for new, security-sensitive components like parsers and protocol handling.

Modern tooling: CI, simulation and hardware-in-the-loop

Practices that are routine in other software fields are becoming normal in firmware:

  • Continuous integration that builds every commit for every target
  • Static analysis and coding standards such as MISRA C, which are especially common in automotive and safety work
  • Unit testing on the host, with hardware access abstracted away
  • Simulation and emulation with tools like QEMU and Renode, so tests can run before boards exist
  • Hardware-in-the-loop (HIL) rigs that flash real devices automatically and run tests against them
  • Digital twins in larger systems, used to test control software against a model of the physical plant

AI-assisted development

AI coding assistants can draft drivers, register definitions, test scaffolding and documentation quickly. In embedded work, though, their output needs particularly careful review. Timing, interrupt handling, DMA, register sequences and power behaviour depend on specific silicon details that general-purpose tools often get wrong. Practical teams treat AI output like code from a new colleague: check it against the datasheet and reference manual, then run it through the same static analysis and tests as everything else.

Trend Student or hobbyist Firmware engineer Product team
On-device AI / TinyML Great project area Growing skill demand Product differentiator
RISC-V Try a board Learn the basics Supply-chain and cost option
Security and regulation Learn the concepts Secure boot, signing, key handling Compliance obligation in some markets
OTA and long-term maintenance Try OTA on an ESP32 Core skill for connected devices Must be designed in from day one
Zephyr / RTOS Learn one RTOS well Increasingly listed in job ads Platform and portability decision
Rust Optional, interesting Worth learning alongside C Evaluate for new security-sensitive code
CI, simulation, HIL Use Git and a debugger Expected in many teams Reduces field failures

If you’re starting out, focus on the fundamentals first: C, one microcontroller family, its peripherals (GPIO, UART, SPI, I2C, ADC, timers), datasheet reading, an RTOS and a debugger. Then add one trend that interests you, such as a TinyML project or an ESP32 MQTT project.

Is AI replacing embedded systems engineers?

There’s no reliable way to predict the job market years ahead, so treat confident claims in either direction with caution. A few observations are reasonably safe:

  • AI tools already speed up routine coding tasks, and that will likely change how embedded engineers spend their time.
  • Embedded work depends heavily on things current AI tools handle poorly without supervision: hardware-specific behaviour, timing, debugging with oscilloscopes and logic analysers, and safety judgement where a bug can damage equipment or hurt someone.
  • Responsibility doesn’t transfer to a tool. In regulated and safety-related products, a human engineer still has to understand, verify and sign off on the code.
  • New trends such as on-device AI and security regulation are creating embedded work, not only automating it.

The likely outcome is a shift in the skill mix rather than the disappearance of the role. Engineers who can review and verify code critically, understand the hardware deeply and work across firmware, security and connectivity are well placed.

FAQ

What is the 40/20/40 rule in software engineering? It’s a rough guideline for splitting project effort: about 40% on requirements, analysis and design, 20% on coding, and 40% on testing and integration. It’s a rule of thumb, not a standard, but it captures a point that holds especially well in embedded work: writing the code is often the smaller part of the job compared with design and verification.

Who earns more, VLSI or embedded engineers? It depends heavily on country, employer, specialisation and experience, so there’s no single answer. Chip design (VLSI) roles at semiconductor companies are often well paid, and so are senior embedded roles in areas like automotive, security and embedded AI. If you’re choosing between them, comparing current job listings in your own region will tell you more than general claims.

What are the four types of embedded systems? Embedded systems are commonly grouped as standalone, real-time, networked and mobile. Our pillar guide on embedded systems and the Internet of Things explains each type with examples.

Which new embedded skill is most worth learning? For most people, security fundamentals and over-the-air updates give the broadest payoff, because nearly every connected product now needs them. On-device AI is the most exciting growth area if you enjoy data and signal processing.

The complete guide to iot and edge ai

Embedded Systems and the Internet of Things
Read the guide

Keep exploring

Guides from other topics

All topics →