IN Brief:
- The kit pairs Cortex-M55 and Cortex-M33 processing with an Ethos-U55 NPU and NNLite accelerator.
- Imaging, microphones, motion sensors, environmental sensing, and 60GHz radar are included within one development platform.
- Multiple processing and sensing domains allow developers to test local inference against realistic power, memory, and latency limits.
Infineon Technologies has introduced an evaluation platform built around its PSoC Edge E84 device, combining heterogeneous Arm processing, neural-network acceleration, wireless communications, imaging, audio, environmental sensing, and 60GHz radar.
Processing is divided between a Cortex-M55 core running at up to 400MHz and a Cortex-M33 operating at up to 200MHz. Machine-learning workloads can be assigned to an Arm Ethos-U55 neural processing unit or Infineon’s NNLite accelerator, enabling developers to balance model performance, energy consumption, memory demand, and response time across the available compute resources.
Alongside Wi-Fi 6 and 6E connectivity and Bluetooth 5.4, the KIT_PSE84_AI board carries external flash and RAM, an inertial measurement unit, magnetometer, pressure sensor, microphones, camera hardware, and a 60GHz radar device. Several sensing modes can therefore be assessed without assembling a stack of unrelated expansion boards before application development begins.
Software support extends across Infineon’s ModusToolbox environment, the Zephyr real-time operating system, and the company’s DEEPCRAFT machine-learning tools. Supplied examples establish peripheral operation, sensor acquisition, model execution, and communications, after which the reference algorithms and datasets can be replaced with application-specific material.
Local inference can reduce the quantity of raw information sent across a network, particularly when cameras, microphones, or radar devices generate continuous streams. Instead of transferring every sample, a product can identify an event, classify an object, or calculate a condition locally before reporting a smaller result to another controller or cloud service.
Multimodal inference increases system complexity
Edge-AI development is shifting from single-input demonstrations towards systems that combine vision, sound, motion, radar, and environmental data. A camera can classify an object, while radar adds range and movement information under poor lighting; microphones can identify acoustic events, and an inertial sensor can distinguish movement of the target from movement of the equipment itself.
Combining those sources can improve confidence, although every sensor operates at a different data rate and carries distinct noise, calibration, and timing characteristics. Samples must be aligned closely enough for a model to associate motion detected by radar with an image or sound, while buffering and preprocessing cannot introduce delays that make the combined result less useful than a simpler detector.
Power management becomes more demanding as additional sensing modes are activated because a camera, radio, radar front end, and application processor cannot remain at full power continuously in most battery-operated systems. Lower-power sensors or timing events can wake the more demanding functions only when closer analysis is needed, with the Cortex-M33, Cortex-M55, NPU, and NNLite accelerator handling different stages of the process.
Memory capacity will constrain the models that can run locally even where sufficient arithmetic performance is available. Weights, intermediate tensors, sensor buffers, operating-system services, and communications stacks all compete for memory, while external RAM increases power consumption and may widen the attack surface. Quantisation, pruning, model selection, and careful buffer reuse remain essential parts of embedded AI implementation.
Those pressures formed part of the wider movement towards edge processing examined in the second-quarter review of electronics investment and computing, where processor capability was increasingly linked to the infrastructure required to deploy and maintain models. Evaluation hardware can make initial inference straightforward, yet production equipment must operate within fixed limits for memory, power, latency, temperature, and component cost.
Model maintenance continues after the hardware has entered service because performance can degrade when lighting, acoustic conditions, machine configurations, or user populations differ from the original training data. Developers need controlled methods for collecting representative datasets, recording model versions, testing updates, and restoring an earlier version when a replacement performs poorly.
Wireless connectivity makes remote updates practical, while secure boot, authenticated firmware, protected credentials, and restricted debug access remain necessary to prevent an altered model or application image from changing system behaviour. Model files deserve the same configuration discipline as executable software, particularly where an inference result influences access control, machine movement, or medical monitoring.
The PSoC Edge E84 kit places a broad collection of compute and sensing resources within one environment, shortening the route to a working multimodal prototype. Converting that prototype into a manufacturable product will require fewer devices, controlled radio performance, application-specific mechanical design, and a software architecture that can be maintained throughout the equipment’s service life.
