Embedded Technology
and produce the best heating profile. In systems that rely on manual control, the model can detect whether an operator has selected an inappropriate program and generate an alert as needed. For these applications, the VectorBlox platform provides a high-performance combination of multicore processing using the hardened RISC-V-based processing system and FPGA logic. VectorBlox comprises a vector-processing unit in combination with a scheduler, both implemented in FPGA fabric. This versatile unit, thanks to the use
of programmable logic, works in concert with an on-board processor.
VectorBlox can take networks in TensorFlow Lite format and convert them to the target hardware with little manual effort. This allows rapid prototyping of the applications, with the option for further customisation to tune the hardware to the AI models’ needs.
For applications such as smart sensors for predictive maintenance, environmental monitoring and healthcare, it makes sense to use a scaled-down platform. We can think of
the target platform as the e-bike of the AI ecosystem. It has low overhead and delivers food and other parcels quickly and easily. With a little extra hardware, you can build smaller modes of transport that can take routes denied for larger vehicles. The microcontroller-based e-bike of the AI world feeds a model with data from a selection of complementary sensors, which may measure temperature, vibration and acceleration. The relatively low data rates and 1D nature of the sensor streams mean the AI model can run on a 32-bit microcontroller. An additional FPGA may assist with data management as well as providing support for other functions such as motor control.
Frameworks such as Stream Analyze and INFXL provide the means to adapt and manage updates to AI models tuned to run on MCU-based targets through a combination of quantisation and pruning. The solution based on INFXL also allows us to implement the MCU C-code directly in FPGA- fabric by using Microchip’s high-level synthesis SmartHLS software. That way it can easily be added on existing FPGA-systems, already doing control of equipment. Scaling in the other direction delivers AI models able to cope not just with inputs from multiple cameras and other sensors but operate the motors and actuators
www.cieonline.co.uk
inside a complex humanoid robot. In this environment, the ability to convert and remap data in real time becomes essential. Converting camera data into different colour planes can help improve CNN throughput. Other organisations may deliver data in a way that suits the memory-access and cache structure of the target AI accelerator. Using one or more FPGAs to act as the conversion bridge brings with it maximum flexibility: models can interface with any sensor with no internal changes. Bridges will often deliver data using one of the many forms of Ethernet, starting from 100M fast Ethernet for lower-performance systems to 10 gigabit Ethernet or more. Microchip’s PolarFire Ethernet Sensor Bridge provides an example of this functionality, letting developers try different hardware and software architectures easily. Carrying pre-processed data to target GPUs using 10 gigabit Ethernet, the Sensor Bridge has inputs for cameras that use the common MIPI CSI-2 port or can receive HDMI 1.4 data. An onboard FMC connector provides access to other high-speed I/O, such as JESD-204B analogue-to-digital converters, using add-on cards.
From the lightest e-bike to the heaviest truck, vehicle design has adapted to support the full range of transportation needs. Physical and edge AI are the same. Suppliers such as Microchip can offer support for the full range of performance options, from microcontrollers to sensor-laden, GPU-based systems. With FPGA hardware such as the PolarFire range of devices, these platforms provide the ability to performance-tune every implementation.
https://www.microchip.com/ Components in Electronics July/August 2026 39
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48
